Fibo-2.0-dev-OrbitQuant-W4A8: OrbitQuant W4A8 transformer and text encoder

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+fonts/DejaVuSans.ttf filter=lfs diff=lfs merge=lfs -text
+processor/tokenizer.json filter=lfs diff=lfs merge=lfs -text

fibo2_orbitquant.py

@@ -0,0 +1,158 @@
+"""Load Fibo 2.0 with OrbitQuant W4A8 components into diffusers' ``BriaFibo2Pipeline``."""
+
+from __future__ import annotations
+
+import dataclasses
+import json
+import os
+from pathlib import Path
+
+import torch
+from torch import nn
+
+GROUPS = ("qkv", "attn_out", "ff_in", "ff_out", "xq", "xkv", "xout")
+# The deepest Qwen3-VL hidden state the pipeline reads is 35, the output of layer 34.
+TEXT_LAYERS = 35
+
+# Negative captions of the original model card: text-to-image steers away from soft, featureless, artifact-laden
+# images; edits use the empty caption the model was trained with.
+NEGATIVE_T2I = '{"short_description":"","objects":[{"description":"the main subject","texture":"smooth, featureless, plastic-like, airbrushed, no fine surface detail","appearance_details":"soft blurred edges, simplified shapes, no fine detail","skin_tone_and_texture":"waxy airbrushed skin, poreless, plastic smooth"}],"background_setting":"a plain seamless studio backdrop, perfectly smooth gradient, blurred and featureless, no visible texture","lighting":{},"aesthetics":{"artifacts":["oversaturated colours","neon colour cast","overexposed highlights","lens flare","slight blur","motion blur","haze"]},"photographic_characteristics":{"depth_of_field":"shallow, heavy bokeh obscuring most of the frame","focus":"soft, slightly out of focus, lacking fine texture"},"style_medium":"","context":"","artistic_style":""}'
+NEGATIVE_EDIT = '{"short_description":"","objects":[],"background_setting":"","lighting":{},"aesthetics":{},"photographic_characteristics":{},"style_medium":"","context":"","artistic_style":""}'
+
+
+def _allocator(device: str) -> None:
+    # Growable segments return the transient blocks of one stage to the next. An explicit setting is left untouched.
+    if not device.startswith("cuda") or {"PYTORCH_CUDA_ALLOC_CONF", "PYTORCH_ALLOC_CONF"} & set(os.environ):
+        return
+    if not torch.cuda.is_initialized():
+        os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+    else:
+        setter = getattr(torch._C, "_accelerator_setAllocatorSettings", None)
+        (setter or torch.cuda.memory._set_allocator_settings)("expandable_segments:True")
+
+
+def _load_transformer(path: Path, device: str, fused: bool, attention: str, activation_bits: int):
+    import orbitquant  # noqa: F401 -- registers the OrbitQuant layers
+    from diffusers import BriaFibo2Transformer2DModel as Transformer
+    from orbitquant.artifacts.manifest import OrbitQuantManifest
+    from orbitquant.config import OrbitQuantConfig
+    from orbitquant.layers import OrbitQuantLinear
+    from orbitquant.modeling import _parent_and_child, _set_child
+    from safetensors.torch import load_file
+
+    config = OrbitQuantConfig.from_dict(json.loads((path / "quantization_config.json").read_text()))
+    manifest = OrbitQuantManifest.from_dict(json.loads((path / "orbitquant_manifest.json").read_text()))
+    with torch.device("meta"):
+        model = Transformer.from_config(Transformer.load_config(str(path)))
+    for name in manifest.quantized_modules:
+        bits = manifest.module_bits.get(name)
+        layer_config = dataclasses.replace(config, weight_bits=bits) if bits and bits != config.weight_bits else config
+        parent, child = _parent_and_child(model, name)
+        _set_child(parent, child, OrbitQuantLinear.empty_from_linear(
+            model.get_submodule(name), config=layer_config, module_name=name))
+    model.load_state_dict(load_file(str(path / "model.safetensors"), device=device), strict=True, assign=True)
+    model = model.to(device).eval()
+    if fused:
+        from orbitquant.fused import fuse
+
+        fuse(model, config=config, attention=attention, activation_bits=dict.fromkeys(GROUPS, activation_bits))
+    return model
+
+
+def _park(model: nn.Module) -> None:
+    """Move a component to host memory right after its GPU build, so the next build does not stack on it."""
+    from orbitquant.fused import release
+
+    release(model)
+    model.to("cpu")
+    torch.cuda.empty_cache()
+
+
+class HostEmbedding(nn.Module):
+    """Token embedding kept in host memory: a prompt reads a few hundred of its 152k rows, so the 0.8 GB table stays
+    off the GPU. The table is not a registered parameter, so moving the text encoder does not move it."""
+
+    def __init__(self, embedding: nn.Embedding):
+        super().__init__()
+        self.__dict__["table"] = embedding.weight.detach().to("cpu")
+        self.num_embeddings, self.embedding_dim = self.table.shape
+
+    @property
+    def weight(self):
+        return self.table
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        return nn.functional.embedding(ids.cpu(), self.table).to(ids.device, non_blocking=True)
+
+
+def _load_text_encoder(path: Path, device: str, fused: bool, activation_bits: int, host_embedding: bool):
+    """Qwen3-VL without the LM head and the unused last decoder layer. The final norm becomes the identity, so the
+    last recorded hidden state is the raw output of layer 34, as in the full model; the vision tower stays BF16 for
+    the images to edit and the drawn caption texts."""
+    import orbitquant  # noqa: F401 -- registers the OrbitQuant backend with Transformers
+    from transformers import Qwen3VLForConditionalGeneration
+
+    full = Qwen3VLForConditionalGeneration.from_pretrained(str(path), dtype=torch.bfloat16, device_map={"": device})
+    model = full.model.eval()
+    language = model.language_model
+    language.layers = language.layers[:TEXT_LAYERS]
+    language.config.num_hidden_layers = TEXT_LAYERS
+    if getattr(language.config, "layer_types", None):
+        language.config.layer_types = language.config.layer_types[:TEXT_LAYERS]
+    language.norm = nn.Identity()
+    del full
+    if fused:
+        from orbitquant.fused import fuse_linears
+
+        fuse_linears(model, activation_bits=activation_bits)
+    if host_embedding:
+        language.embed_tokens = HostEmbedding(language.embed_tokens)
+    torch.cuda.empty_cache()
+    return model
+
+
+def load_pipeline(path: str | Path, device: str = "cuda", attention: str = "cudnn", offload: bool = False,
+                  fused: bool | None = None, activation_bits: int = 8, host_embedding: bool = True):
+    """Fibo 2.0 pipeline with the W4A8 transformer and text encoder; call it as the original ``BriaFibo2Pipeline``.
+
+    ``attention`` is ``cudnn`` (default), ``flash``, ``sdpa`` or ``int8`` (faster, changes the image slightly).
+    ``offload`` keeps the components in host memory and moves each to the GPU only while it runs. ``fused=None``
+    fuses the projections on CUDA devices with compute capability 8.0 or newer (needed for 8-bit activations)."""
+    path = Path(path)
+    _allocator(device)
+    from diffusers import AutoencoderKLFlux2, BriaFibo2Pipeline, FlowMatchEulerDiscreteScheduler
+    from transformers import Qwen3VLProcessor
+
+    if fused is None:
+        fused = device.startswith("cuda") and torch.cuda.get_device_capability(device) >= (8, 0)
+    # The transformer is fused on the GPU even with offload: on the host the group build runs many small parallel
+    # copies, which stall for minutes when a container's CPU quota is far below its visible core count.
+    transformer = _load_transformer(path / "transformer", device, fused, attention, activation_bits)
+    if offload:
+        _park(transformer)
+    text_encoder = _load_text_encoder(path / "text_encoder", device, fused, activation_bits, host_embedding)
+    if offload:
+        _park(text_encoder)
+    vae = AutoencoderKLFlux2.from_pretrained(str(path / "vae"), dtype=torch.bfloat16)
+    vae = vae.to("cpu" if offload else device).eval()
+    pipeline = BriaFibo2Pipeline(
+        transformer=transformer,
+        scheduler=FlowMatchEulerDiscreteScheduler.from_pretrained(str(path / "scheduler")),
+        vae=vae,
+        text_encoder=text_encoder,
+        processor=Qwen3VLProcessor.from_pretrained(str(path / "processor")),
+    )
+    if offload:
+        pipeline.enable_model_cpu_offload(device=device)
+    torch.cuda.empty_cache()
+    return pipeline
+
+
+def fonts(path: str | Path) -> list[str]:
+    """The fonts Fibo 2.0's training drew caption texts with (pass them as ``fonts=``)."""
+    return [str(Path(path) / "fonts" / name) for name in ("Aileron-Regular.otf", "DejaVuSans.ttf")]
+
+
+def sampling(path: str | Path) -> dict:
+    """Default steps and guidance of the release (Dev: 50 and 5.0; Edit Turbo: 4 and 1.0)."""
+    return json.loads((Path(path) / "manifest.json").read_text())["sampling"]

fonts/DejaVuSans.ttf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:437a5ecec2ec177cba9284cbffc56055311347f3e9fe4f2b4b40323d0290a6db
+size 757076

fonts/LICENSE-Aileron.txt

@@ -0,0 +1,12 @@
+Aileron Regular, version 1.102 (Aileron-Regular.otf)
+Author: Sora Sagano (dot colon), https://www.dotcolon.net/font/aileron
+Downloaded 2026-09-15 from https://www.dotcolon.net/files/fonts/aileron_0102.zip
+
+Licence: CC0 1.0 Universal (Public Domain Dedication)
+https://creativecommons.org/publicdomain/zero/1.0/
+
+The font's own name table (nameID 0) reads: "No Rights Reserved."
+
+To the extent possible under law, the author has waived all copyright and related or neighboring
+rights to this work. You can copy, modify, distribute and perform the work, even for commercial
+purposes, all without asking permission. The work is provided "as is", without warranty of any kind.

fonts/LICENSE-DejaVu.txt

@@ -0,0 +1,78 @@
+Format: http://www.debian.org/doc/packaging-manuals/copyright-format/1.0/
+Upstream-Name: DejaVu fonts
+Upstream-Author: Stepan Roh <src@users.sourceforge.net> (original author),
+                  see /usr/share/doc/ttf-dejavu/AUTHORS for full list
+Source: http://dejavu-fonts.org/
+
+Files: *
+Copyright: Copyright (c) 2003 by Bitstream, Inc. All Rights Reserved. 
+ Bitstream Vera is a trademark of Bitstream, Inc.
+ DejaVu changes are in public domain.
+License: bitstream-vera
+ Permission is hereby granted, free of charge, to any person obtaining a copy
+ of the fonts accompanying this license ("Fonts") and associated
+ documentation files (the "Font Software"), to reproduce and distribute the
+ Font Software, including without limitation the rights to use, copy, merge,
+ publish, distribute, and/or sell copies of the Font Software, and to permit
+ persons to whom the Font Software is furnished to do so, subject to the
+ following conditions:
+ .
+ The above copyright and trademark notices and this permission notice shall
+ be included in all copies of one or more of the Font Software typefaces.
+ .
+ The Font Software may be modified, altered, or added to, and in particular
+ the designs of glyphs or characters in the Fonts may be modified and
+ additional glyphs or characters may be added to the Fonts, only if the fonts
+ are renamed to names not containing either the words "Bitstream" or the word
+ "Vera".
+ .
+ This License becomes null and void to the extent applicable to Fonts or Font
+ Software that has been modified and is distributed under the "Bitstream
+ Vera" names.
+ .
+ The Font Software may be sold as part of a larger software package but no
+ copy of one or more of the Font Software typefaces may be sold by itself.
+ .
+ THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF MERCHANTABILITY,
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT OF COPYRIGHT, PATENT,
+ TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL BITSTREAM OR THE GNOME
+ FOUNDATION BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, INCLUDING
+ ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES,
+ WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
+ THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM OTHER DEALINGS IN THE
+ FONT SOFTWARE.
+ .
+ Except as contained in this notice, the names of Gnome, the Gnome
+ Foundation, and Bitstream Inc., shall not be used in advertising or
+ otherwise to promote the sale, use or other dealings in this Font Software
+ without prior written authorization from the Gnome Foundation or Bitstream
+ Inc., respectively. For further information, contact: fonts at gnome dot
+ org.
+
+Files: debian/*
+Copyright: (C) 2005-2006 Peter Cernak <pce@users.sourceforge.net> 
+           (C) 2006-2011 Davide Viti <zinosat@tiscali.it>
+           (C) 2011-2013 Christian Perrier <bubulle@debian.org>
+           (C) 2013 Fabian Greffrath <fabian+debian@greffrath.com>
+License: GPL-2+
+ This program is free software; you can redistribute it
+ and/or modify it under the terms of the GNU General Public
+ License as published by the Free Software Foundation; either
+ version 2 of the License, or (at your option) any later
+ version.
+ .
+ This program is distributed in the hope that it will be
+ useful, but WITHOUT ANY WARRANTY; without even the implied
+ warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR
+ PURPOSE.  See the GNU General Public License for more
+ details.
+ .
+ You should have received a copy of the GNU General Public
+ License along with this package; if not, write to the Free
+ Software Foundation, Inc., 51 Franklin St, Fifth Floor,
+ Boston, MA  02110-1301 USA
+ .
+ On Debian systems, the full text of the GNU General Public
+ License version 2 can be found in the file
+ /usr/share/common-licenses/GPL-2'.

fonts/README.md

@@ -0,0 +1,21 @@
+# fonts
+
+Fallback fonts for drawing the images of a caption's text (`render_text_images=True` in
+`BriaFibo2Pipeline`), the way fibo-2's training drew them.
+
+Each `text_render` text is drawn black on white at size 20. A character is drawn with the first of
+these that has a glyph for it:
+
+| order | font | source | licence |
+|---|---|---|---|
+| 1 | Pillow's built-in font, `ImageFont.load_default(20)` (a subset of Aileron Regular) | bundled with Pillow, not in this folder | — |
+| 2 | `Aileron-Regular.otf`, Aileron Regular 1.102 | https://www.dotcolon.net/font/aileron | CC0 1.0, `LICENSE-Aileron.txt` |
+| 3 | `DejaVuSans.ttf`, DejaVu Sans 2.37 | https://dejavu-fonts.github.io | Bitstream Vera licence, `LICENSE-DejaVu.txt` |
+
+| file | bytes | md5 |
+|---|---|---|
+| `Aileron-Regular.otf` | 27,644 | `d321fa78bb7190a8ca7e14213ef63203` |
+| `DejaVuSans.ttf` | 757,076 | `517a360397bbb5a7573c2558e78fa4ea` |
+
+Training used Pillow 12.2. Characters none of the three fonts covers (CJK, emoji) are drawn as the
+built-in font's missing-glyph box, as in training.

manifest.json

@@ -0,0 +1,41 @@
+{
+  "format": "fibo2-orbitquant",
+  "format_version": 1,
+  "source_model": "briaai/Fibo-2.0-dev",
+  "source_revision": "fb2b781f9e9e5572a6c348a5af7a4d155a2b1ed0",
+  "source_license": "cc-by-nc-4.0",
+  "runtime": {
+    "orbitquant": "0.15.0",
+    "torch": "2.12.1",
+    "transformers": "5.20.1",
+    "diffusers": "git+https://github.com/huggingface/diffusers@48cdbaef5ee3d718aaee392b44f2099ae511c4a5"
+  },
+  "sampling": {
+    "steps": 50,
+    "guidance_scale": 5.0
+  },
+  "components": {
+    "transformer": {
+      "precision": "OrbitQuant 4-bit weights; 8-bit activations at run time",
+      "quantized_projections": 364,
+      "bf16": [
+        "x_embedder",
+        "t_embedder",
+        "perceiver.latent_queries",
+        "adaLN_modulation (every block)",
+        "final_layer",
+        "norms"
+      ],
+      "bytes": 4329652163
+    },
+    "text_encoder": {
+      "precision": "OrbitQuant 4-bit weights for the language model; 8-bit activations at run time; BF16 vision tower and token embeddings",
+      "quantized_projections": 252,
+      "bytes": 3428012665
+    },
+    "vae": {
+      "precision": "original (FP32 file, loaded in BF16 as in the original pipeline)",
+      "bytes": 336214465
+    }
+  }
+}

processor/chat_template.jinja

@@ -0,0 +1,120 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' }}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- for message in messages %}
+    {%- if message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content_item in message.content %}
+                {%- if 'text' in content_item %}
+                    {{- content_item.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and message.content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

processor/processor_config.json

@@ -0,0 +1,61 @@
+{
+  "image_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "merge_size": 2,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "size": {
+      "longest_edge": 1605632,
+      "shortest_edge": 3136
+    },
+    "temporal_patch_size": 2
+  },
+  "processor_class": "Qwen3VLProcessor",
+  "video_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "do_sample_frames": true,
+    "fps": 2,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_frames": 768,
+    "max_video_tokens": 768,
+    "merge_size": 2,
+    "min_frames": 4,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "return_metadata": false,
+    "size": {
+      "longest_edge": 25165824,
+      "shortest_edge": 4096
+    },
+    "temporal_patch_size": 2,
+    "video_processor_type": "Qwen3VLVideoProcessor"
+  }
+}

processor/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650

processor/tokenizer_config.json

@@ -0,0 +1,31 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen3VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

requirements.txt

@@ -0,0 +1,8 @@
+orbitquant[hf]==0.15.0
+diffusers @ git+https://github.com/huggingface/diffusers@48cdbaef5ee3d718aaee392b44f2099ae511c4a5
+transformers==5.20.1
+accelerate
+safetensors
+huggingface_hub
+pillow
+numpy

run_fibo2.py

@@ -0,0 +1,67 @@
+"""Generate or edit one image with Fibo 2.0 OrbitQuant W4A8.
+
+python run_fibo2.py --prompt caption.json --out image.png [--width 1024 --height 1024]
+python run_fibo2.py --prompt edit.json --image photo.jpg [--image ref.png] [--mask mask.png] --out edited.png
+--prompt takes a VGL caption (a JSON file or string) or plain text. Steps and guidance default to the release's.
+"""
+import argparse
+import json
+import os
+import sys
+import time
+from pathlib import Path
+
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+import torch  # noqa: E402
+from PIL import Image  # noqa: E402
+
+HERE = Path(__file__).resolve().parent
+sys.path.insert(0, str(HERE))
+from fibo2_orbitquant import NEGATIVE_EDIT, NEGATIVE_T2I, fonts, load_pipeline, sampling  # noqa: E402
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--model", default=str(HERE), help="release directory")
+ap.add_argument("--prompt", required=True, help="VGL caption: a .json file or a JSON string; plain text also works")
+ap.add_argument("--image", action="append", default=[], help="image to edit; repeat for up to 4 references")
+ap.add_argument("--mask", default=None, help="white marks the region to regenerate (one image)")
+ap.add_argument("--negative-prompt", default=None, help="default: the model card's negative for the mode")
+ap.add_argument("--width", type=int, default=None)
+ap.add_argument("--height", type=int, default=None)
+ap.add_argument("--steps", type=int, default=None)
+ap.add_argument("--cfg", type=float, default=None)
+ap.add_argument("--seed", type=int, default=0)
+ap.add_argument("--attention", default="cudnn", choices=["cudnn", "flash", "sdpa", "int8"])
+ap.add_argument("--offload", action="store_true", help="keep components in host memory between stages")
+ap.add_argument("--out", default="image.png")
+args = ap.parse_args()
+
+prompt = Path(args.prompt).read_text() if args.prompt.endswith(".json") and Path(args.prompt).is_file() else args.prompt
+if prompt.lstrip().startswith("{"):
+    prompt = json.dumps(json.loads(prompt), ensure_ascii=False)
+defaults = sampling(args.model)
+steps = args.steps or defaults["steps"]
+cfg = defaults["guidance_scale"] if args.cfg is None else args.cfg
+images = [Image.open(p).convert("RGB") for p in args.image]
+kwargs = dict(prompt=prompt, num_inference_steps=steps, guidance_scale=cfg, fonts=fonts(args.model),
+              generator=torch.Generator("cpu").manual_seed(args.seed))
+if images:
+    kwargs["image"] = images if len(images) > 1 else images[0]
+if args.mask:
+    kwargs["mask"] = Image.open(args.mask).convert("L")
+if args.width or not images:
+    kwargs["width"] = args.width or 1024
+if args.height or not images:
+    kwargs["height"] = args.height or 1024
+if cfg > 1:
+    kwargs["negative_prompt"] = args.negative_prompt or (NEGATIVE_EDIT if images else NEGATIVE_T2I)
+
+t0 = time.perf_counter()
+pipe = load_pipeline(args.model, attention=args.attention, offload=args.offload)
+print(f"loaded in {time.perf_counter() - t0:.1f} s", flush=True)
+t0 = time.perf_counter()
+image = pipe(**kwargs).images[0]
+torch.cuda.synchronize()
+print(f"generated in {time.perf_counter() - t0:.1f} s, device peak {torch.cuda.max_memory_allocated() / 2**30:.2f} GiB",
+      flush=True)
+image.save(args.out)

scheduler/scheduler_config.json

@@ -0,0 +1,18 @@
+{
+  "_class_name": "FlowMatchEulerDiscreteScheduler",
+  "_diffusers_version": "0.41.0.dev0",
+  "base_image_seq_len": 256,
+  "base_shift": 0.5,
+  "invert_sigmas": false,
+  "max_image_seq_len": 4096,
+  "max_shift": 1.15,
+  "num_train_timesteps": 1000,
+  "shift": 3.0,
+  "shift_terminal": null,
+  "stochastic_sampling": false,
+  "time_shift_type": "exponential",
+  "use_beta_sigmas": false,
+  "use_dynamic_shifting": true,
+  "use_exponential_sigmas": false,
+  "use_karras_sigmas": false
+}

styles/boring_reality/negative_prompt.json

@@ -0,0 +1 @@
+{"short_description":"","objects":[{"description":"the main subject","texture":"smooth, featureless, plastic-like, airbrushed, no fine surface detail","appearance_details":"soft blurred edges, simplified shapes, no fine detail","skin_tone_and_texture":"waxy airbrushed skin, poreless, plastic smooth"}],"background_setting":"a plain seamless studio backdrop, perfectly smooth gradient, blurred and featureless, no visible texture","lighting":{},"aesthetics":{"artifacts":["oversaturated colours","neon colour cast","overexposed highlights","lens flare","slight blur","motion blur","haze"]},"photographic_characteristics":{"depth_of_field":"shallow, heavy bokeh obscuring most of the frame","focus":"soft, slightly out of focus, lacking fine texture"},"style_medium":"","context":"","artistic_style":""}

styles/boring_reality/system_prompt.txt

@@ -0,0 +1,75 @@
+You are a documentary photo editor at a leading Generative AI company.
+Your job is to take a user's short idea and write, in JSON, a precise account of a single real photograph of that idea — a photograph that already exists and that you are describing, not one you are commissioning.
+You must creatively supply the details the user did not state: the setting, the light, the other objects present, the surfaces and their condition. Everything you supply must be the kind of detail a camera records, and the whole description must hold together as one photograph.
+Adhere strictly to the following structure and guidelines:
+The output MUST be ONLY a valid JSON object. Do not include any text before or after the JSON object (e.g., no "Here is the JSON:", no explanations, no apologies).
+
+THE ONE RULE THAT MATTERS MOST: describe, do not art-direct. Every field records something that is physically there. A field that expresses an intention, a purpose, or an opinion produces glossy commercial imagery instead of a photograph.
+* BANNED — commercial framing, in every field: "advertisement", "ad campaign", "greeting card", "calendar", "magazine", "editorial", "stock photo", "poster", "brochure", "concept for", "suitable for", "intended for", "marketing", "campaign", "brand".
+* BANNED — evaluative words, in every field: "beautiful", "stunning", "gorgeous", "breathtaking", "perfect", "flawless", "pristine", "immaculate", "high-quality", "professional", "masterful", "captivating", "charming", "lovely", "idyllic", "picturesque", "vibrant", "lush", "sun-drenched", "glowing", "radiant", "magical", "whimsical", "heartwarming", "serene", "innocent", "inspiring".
+* Replace every banned word with the physical fact behind it. Not "beautiful weathered wood" but "wood with raised grain and flaking grey paint". Not "vibrant market" but "crates of tomatoes and packed shoppers under strip lighting".
+
+DENSITY: thin JSON produces an under-specified image, and whatever you leave out the model invents. Fill 3-6 objects, fill every optional field of every object that the scene determines, and write a `short_description` of 400-600 characters. The training captions this model learned from average about 5,300 characters in total — aim for that scale.
+
+IMPERFECTION IS THE EVIDENCE OF REALITY: real scenes are uneven. Include wear, dust, fingerprints, scuffs, creases, stray hairs, moisture, asymmetry, a thing slightly out of place, skin with pores and fine lines. A scene where everything is clean, new and perfectly arranged reads as a render.
+
+OPTICS: most photographs are mostly in focus, lit by ordinary light, taken from an ordinary height with an ordinary lens. Prefer deep or moderate depth of field; use shallow depth of field only when the user asks for a blurred background. Prefer overcast, open shade, indoor daylight or plain sun over golden hour unless the prompt implies otherwise.
+
+INTERNAL CONSISTENCY: every field describes ONE photograph. One lighting condition, not two. One composition, not two. If the subject faces left, every field that mentions direction says left. If the light comes from the right, the shadows fall to the left. Contradictions between fields are the most common way these descriptions fall apart.
+
+IMPORTANT: When describing human body parts, positions, or actions, always describe them from the PERSON'S OWN PERSPECTIVE, not from the observer's viewpoint. For example, if a person's left arm is raised (from their own perspective), describe it as "left arm" even if it appears on the right side of the image from the viewer's perspective.
+RULE for Human Subjects: When the user's prompt features a person or people as the main subject, you MUST default to a composition that frames them prominently. Aim for compositions where their face and upper body are a primary focus (e.g., 'medium shot', 'close-up'). Avoid defaulting to 'wide-angle' or 'full-body' shots where the face is small, unless the user's prompt specifically implies a large scene (e.g., "a person standing on a mountain").
+Unless the user's prompt explicitly requests a different style (e.g., 'painting', 'cartoon', 'illustration'), you MUST default to `style_medium: "photograph"` and aim for the highest degree of photorealism. In such cases, `artistic_style` should be "realistic".
+COLOUR: also fill `color_palette` (the whole image, 8 entries, ordered most to least dominant) and, for each object, `colors` (2-4 entries, most to least dominant). Every entry is an object with a single `name` field, chosen from the enumerated colour vocabulary. Choose the specific named colour that genuinely matches the surface you described, not the nearest basic colour: a bay horse is "Coffee" or "Café Noir", not "Brown". The palette must agree with the objects, the background and the lighting you have already described.
+Choose every colour name from this vocabulary, which is the one the model was trained on (most common first):
+Eerie Black, White, Raisin Black, Coyote Brown, Spanish Gray, Silver Chalice, Charcoal, Café Au Lait, Copper, Isabelline, Davy's Grey, Dark Jungle Green, Silver Pink, Café Noir, Platinum, Silver, Jet, Camel, Dark Liver, Rifle Green, Bistre, Black, Umber, Olive Drab #7, Coffee, Shadow, Dark Sky Blue, Light Gray, Lion, Gunmetal, Chestnut, Dark Olive Green, Granite Gray, Desert Sand, Smoky Black, Bone, Dark Slate Gray, Red, Black Olive, Space Cadet, Zinnwaldite Brown, Gainsboro, Beaver, Dark Lava, Rosy Brown, Blue, Carnelian, Deep Taupe, Slate Gray, Ebony, Yellow, Brown Sugar, Light Steel Blue, Dark Electric Blue, Sonic Silver, Maximum Green, Tan, Cultured Pearl, Moss Green, Ash Gray, Iceberg, Metallic Sunburst, Orange Peel, Queen Blue, Light Sky Blue, French Bistre, Patriarch, Saffron, Steel Blue, Antique Bronze, Olive Green, Cedar Chest, Light French Beige, Independence, Light Blue, Sage, Raw Umber, Lavender Gray, Cinnabar, YInMn Blue, Liver Chestnut, Satin Sheen Gold, Pink, May Green, Taupe Gray, Dim Gray, Fern Green, Hunter Green, Spanish Bistre, Middle Grey, Green, Drab Dark Brown, Dark Purple, Glaucous, India Green, Cinereous, Copper Red, Wild Blue Yonder, Gold Fusion, Shadow Blue, Saddle Brown, Antique Brass, Xanadu, Kobicha, Maximum Blue, Quick Silver, Fuchsia, Magnolia, Battleship Grey, Van Dyke Brown, Fawn, Silver Sand, Pewter Blue, Bronze, Blue Jeans, Copper Penny, Gold, Bud Green, True Blue, Opal.
+If no name here fits a surface, use the closest one rather than inventing a new name.
+
+The JSON object must contain the following keys precisely:
+1.  `short_description`: (String) A factual summary of what is visible in the photograph, 400-600 characters. Describe the scene as an observer would report it, not as a caption written to sell it.
+2.  `objects`: (Array of Objects) List 3 to 6 prominent objects. Include the things the setting really contains rather than inventing clutter to reach a count, but do not stop at the single main subject — a photograph of one thing still has a surface it rests on, things beside it, and things behind it. For each object, include:
+    * `description`: (String) A detailed description of the object, 100 words maximum. Name materials, construction, wear and condition.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `relative_size`: (String) E.g., "small", "medium", "large within frame". (If a person is the main subject, this should be "medium-to-large" or "large within frame").
+    * `shape_and_color`: (String) The object's actual silhouette and its specific colour. Name the form ("a low four-legged stance", "a tapered cylinder", "a squat rectangular body") and a specific colour ("dark reddish-brown", "faded olive green"). Never "irregular", "rounded" or "basic shape" on their own, and never a bare colour word where a specific one exists.
+    * `texture`: (String) The fine surface character, at the scale a viewer would see it up close. E.g., "coarse wire-like guard hairs over a dense undercoat", "open-grain oak with raised ridges", "brushed steel with fine circular scratches". Avoid the bare word "smooth" unless the surface genuinely is polished.
+    * `appearance_details`: (String) The specific marks that make this object a real one rather than a generic one — chips, stains, scuffs, dust, moisture, repairs, uneven wear, small asymmetries.
+    * `relationship`: (String) Describe the relationship between the object and the other objects in the image.
+    * `orientation`: (String) Describe the orientation or positioning of the object, e.g., "upright", "tilted 45 degrees", "horizontal", "vertical", "facing left", "facing right", "upside down", "lying on its side".
+    * If the object is a human or a human-like object, include the following:
+        * `pose`: (String) Describe the body position.
+        * `expression`: (String) The facial expression actually on the face. Ordinary expressions are the common case: "neutral, mouth slightly open in concentration", "eyes narrowed against the light", "mid-blink". Avoid posed or theatrical expressions unless the prompt asks for one.
+        * `clothing`: (String) Describe attire, including fit, fabric, and how worn or creased it is.
+        * `action`: (String) Describe the action of the human.
+        * `gender`: (String) Describe the gender of the human.
+        * `skin_tone_and_texture`: (String) The skin tone AND its real surface: visible pores, fine lines, uneven colouring, stubble, freckles, flushing, shine where skin is oily and matte where it is not. Never describe skin as flawless, poreless, glowing or airbrushed.
+    * `colors`: (Array of Objects) The 2-4 dominant colours of THIS object, most dominant first, each an object with a single `name` field drawn from the colour vocabulary.
+    * If the object is a cluster of objects, include the following:
+        * `number_of_objects`: (Integer) The number of objects in the cluster.
+3.  `background_setting`: (String) Describe the overall environment, setting, or background, including any notable background elements that are not part of the `objects` section. Give it real structure and texture — a background is a place, not a blur.
+4.  `lighting`: (Object) Name ONE lighting condition. Two at once ("bright daylight, golden hour") describes two different photographs.
+    * `conditions`: (String) E.g., "overcast daylight", "open shade", "indoor daylight through a north window", "plain midday sun", "a single overhead fluorescent". Ordinary light is usually the truthful choice; use golden hour only when the prompt implies it.
+    * `direction`: (String) E.g., "front-lit", "backlit", "side-lit from left".
+    * `shadows`: (String) Describe the presence and quality of shadows, e.g., "long, soft shadows", "sharp, defined shadows", "minimal shadows". These must be consistent with `direction`: light from the right puts shadows on the left.
+5.  `aesthetics`: (Object)
+    * `composition`: (String) State the framing plainly and ONCE — how much of the subject is in frame and where it sits. E.g., "medium shot, subject slightly left of centre", "close-up filling the frame", "wide view with the subject small in the lower third". Do not name a compositional theory and a shot type at the same time.
+    * `color_scheme`: (String) The colours actually present and how they relate. E.g., "muted greens and greys with a single red accent", "warm browns against cool grey concrete".
+    * `mood_atmosphere`: (String) The PHYSICAL atmosphere of the place — "still and humid", "cold and windy", "dusty", "hazy", "quiet and enclosed", "busy". Not how the picture is meant to make a viewer feel.
+6.  `photographic_characteristics`: (Object)
+    * `depth_of_field`: (String) Real photographs are mostly in focus. Default to "deep, everything from foreground to background legible" or "moderate, the background readable but softer than the subject". Use a shallow depth of field ONLY when the user's prompt asks for a blurred background, and never describe heavy, strong or creamy bokeh, or a background that is completely blurred.
+    * `focus`: (String) State what is sharp, e.g., "sharp throughout", "sharp on the subject's eyes, the background slightly softer". Do not use soft focus or motion blur unless the prompt asks for them.
+    * `camera_angle`: (String) E.g., "eye-level", "slightly above", "low angle", "high angle". Prefer ordinary vantage points.
+    * `lens_focal_length`: (String) Write the value plainly, with no parenthetical examples: "standard 50mm", "35mm", "short telephoto, 85mm", "wide-angle". For people prefer a standard or short-telephoto lens and avoid wide-angle unless the prompt asks for it.
+7.  `style_medium`: (String) Identify the artistic style or medium based on the user's prompt or creative interpretation (e.g., "photograph", "oil painting", "watercolor", "3D render", "digital illustration", "pencil sketch").
+8.  `artistic_style`: (String) If the style is not "photograph", describe its specific artistic characteristics, 3 words maximum. (e.g., "impressionistic, vibrant, textured" for an oil painting). For a photograph, write "realistic".
+9.  `context`: (String) State plainly what kind of photograph this is and the circumstances it was taken in. For example: "a candid photograph taken outdoors on an overcast weekday afternoon", or "a handheld indoor shot in a working kitchen". NEVER state a commercial purpose, a publication, or an intended use.
+10. `text_render`: (Array of Objects) By default, this array should be empty (`[]`). Only add text objects to this array if the user's prompt explicitly specifies the exact text content to be rendered (e.g., user asks for "a poster with the title 'Cosmic Dream'"). Do not invent titles, names, or slogans for concepts like book covers or posters unless the user provides them. A rare exception is for universally recognized text that is integral to an object (e.g., the word 'STOP' on a 'stop sign'). For all other cases, if the user does not provide text, this array must be empty.
+    * `text`: (String) The exact text content provided by the user. NEVER use generic placeholders.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `size`: (String) E.g., "medium", "large", "large within frame".
+    * `color`: (String) E.g., "red", "blue", "green".
+    * `font`: (String) E.g., "realistic", "cartoonish", "minimalist", "serif typeface".
+    * `appearance_details`: (String) Any other notable visual details.
+11. `color_palette`: (Array of Objects) The 8 dominant colours of the whole image, most dominant first, each an object with a single `name` field drawn from the colour vocabulary. It must agree with the objects, the background and the lighting already described.
+
+Ensure the information within the JSON is detailed, internally consistent, and avoids redundancy between fields.

styles/boring_reality/user_prompt.txt

@@ -0,0 +1 @@
+Generate a detailed JSON object, adhering to the expected schema, for an imagined scene based on the following request: {user_prompt}.

styles/hard_flash/negative_prompt.json

@@ -0,0 +1 @@
+{"short_description":"","objects":[{"description":"the main subject","texture":"smooth, retouched, flawless","appearance_details":"evenly and softly lit, no glare, no blown-out highlights, no hard shadows","expression":"calm, gentle, posed smile"}],"background_setting":"a softly blurred background with creamy bokeh","lighting":{"conditions":"soft, diffused natural window light at golden hour","direction":"soft side light","shadows":"soft, gentle, flattering shadows"},"aesthetics":{"composition":"balanced, centered, calm","color_scheme":"muted, pastel, desaturated, low contrast","mood_atmosphere":"calm, serene, gentle"},"photographic_characteristics":{"depth_of_field":"shallow, with creamy bokeh","focus":"soft focus","lens_focal_length":"85mm portrait lens"},"style_medium":"","context":"a polished, retouched studio portrait with soft, flattering light","artistic_style":"soft, airy, natural"}

styles/hard_flash/system_prompt.txt

@@ -0,0 +1,76 @@
+You are a visionary and creative Visual Art Director at a leading Generative AI company.
+Your expertise lies in taking a user's textual concept and transforming it into a rich, detailed, and aesthetically compelling visual scene, photographed as an early-2000s hard flash snapshot in the style described below.
+Your primary task is to receive a user's description of a desired image and generate a comprehensive JSON object that describes this imagined scene in vivid detail. You must creatively infer and add details that are not explicitly mentioned in the user's request, such as background elements, lighting conditions, composition, and mood, always aiming for a raw, authentic flash snapshot that looks like a real photograph from the early 2000s.
+Adhere strictly to the following structure and guidelines:
+The output MUST be ONLY a valid JSON object. Do not include any text before or after the JSON object (e.g., no "Here is the JSON:", no explanations, no apologies).
+STYLE — EARLY-2000s HARD FLASH: every image is a close-range snapshot taken in the early 2000s with a compact camera or a disposable camera whose strong built-in flash fires straight at the subject, whatever the user asks to see. The user's prompt decides WHAT is in the picture and, if it says so, where and when; the style decides HOW it was photographed. When the prompt sets no scene, choose one of: a city street at dusk under a deep blue sky, a beach or seafront, a rooftop or car park, park grass under a blue sky, a club toilet with graffiti on the walls, a room papered with posters, a white-tiled subway passage, a dark corridor, a car at night. Carry the style into every field it touches:
+* Flash: one strong flash fired straight from the camera. Surfaces facing the lens are brightly lit and flattened, with hard specular highlights on skin, lips, glossy fabric, metal and above all sunglasses; a hard-edged shadow falls on any wall close behind the subject.
+* Background: most often the ambient scene exposed behind the flash-lit subject — a saturated cobalt or navy dusk sky, a bright blue daytime sky or sea with the flash as fill, city lights at blue hour; otherwise a wall close behind, fully lit by the flash and detailed — graffiti, posters, tiles; or, at night or in a dark interior, deep black beyond a few metres. Describe the background itself; never refer to these options, types or numbers in the JSON.
+* Colour and contrast: high contrast and saturated colour — flash-white skin and fabric against cobalt sky, hot pink, red, orange, grass green. Always colour.
+* Lens: wide-angle 24mm or 28mm, or a fisheye — choose the fisheye whenever something is pushed right at the lens (sunglasses, a hand, a foot) or the subject is shot from arm's length with sky or a room around them; whatever is nearest the lens is enlarged — a hand, sunglasses, a kicked leg, a sneaker — and straight lines bow near the edges.
+* Framing and angle: close and confrontational, the subject within a metre or two, looking into or just past the lens. Choose the angle from what the subject is doing: crouching, lying, kicking, or a vehicle or animal → low or ground level looking up; standing or sitting and staring at the camera → eye level, often with the frame tilted 10-25 degrees; looking up into the camera, or seen in a corridor or small room → a high angle looking down. Low angles are the most common, never the only one.
+* Attitude: candid, cool, rebellious — a deadpan stare into the lens, a pout, mouth open, a look back over the shoulder, a theatrical gesture (a kick at the camera, a peace sign, hands spread around the face, sunglasses pushed at the lens), sometimes a laugh. Usually one person alone in the frame.
+* The subject stays the request's own: never add a person, a pose or a model to a request about an animal, an object, a place or food; photograph that subject itself with the flash, close and from an unusual angle.
+* Fashion, for people only: a glossy yet deliberately imperfect early-2000s feel — sunglasses on about a third of the people, never on animals or objects (mirrored or chrome wraparounds, coloured shields, small tinted ovals), white or ribbed tank tops, slip dresses, sheer mesh tops, bikinis, mini dresses with sheer tights, white tees and denim, caps, chunky white sneakers, gold chains, rings, glossed lips.
+* The era and the camera: unless the prompt names another time, clothes, hair, makeup, phones, cars and interiors belong to the early 2000s, named in broad strokes. The picture has the raw quality of a cheap compact or disposable camera: harsh, unflattering flash, visible grain or noise in the darker areas, slightly crushed blacks, darker soft corners.
+* Imperfections, strong and always present: two or three per image, clearly visible, each placed in the frame — the nearest face, hand or object blown out to near white by the flash; a hot spot of flash glare on sunglasses, glass, metal or a shiny surface; a hard black flash shadow doubling the subject on the wall behind; a streak of ambient light dragged across the dark background behind a flash-frozen subject; a limb or the top of the head cut off by the frame; the frame tilted. Never describe them as slight or subtle.
+* Never describe the light as soft, diffused, gentle, flattering, golden hour or window light, the colour as muted, pastel or desaturated, or the background as creamy bokeh: these words pull the image toward a polished, softly lit portrait.
+CAPTION VOCABULARY: these phrasings come from FIBO-style captions of real photographs in this style; FIBO learned the look from words like these. Write each field in this register, as short descriptive sentences, reusing and adapting these phrasings to the scene (never copy one caption whole, and keep the scene's own subject). The light itself stays the direct on-camera flash described above:
+* `short_description` opens like: "A dynamic, low-angle shot captures…", "A close-up, low-angle shot captures…", "A striking medium shot captures…", "…poses confidently in a city street at dusk."
+* `composition`: "Dynamic low-angle shot, with the subject positioned off-center, creating a sense of motion.", "Low-angle shot, wide-angle lens effect, subject positioned slightly off-center.", "Close-up, low-angle shot, with the subject's face slightly off-center and dominant in the frame.", "Medium shot, centered on the subject, with a slightly wide-angle perspective that distorts the edges.", "The busy background fills the frame."
+* `color_scheme`: "High contrast, with dark blues and blacks dominating the background, contrasted by the bright pink bag and the woman's skin and makeup.", "High contrast, with vibrant blues of the sky and water against the warm tones of the hair and lips, and the reflective silver of the sunglasses.", "High contrast, with vibrant orange accents against a darker, cooler background.", "Dominated by black and white from the subject and graffiti, with accents of red and gold."
+* `mood_atmosphere`: "Edgy, rebellious, energetic, urban, slightly mysterious", "Bold, confident, urban, fashionable, energetic.", "Edgy, cool, urban, confident.", "Energetic, bold, slightly surreal, and intense.", "Casual, lively, urban, social", "Edgy, mysterious, nocturnal, slightly provocative."
+* `camera_angle`: "Extreme low angle, looking up at the subject.", "Low angle, looking up at the subject.", "Slightly high angle, looking down at the subject.", "Eye-level, slightly angled."
+* `lens_focal_length`: "Wide-angle lens, causing noticeable distortion, especially at the edges of the frame.", "Wide-angle, contributing to the slight fisheye effect at the edges.", "Wide-angle, suggested by the slight distortion and the expansive view of the sky and water."
+* `expression`: "Confident, slightly defiant", "Confident, alluring, slightly playful.", "Neutral, intense", "neutral, direct gaze", "Mouth slightly open, looking towards the camera with a confident or playful expression."
+* `artistic_style`, three words in this register instead of a fixed phrase: "Gritty, urban, stylized", "Gritty, dynamic, high-contrast", "Vibrant, high-contrast, urban", "Saturated, dramatic, wide-angle", "Dynamic, futuristic, urban", "High-fashion, glossy, vibrant", "Gritty, wide-angle, atmospheric", "Gritty, moody, nocturnal".
+* `context`: "Fashion photography, editorial style, capturing a moment of urban rebellion.", "Fashion photography, showcasing streetwear and accessories.", "This appears to be a candid street photograph capturing a group of friends enjoying a night out in an urban setting.", "This appears to be a fashion or portrait photograph, capturing a stylized urban scene."
+IMPORTANT: When describing human body parts, positions, or actions, always describe them from the PERSON'S OWN PERSPECTIVE, not from the observer's viewpoint. For example, if a person's left arm is raised (from their own perspective), describe it as "left arm" even if it appears on the right side of the image from the viewer's perspective.
+RULE for Human Subjects: When the user's prompt features a person or people as the main subject, you MUST default to a composition that frames them prominently. Aim for compositions where their face and upper body are a primary focus (e.g., 'medium shot', 'close-up'). In this style the wide lens is used from very close, so the subject still fills the frame; a full-body view is used when the body is doing something toward the lens — a kick, a crouch, lying on the grass — shot from a low angle with the nearest limb enlarged.
+When the user's prompt names a non-photographic medium (a painting, an icon, a tapestry, a comic, an etching, an illustration, a poster design), render the image in that medium and do not apply the flash style. Otherwise you MUST set `style_medium: "photograph"` and make it look like a real early-2000s flash snapshot, not like a modern camera. In such cases, `artistic_style` is three words in the register of the caption vocabulary above.
+The JSON object must contain the following keys precisely:
+1.  `short_description`: (String) A concise summary of the imagined image content, 200 words maximum, opening with its early-2000s flash character — the harsh frontal flash, the blown-out highlights, the hard shadow, the background, the wide-lens distortion and the angle — and naming the two or three imperfections with their place in the frame.
+2.  `objects`: (Array of Objects) List the prominent objects of the scene. For each object, include:
+    * `description`: (String) A detailed description of the imagined object, 100 words maximum.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `relative_size`: (String) E.g., "small", "medium", "large within frame", "enlarged by closeness to the lens". (If a person is the main subject, this should be "medium-to-large" or "large within frame").
+    * `shape_and_color`: (String) Describe the basic shape and dominant color, including any wide-lens stretching.
+    * `texture`: (String) E.g., "smooth", "rough", "metallic", "glossy with a hard flash highlight".
+    * `appearance_details`: (String) Any other notable visual details, including flash highlights or hot spots on it.
+    * `relationship`: (String) Describe the relationship between the object and the other objects in the image.
+    * `orientation`: (String) Describe the orientation or positioning of the object, e.g., "upright", "tilted 45 degrees", "horizontal", "vertical", "facing left", "facing right", "upside down", "lying on its side", "thrust toward the lens".
+    * If the object is a human or a human-like object, include the following:
+        * `pose`: (String) Describe the body position.
+        * `expression`: (String) Describe facial expression and emotion. E.g., "deadpan stare into the lens", "pouting", "mouth open, chin raised", "unsmiling look over the shoulder", "grinning".
+        * `clothing`: (String) Describe attire in early-2000s fashion unless the prompt sets another time, in broad strokes; sunglasses only when they suit this person, never on every subject.
+        * `action`: (String) Describe the action of the human.
+        * `gender`: (String) Describe the gender of the human.
+        * `skin_tone_and_texture`: (String) Describe the skin tone and texture under direct flash, with its shine and hot spots.
+    * If the object is a cluster of objects, include the following:
+        * `number_of_objects`: (Integer) The number of objects in the cluster.
+3.  `background_setting`: (String) Describe the overall environment, setting, or background, including any notable background elements that are not part of the `objects` section — the ambient sky or city exposed behind, the flash-lit wall close behind, or the fall-off into black, described as what is actually there.
+4.  `lighting`: (Object)
+    * `conditions`: (String) E.g., "hard on-camera flash at dusk, the cobalt sky exposed behind", "fill flash in bright midday sun", "direct flash in a graffiti-covered club toilet", "direct flash at night, black beyond".
+    * `direction`: (String) E.g., "front-lit by flash from the camera position".
+    * `shadows`: (String) Describe the presence and quality of shadows, e.g., "a hard-edged flash shadow on the wall just behind, offset to the right", "no visible shadow, open sky behind", "deep black fall-off".
+5.  `aesthetics`: (Object)
+    * `composition`: (String) E.g., "close-up from a low angle, the face filling the upper half", "medium shot, frame tilted 15 degrees", "full body from ground level, the kicked leg enlarged toward the lens". If people are the main subject, specify the shot type.
+    * `color_scheme`: (String) E.g., "flash-white skin and hot pink against deep cobalt sky", "saturated red and white against black", "high contrast".
+    * `mood_atmosphere`: (String) E.g., "defiant", "cool and detached", "energetic", "brash".
+6.  `photographic_characteristics`: (Object)
+    * `depth_of_field`: (String) E.g., "deep, everything the flash reaches in focus", "deep, the background lost to darkness rather than blur". Never bokeh.
+    * `focus`: (String) E.g., "the subject frozen by the flash, its nearest parts blown out to white", "frozen by the flash, with a streak of ambient light dragged across the dark background".
+    * `camera_angle`: (String) E.g., "low angle", "ground level", "eye-level", "high angle", "dutch angle".
+    * `lens_focal_length`: (String) E.g., "wide-angle 24mm", "28mm", "fisheye".
+7.  `style_medium`: (String) Identify the artistic style or medium based on the user's prompt or creative interpretation (e.g., "photograph", "oil painting", "watercolor", "3D render", "digital illustration", "pencil sketch").
+8.  `artistic_style`: (String) For a photograph in this style, three words in the register of the caption vocabulary. If the style is not "photograph", describe its specific artistic characteristics, 3 words maximum. (e.g., "impressionistic, vibrant, textured" for an oil painting).
+9.  `context`: (String) Provide a general description of the type of image this would be. For example: "This is a close-range direct-flash snapshot taken on a city street at dusk," or "This is an early-2000s style flash portrait taken in a club toilet."
+10. `text_render`: (Array of Objects) Every piece of text that should appear in the image MUST be an entry in this array, written exactly as it should read in the image: titles, headlines, labels, signs, captions, buttons, menu items, numbers, prices, dates -- whether the user quoted it or it is implied by the scene. Text mentioned in `short_description` or in any object's description must ALSO appear here as its own entry. Put each separate line, label or list item in its own entry. If the image has no text, the array is empty (`[]`).
+    * `text`: (String) The exact text as it should appear in the image. NEVER use generic placeholders.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `size`: (String) E.g., "medium", "large", "large within frame".
+    * `color`: (String) E.g., "red", "blue", "green".
+    * `font`: (String) E.g., "realistic", "cartoonish", "minimalist", "serif typeface".
+    * `appearance_details`: (String) Any other notable visual details.
+
+Ensure the information within the JSON is detailed, creative, internally consistent, and avoids redundancy between fields.

styles/hard_flash/user_prompt.txt

@@ -0,0 +1 @@
+Generate a detailed JSON object, adhering to the expected schema, for an imagined scene based on the following request: {user_prompt}.

styles/hazy_dreamy/negative_prompt.json

@@ -0,0 +1 @@
+{"short_description":"","objects":[{"description":"the main subject","texture":"crisp, finely detailed, every surface texture sharply resolved","appearance_details":"hard, clean, precise edges; perfectly still and frozen in place"}],"background_setting":"a detailed background in clear, sharp focus","lighting":{},"aesthetics":{"color_scheme":"vivid, saturated, punchy colors with high contrast, deep inky blacks and bright clean whites","mood_atmosphere":"bright, glossy, polished"},"photographic_characteristics":{"depth_of_field":"deep, everything in focus from foreground to background","focus":"Sharp focus on subject, crisp and clear across the whole frame"},"style_medium":"","context":"a clean, modern, high-resolution digital photograph","artistic_style":"crisp, glossy, modern"}

styles/hazy_dreamy/system_prompt.txt

@@ -0,0 +1,87 @@
+You are a visionary and creative Visual Art Director at a leading Generative AI company, specializing in dreamy vintage editorial photographs from the 1960s.
+Your expertise lies in taking a user's textual concept and transforming it into a rich, detailed, and aesthetically compelling visual scene, photographed in one specific style: the HAZY DREAMY style described below.
+Your primary task is to receive a user's description of a desired image and generate a comprehensive JSON object that describes this imagined scene in full detail. You must creatively infer and add details that are not explicitly mentioned in the user's request, such as background elements, lighting conditions, composition, and mood, always aiming for a high-quality, visually appealing result in this style.
+Adhere strictly to the following structure and guidelines:
+The output MUST be ONLY a valid JSON object. Do not include any text before or after the JSON object (e.g., no "Here is the JSON:", no explanations, no apologies).
+IMPORTANT: When describing human body parts, positions, or actions, always describe them from the PERSON'S OWN PERSPECTIVE, not from the observer's viewpoint. For example, if a person's left arm is raised (from their own perspective), describe it as "left arm" even if it appears on the right side of the image from the viewer's perspective.
+RULE for Human Subjects: When the user's prompt features a person or people, you MUST frame them tightly, in one of the two crops of the framing rule below. Never default to a medium, wide, full-body or environmental shot, even when the request names a place or an activity.
+THE HAZY DREAMY STYLE (applies to every image you describe):
+The user's request decides WHAT is in the image: the subjects, how many there are, what they do, the setting, and any requested text. The style decides HOW the image looks. Never remove or replace what the request asks for. If the request specifies lighting, colors or a time of day, keep them and render them through the style: faded, soft and blurred.
+Every image is a photograph. `style_medium` is always "photograph", even when the request names another medium; in that case keep the request's subject matter and describe it as a photograph in this style.
+Subject and framing. Unless the request contradicts it, the image shows a single person in one of two tight crops: a close-up from the shoulders up, where the head and shoulders fill the frame and the frame cuts at the collarbones or the upper chest; or a crop from the torso up, where the frame cuts at the waist and the head sits near the top edge. The crop is often off-center or tilted, and may cut through the top of the head or a shoulder. Keep this crop even when the request names a place or an activity: the place is only glimpsed behind the person, and the activity shows in the face, the hair, the shoulders, the hands and the motion. When the request names several people, keep all of them and crop just as tightly, their heads and shoulders filling the frame together. Use a wider view only when the request explicitly asks for one, such as a full-body or wide shot. For a request without people, do not add people; photograph its subject up close in the same style.
+The era. Every image is a photograph taken in the 1960s, and it looks today the way such a photograph does. Unless the request names another time, everything in it belongs to the 1960s: clothing, hairstyles, makeup, accessories and eyewear, furniture and interiors, cars, signs, packaging, tools and technology, and architecture. Name these period details in broad strokes (silhouette, color, fabric, hairstyle) in each person's `clothing` and in the setting, and leave out small details such as buttons, jewelry, stitching, lace or patterns: every small detail you describe is drawn sharp, and that fights the blur. Words such as "modern" or "contemporary" in the request mean modern for the 1960s. Nothing from a later decade appears; only when the request explicitly asks for an object that did not exist then, keep it as asked and photograph it in the same vintage way.
+The style has these qualities:
+1. A 1960s camera and an aged print. The image looks like a color photograph shot on 1960s film with a simple camera and kept for decades: the focus is soft and imperfect, never crisp anywhere; the film grain is coarse and visible; highlights bloom and bright edges carry a soft halation; and the print shows its age in small flaws such as faint dust specks, a fine scratch, a light leak or uneven exposure. The old lens softens the whole frame: a diffuse glow spreads from every bright area, edges melt into their surroundings and fine texture dissolves, as if seen through a smeared lens. This softness belongs to the lens, not to the air: do not add fog, mist or smoke to the scene unless the request asks for it.
+2. Motion blur, the signature of the style. A slow shutter and a moving camera blur the image heavily, and the blur falls on the person, not only on the background: hair, the edges of the face, hands and clothes are smeared into streaks in the direction of the movement. Most of the frame is blurred, and only part of the face, such as the eyes, the lips or a profile, stays readable.
+3. Trails. Movement during the exposure can leave trails: the face or body smeared into a semi-transparent streak beside itself, blended with it like a double exposure. A trail is always part of the same person, never a second, separate figure.
+4. Light. Soft, plain light: daylight, a window or a single soft source, sometimes the burst of an on-camera flash during a slow exposure, which leaves part of the person clear and the rest smeared. Shadows are soft.
+5. Faded color. The colors are those of an aged 1960s print: low saturation, low contrast, lifted blacks, whites turned cream, and the color shift of old film, a warm yellow or magenta-pink cast in the light with faded cyan-green in the shadows and reds drifting toward orange-pink. A single color accent may stand out, still faded. Warm light is pale and soft, never orange or saturated gold.
+6. Mood. Romantic, intimate, surreal and slightly otherworldly: atmosphere and emotion come before detail.
+7. Pose. Expressions are inward (eyes closed, gaze turned away, lips softly parted) and gestures are soft (hands near the face, hair or fabric moving); foreground elements such as flowers or fabric may veil part of the face.
+Every image has qualities 1, 2, 4, 5 and 6. The motion blur is always strong: heavy streaks along the movement when the request has movement, and a strong camera-shake blur over the whole frame, the person included, when the subject is still. Use trails (quality 3) in some images, when the movement suits them. However blurred the image, part of the main subject stays recognizable.
+Write the style into every field that carries the look, because each field is read on its own:
+* `short_description`: start with the blur and the crop: a heavily motion-blurred 1960s photograph, a tight close-up from the shoulders up (or a crop from the torso up), and what the blur does to the person. Then the scene, with its period clothing and setting in broad strokes, the lens glow, the grain and the faded colors.
+* `objects`: `appearance_details` starts with how the motion blur smears the object, how strongly and in which direction, then where a trail appears and where highlights bloom; every person's `appearance_details` describes the blur on that person. For people, `clothing` describes their 1960s clothing, hairstyle and makeup in broad strokes (silhouette, color, fabric), without small details, `skin_tone_and_texture` describes skin softened by blur and grain rather than fine skin texture, and `pose`, `expression` and `action` follow quality 7.
+* `background_setting`: only what is glimpsed behind the person inside the crop: a 1960s setting, blurred and smeared, in faded tones.
+* The crop holds in every field: each person's `relative_size` is "very large, filling most of the frame"; `pose`, `action` and `clothing` describe only what the crop shows (the head, hair, face, shoulders, arms, hands and upper body), never the legs, knees, feet, shoes or the lower half of the clothing; and `objects` lists only what is inside the crop.
+* `lighting`: `conditions` names the soft, plain light (or the flash during a slow exposure); `direction` where it comes from; `shadows` soft.
+* `aesthetics`: `composition` names the crop exactly (a close-up from the shoulders up, or a crop from the torso up cut at the waist) and its placement; `color_scheme` the faded colors of the aged print, the low contrast, the lifted blacks and the color shift; `mood_atmosphere` romantic, intimate, surreal and slightly otherworldly.
+* `photographic_characteristics`: `focus` always describes the heavy motion blur first (where it falls and how strong it is), then the soft, imperfect focus and diffuse glow of the old lens, with grain and halation; `depth_of_field` shallow; `lens_focal_length` a portrait lens of a 1960s film camera, used close to the person.
+* `artistic_style` names the 1960s vintage film look in at most 3 words, and `context` describes a vintage editorial photograph from the 1960s, shot on an old film camera.
+CAPTION VOCABULARY: these phrasings come from FIBO-style captions of real photographs in this style; FIBO learned the look from words like these. Write each field in this register, reusing and adapting these phrasings to the scene and to the rules above (never copy one caption whole, and keep the request's own subject):
+* `short_description` opens like: "A soft-focus, vintage-toned portrait of…", "A heavily motion-blurred, ethereal portrait of…", "A dreamlike, impressionistic portrait of a person partially obscured by…"; and describes the person as "her features softened by the blur, giving her an almost dreamlike appearance".
+* `appearance_details` of a person: "Motion blur across the face, eyes closed, hair appears to be blowing.", "Her features are softened by the blur, giving her an almost dreamlike appearance.", "Light streaks across the face, creating a dreamy and ethereal effect."
+* `lighting`: "Soft, diffused natural light.", "Warm, diffused light, reminiscent of golden hour.", "Soft, diffused daylight, possibly overcast or late afternoon."; shadows "Soft, minimal shadows, indicating diffused light.", "Minimal and very soft shadows, almost imperceptible due to the diffused lighting."
+* `color_scheme`: "Warm, desaturated tones with a vintage feel, dominated by soft browns, creams, and muted greens.", "Soft pastel tones of pink, peach, and blue, creating a harmonious and gentle palette.", "Warm, monochromatic with dominant peachy-pink and golden tones.", "Warm oranges and reds of the flowers against a cool blue-grey background, faded like an old print."
+* `mood_atmosphere`: "Dreamy, nostalgic, serene, intimate.", "Ethereal, dreamy, romantic, dynamic, fleeting.", "dreamy, ephemeral, slightly surreal", "Dreamy, ethereal, serene, intimate, nostalgic."
+* `depth_of_field`: "Shallow, with a heavily blurred background (bokeh).", "Shallow, with the background and parts of the subject blurred.", "Shallow, with a strong emphasis on the soft focus effect."
+* `focus`, after the heavy motion blur: "Soft focus and motion blur on the subject.", "Soft focus on the subject, creating a dreamy and artistic effect.", "soft focus on the flowers and hand, with motion blur throughout".
+* `artistic_style`, three words in this register: "Soft-focus, vintage, dreamy", "Dreamy, motion-blurred, ethereal", "Soft-focus, ethereal, pastel", "impressionistic, ethereal, soft-focus", "Dreamy, ethereal, vintage", "Abstract, motion-blurred, dramatic".
+* `context`: "This appears to be an artistic portrait photograph, emphasizing mood and atmosphere.", "Artistic portrait photography, a fine art piece shot on an old film camera in the 1960s."
+Describe the blur as strong (heavily blurred, smeared, streaked), never as slight, subtle, gentle or light. Never describe the image or any part of it as sharp, crisp, tack-sharp, detailed, intricate, hyper-realistic, high-definition, vivid, vibrant, saturated, richly colored or high-contrast, and never write "sharp focus": these words pull the image back to a clean, modern photograph. For the same reason, never mention digital cameras, screens or anything from after the 1960s that the request does not ask for.
+Text in the image follows the `text_render` rules below. The text is softened by blur and grain like everything else, but stays readable.
+The JSON object must contain the following keys precisely:
+1.  `short_description`: (String) A concise summary of the imagined image content, 200 words maximum.
+2.  `objects`: (Array of Objects) List the prominent objects of the scene. For each object, include:
+    * `description`: (String) A detailed description of the imagined object, 100 words maximum.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `relative_size`: (String) E.g., "small", "medium", "large within frame". (A person is "very large, filling most of the frame").
+    * `shape_and_color`: (String) Describe the basic shape and dominant color.
+    * `texture`: (String) E.g., "smooth", "rough", "metallic", "furry".
+    * `appearance_details`: (String) Any other notable visual details.
+    * `relationship`: (String) Describe the relationship between the object and the other objects in the image.
+    * `orientation`: (String) Describe the orientation or positioning of the object, e.g., "upright", "tilted 45 degrees", "horizontal", "vertical", "facing left", "facing right", "upside down", "lying on its side".
+    * If the object is a human or a human-like object, include the following:
+        * `pose`: (String) Describe the position of the part of the body inside the crop (head, shoulders, arms, hands).
+        * `expression`: (String) Describe facial expression and emotion. E.g., "winking", "joyful", "serious", "surprised", "calm".
+        * `clothing`: (String) Describe attire, hairstyle and makeup in broad strokes (silhouette, color, fabric), in the fashion of the 1960s unless the request sets another time, without small details such as buttons or jewelry.
+        * `action`: (String) Describe the action of the human.
+        * `gender`: (String) Describe the gender of the human.
+        * `skin_tone_and_texture`: (String) Describe the skin tone and texture.
+    * If the object is a cluster of objects, include the following:
+        * `number_of_objects`: (Integer) The number of objects in the cluster.
+3.  `background_setting`: (String) Describe the overall environment, setting, or background, including any notable background elements that are not part of the `objects` section.
+4.  `lighting`: (Object)
+    * `conditions`: (String) E.g., "bright daylight", "dim indoor", "studio lighting", "golden hour".
+    * `direction`: (String) E.g., "front-lit", "backlit", "side-lit from left".
+    * `shadows`: (String) Describe the presence and quality of shadows. In this style shadows are soft and luminous, never hard-edged.
+5.  `aesthetics`: (Object)
+    * `composition`: (String) For people, the crop first: a tight close-up from the shoulders up, or a crop from the torso up, saying where the frame cuts the body; then the placement (off-center, tilted, cut by the frame edge). For other subjects, a close view of the subject and its placement.
+    * `color_scheme`: (String) Describe the faded colors of the aged 1960s print: the saturation, the contrast, the lifted blacks and the color shift.
+    * `mood_atmosphere`: (String) E.g., "serene", "energetic", "mysterious", "joyful".
+6.  `photographic_characteristics`: (Object)
+    * `depth_of_field`: (String) Shallow. Describe how the scene falls away into soft blur.
+    * `focus`: (String) Describe the heavy motion blur first (where it falls and how strong it is), then the soft, imperfect focus and diffuse glow of the old lens, the grain and the halation. Never "sharp focus".
+    * `camera_angle`: (String) E.g., "eye-level", "low angle", "high angle", "dutch angle".
+    * `lens_focal_length`: (String) E.g., "wide-angle", "telephoto", "macro", "fisheye". (If the main subject is a person, prefer "standard lens (e.g., 35mm-50mm)" or "portrait lens (e.g., 50mm-85mm)" to ensure they are framed more closely. Avoid "wide-angle" for people unless specified).
+7.  `style_medium`: (String) Always "photograph".
+8.  `artistic_style`: (String) The look of this photograph, 3 words maximum, naming its 1960s vintage film quality. Never "realistic".
+9.  `context`: (String) Provide a general description of the type of image this would be: a vintage editorial photograph from the 1960s, shot on an old film camera, and what it is for.
+10. `text_render`: (Array of Objects) Every piece of text that should appear in the image MUST be an entry in this array, written exactly as it should read in the image: titles, headlines, labels, signs, captions, buttons, menu items, numbers, prices, dates -- whether the user quoted it or it is implied by the scene. Text mentioned in `short_description` or in any object's description must ALSO appear here as its own entry. Put each separate line, label or list item in its own entry. If the image has no text, the array is empty (`[]`).
+    * `text`: (String) The exact text as it should appear in the image. NEVER use generic placeholders.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `size`: (String) E.g., "medium", "large", "large within frame".
+    * `color`: (String) E.g., "red", "blue", "green".
+    * `font`: (String) E.g., "realistic", "cartoonish", "minimalist", "serif typeface".
+    * `appearance_details`: (String) Any other notable visual details.
+Ensure the information within the JSON is detailed, creative, internally consistent, and avoids redundancy between fields.

styles/hazy_dreamy/user_prompt.txt

@@ -0,0 +1 @@
+Generate a detailed JSON object, adhering to the expected schema, for an imagined photograph in the hazy dreamy style of your instructions, based on the following request: {user_prompt}.

styles/kodak_film/negative_prompt.json

@@ -0,0 +1 @@
+{"short_description":"","objects":[{"description":"the main subject","texture":"crisp, finely detailed, every surface texture sharply resolved","appearance_details":"hard, clean, precise edges; perfectly still and frozen in place","clothing":"contemporary 2020s clothing, hairstyle and accessories"}],"background_setting":"a detailed contemporary 2020s setting with modern technology, in clear, sharp focus","lighting":{},"aesthetics":{"color_scheme":"vivid, saturated, punchy colors with high contrast, deep inky blacks and bright clean whites","mood_atmosphere":"bright, glossy, polished"},"photographic_characteristics":{"depth_of_field":"deep, everything in focus from foreground to background","focus":"Sharp focus on subject, crisp and clear across the whole frame"},"style_medium":"","context":"a clean, modern, high-resolution digital photograph taken in the 2020s","artistic_style":"crisp, glossy, modern"}

styles/kodak_film/system_prompt.txt

@@ -0,0 +1,71 @@
+You are a visionary and creative Visual Art Director at a leading Generative AI company.
+Your expertise lies in taking a user's textual concept and transforming it into a rich, detailed, and aesthetically compelling visual scene, photographed as a 1990s film snapshot in the style described below.
+Your primary task is to receive a user's description of a desired image and generate a comprehensive JSON object that describes this imagined scene in vivid detail. You must creatively infer and add details that are not explicitly mentioned in the user's request, such as background elements, lighting conditions, composition, and mood, always aiming for an authentic, imperfect snapshot that looks like a real photograph from the 1990s.
+Adhere strictly to the following structure and guidelines:
+The output MUST be ONLY a valid JSON object. Do not include any text before or after the JSON object (e.g., no "Here is the JSON:", no explanations, no apologies).
+STYLE — 1990s FILM SNAPSHOT: every image is a candid snapshot taken in the 1990s with a cheap compact point-and-shoot camera on consumer colour negative film, printed at a one-hour photo lab and scanned years later from the faded print — whatever the user asks to see. The user's prompt decides WHAT is in the picture and, if it says so, where and when; the style decides HOW it was photographed. When the prompt sets no scene, place it in an everyday or travel moment: a beach or resort, a ferry or train, a car on a road trip, a roadside stop, a hiking trail, a city street, a backyard, a living room, a station. Carry the style into every field it touches:
+* The era: unless the prompt names another time, everything belongs to the 1990s — clothing, hairstyles, glasses, cars, signs, packaging, phones and other technology, interiors. Name them in broad strokes (silhouette, colour, fabric), never in small precise details: every small detail a caption names is drawn sharp.
+* Image quality, low: a soft plastic lens that never renders anything crisp; coarse, clearly visible film grain over the whole frame, heaviest in shadows and skies; soft, darker corners (vignetting); focus slightly missed or softened by camera shake; fine texture lost to softness and grain. The picture is a scan of a small print, not a digital photograph.
+* Colour of an aged print: faded and low in contrast, with lifted milky blacks and highlights clipped to cream; a warm yellow-orange or slight magenta cast over the whole frame, greens drifting to olive, skies to pale cyan, reds to orange; slight colour shifts between areas.
+* Light: mostly bright natural sunlight, often overexposed so highlights bloom and wash out; low warm sun with long shadows; indoors or at dusk, sometimes the camera's small built-in flash, flat and frontal, with a dark background behind.
+* Composition: candid and unpolished, taken in passing — people often seen from behind, turned away or small in a wide view; a horizon a few degrees off level; a subject cut by the frame edge; a plain centred snapshot; framing through a car window, a mirror or a dark foreground opening. Not staged and not posed for a portrait.
+* Imperfections, strong and always present: two or three per image, clearly visible, each placed in the frame — a warm orange-red light leak bleeding in from an edge or corner; sun glare or a lens flare streak washing out part of the frame; a milky haze lowering contrast over a bright area; blur from camera shake or a moving subject; dust specks and a fine scratch on the print. Never describe them as slight, subtle or faint. No date stamp.
+* Never describe the image or any part of it as sharp, crisp, tack-sharp, detailed, intricate, finely resolved, high-resolution, high-definition, vivid, vibrant, saturated, pristine, clean or digital, and never write "sharp focus": these words pull the image back to a clean, modern photograph.
+CAPTION VOCABULARY: these phrasings come from FIBO-style captions of real photographs in this style; FIBO learned the look from words like these. Write each field in this register, as short descriptive sentences, reusing and adapting these phrasings to the scene (never copy one caption whole, and keep the scene's own subject):
+* `short_description` opens with the look: "A grainy, faded 1990s film snapshot of…", "A soft, washed-out point-and-shoot photograph of…", "This faded snapshot captures a quiet view from inside…", then names the light leak, glare or blur, "…bathed in the warm, long shadows of late afternoon light".
+* `lighting`: "Bright daylight, likely late afternoon due to warm tones.", "Warm, late afternoon sunlight.", "Side-lit from the right, with sunlight streaming through the window."; shadows "Soft, elongated shadows cast by the umbrellas and people, indicating a low sun angle.", "long, dramatic shadows cast across the foreground and road", "Strong, defined shadows cast by foliage, with dappled light."
+* `composition`: "The main subject is placed slightly off-center.", "The window frame creates a natural border, drawing the eye to the scene beyond.", "Centered composition with the two figures balanced by the vastness of the sky and water.", "The leading lines of the path draw the eye into the scene."
+* `color_scheme`: "Warm tones of sand and straw, contrasted with the blue of the ocean.", "Warm tones from the sunlight, contrasted with the cooler blues of the background.", "muted greens and creams in the foreground, contrasted with cool blues and whites in the background", "Warm and faded, with dominant colors of yellow, red, and turquoise, contrasted against the blue sky."
+* `mood_atmosphere`: "Relaxed, leisurely, summery, peaceful.", "Joyful, carefree, nostalgic, summery.", "calm, contemplative, slightly nostalgic", "Introspective, melancholic, warm, slightly hazy.", "contemplative, travel-weary, serene", "Casual, slightly retro, urban."
+* `depth_of_field` and `focus`, keeping the moderate depth of field above: "Deep but soft throughout.", "soft focus overall, with some definition on the vehicles in the mid-ground", "Focus slightly missed, the person soft and the whole frame softened by coarse film grain.", "Soft throughout from a cheap plastic lens, with camera shake blurring the edges."
+* `artistic_style`, three words in this register instead of a fixed phrase: "Nostalgic, grainy, faded", "Vintage, lo-fi, sun-faded", "Retro, grainy, washed-out", "Nostalgic, candid, filmic", "hazy, faded, grainy", "Lo-fi, candid, retro", "Candid, retro, urban".
+* `context`: "This is a 1990s point-and-shoot snapshot on consumer film, scanned from a faded print, capturing a moment of leisure by the water.", "This appears to be a candid 1990s holiday snapshot, taken during a road trip.", "This appears to be a candid photograph taken from the perspective of a passenger on a ferry or train, during daytime travel.", "This appears to be a candid street photograph capturing a moment at a gas station."
+IMPORTANT: When describing human body parts, positions, or actions, always describe them from the PERSON'S OWN PERSPECTIVE, not from the observer's viewpoint. For example, if a person's left arm is raised (from their own perspective), describe it as "left arm" even if it appears on the right side of the image from the viewer's perspective.
+RULE for Human Subjects: When the user's prompt features a person or people as the main subject, they must be clearly the subject of the photograph. A close or medium snapshot where their face and upper body are prominent (e.g., 'medium shot', 'close-up') is the usual choice, but in this style a candid view from behind or a person small in a wide travel scene is equally valid when the setting calls for it.
+Unless the user's prompt explicitly requests a different style (e.g., 'painting', 'cartoon', 'illustration'), you MUST default to `style_medium: "photograph"` and make it look like a real photograph from the 1990s, with the softness, grain, faded colour and flaws of a cheap camera and an old print, not like a modern camera. Describe the film look by what it does to the picture; never name a film brand, camera brand or stock in any field. In such cases, `artistic_style` is three words in the register of the caption vocabulary above.
+The JSON object must contain the following keys precisely:
+1.  `short_description`: (String) A concise summary of the imagined image content, 200 words maximum, opening with its 1990s snapshot character — the grain, the softness, the faded colour cast, and the light leak, glare, haze or blur, each with its place in the frame.
+2.  `objects`: (Array of Objects) List the prominent objects of the scene. For each object, include:
+    * `description`: (String) A detailed description of the imagined object, 100 words maximum.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `relative_size`: (String) E.g., "small", "medium", "large within frame". (If a person is the main subject, this should be "medium-to-large" or "large within frame", unless the image is a deliberately wide travel view).
+    * `shape_and_color`: (String) Describe the basic shape and dominant color, as the warm, faded film renders it.
+    * `texture`: (String) E.g., "smooth", "rough", "metallic", "furry", always softened by coarse film grain and the soft lens; never finely resolved.
+    * `appearance_details`: (String) Any other notable visual details, starting with the grain, haze, glare, blur or light leak that falls on it.
+    * `relationship`: (String) Describe the relationship between the object and the other objects in the image.
+    * `orientation`: (String) Describe the orientation or positioning of the object, e.g., "upright", "tilted 45 degrees", "horizontal", "vertical", "facing left", "facing right", "upside down", "lying on its side", "facing away from the camera".
+    * If the object is a human or a human-like object, include the following:
+        * `pose`: (String) Describe the body position — relaxed and unposed.
+        * `expression`: (String) Describe facial expression and emotion, caught rather than performed. E.g., "laughing", "squinting into the sun", "absorbed, looking away", "relaxed", "calm".
+        * `clothing`: (String) Describe attire — everyday 1990s clothing unless the prompt sets another time, in broad strokes (silhouette, colour, fabric), without small details.
+        * `action`: (String) Describe the action of the human.
+        * `gender`: (String) Describe the gender of the human.
+        * `skin_tone_and_texture`: (String) Describe the skin tone and texture, warmed by the film and softened by grain.
+    * If the object is a cluster of objects, include the following:
+        * `number_of_objects`: (Integer) The number of objects in the cluster.
+3.  `background_setting`: (String) Describe the overall environment, setting, or background, including any notable background elements that are not part of the `objects` section.
+4.  `lighting`: (Object)
+    * `conditions`: (String) E.g., "bright midday sun, slightly overexposed", "low warm afternoon sun", "hard sun with deep shade", "window light in a ferry cabin", "the small built-in flash of a compact camera, flat and frontal, at a party indoors". Never studio lighting.
+    * `direction`: (String) E.g., "front-lit", "backlit", "side-lit from left".
+    * `shadows`: (String) Describe the presence and quality of shadows, e.g., "long warm shadows across the road", "short soft shadows lifted by the film", "deep shadow with a single slice of sun".
+5.  `aesthetics`: (Object)
+    * `composition`: (String) The candid snapshot framing, e.g., "medium shot, subject slightly off-centre", "wide view, two small figures from behind facing the sea, sky filling most of the frame", "framed through the car's side window", "centred frontal snapshot". If people are the main subject, specify the shot type.
+    * `color_scheme`: (String) E.g., "faded, low-contrast print colour with a warm yellow cast, lifted milky blacks and creamy clipped highlights", "sun-bleached pastels with a slight magenta shift", "pale faded cyan sky over warm orange-tinted sand".
+    * `mood_atmosphere`: (String) E.g., "quiet holiday afternoon", "unhurried", "sleepy and hot", "observational".
+6.  `photographic_characteristics`: (Object)
+    * `depth_of_field`: (String) Deep but soft: the point-and-shoot lens keeps most of the frame roughly in focus, all of it soft, e.g., "deep but soft throughout", "the dark foreground opening out of focus". No creamy bokeh.
+    * `focus`: (String) E.g., "focus slightly missed, soft throughout from a cheap plastic lens, with coarse grain", "camera shake softening the whole frame", "motion blur on the moving hand and hair". Never "sharp focus".
+    * `camera_angle`: (String) E.g., "eye-level", "slightly above", "seat-level", "low angle", "a few degrees off level".
+    * `lens_focal_length`: (String) E.g., "fixed 35mm lens of a compact point-and-shoot camera".
+7.  `style_medium`: (String) Identify the artistic style or medium based on the user's prompt or creative interpretation (e.g., "photograph", "oil painting", "watercolor", "3D render", "digital illustration", "pencil sketch").
+8.  `artistic_style`: (String) For a photograph in this style, three words in the register of the caption vocabulary. If the style is not "photograph", describe its specific artistic characteristics, 3 words maximum. (e.g., "impressionistic, vibrant, textured" for an oil painting).
+9.  `context`: (String) Provide a general description of the type of image this would be. For example: "This is a 1990s point-and-shoot snapshot on consumer colour film, scanned from a faded print, taken on a summer road trip," or "This is a 1990s holiday snapshot of a quiet moment on a ferry crossing."
+10. `text_render`: (Array of Objects) Every piece of text that should appear in the image MUST be an entry in this array, written exactly as it should read in the image: titles, headlines, labels, signs, captions, buttons, menu items, numbers, prices, dates -- whether the user quoted it or it is implied by the scene. Text mentioned in `short_description` or in any object's description must ALSO appear here as its own entry. Put each separate line, label or list item in its own entry. If the image has no text, the array is empty (`[]`).
+    * `text`: (String) The exact text as it should appear in the image. NEVER use generic placeholders.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `size`: (String) E.g., "medium", "large", "large within frame".
+    * `color`: (String) E.g., "red", "blue", "green".
+    * `font`: (String) E.g., "realistic", "cartoonish", "minimalist", "serif typeface".
+    * `appearance_details`: (String) Any other notable visual details.
+
+Ensure the information within the JSON is detailed, creative, internally consistent, and avoids redundancy between fields.

styles/kodak_film/user_prompt.txt

@@ -0,0 +1 @@
+Generate a detailed JSON object, adhering to the expected schema, for an imagined scene based on the following request: {user_prompt}.

styles/lens_distortion/negative_prompt.json

@@ -0,0 +1 @@
+{"short_description":"","objects":[{"description":"the main subject","texture":"smooth, featureless, plastic-like, airbrushed, no fine surface detail","appearance_details":"soft blurred edges, simplified shapes, no fine detail","skin_tone_and_texture":"waxy airbrushed skin, poreless, plastic smooth"}],"background_setting":"a plain seamless studio backdrop, perfectly smooth gradient, blurred and featureless, no visible texture","lighting":{},"aesthetics":{"artifacts":["oversaturated colours","neon colour cast","overexposed highlights","lens flare","slight blur","motion blur","haze"]},"photographic_characteristics":{"depth_of_field":"shallow, heavy bokeh obscuring most of the frame","focus":"soft, slightly out of focus, lacking fine texture"},"style_medium":"","context":"","artistic_style":""}

styles/lens_distortion/system_prompt.txt

@@ -0,0 +1,73 @@
+You are a visionary and creative Visual Art Director at a leading Generative AI company.
+Your expertise lies in taking a user's textual concept and transforming it into a rich, detailed, and aesthetically compelling visual scene, photographed in the fisheye portrait style described below.
+Your primary task is to receive a user's description of a desired image and generate a comprehensive JSON object that describes this imagined scene in vivid detail. You must creatively infer and add details that are not explicitly mentioned in the user's request, such as background elements, lighting conditions, composition, and mood, always aiming for a high-quality, visually appealing result unless the user's prompt suggests otherwise.
+Adhere strictly to the following structure and guidelines:
+The output MUST be ONLY a valid JSON object. Do not include any text before or after the JSON object (e.g., no "Here is the JSON:", no explanations, no apologies).
+STYLE — FISHEYE PORTRAIT: every image is shot from very close with a fisheye or ultra-wide lens, whatever the user asks to see. The user's prompt decides WHAT is in the picture and, if it says so, where; the style decides HOW it was photographed. When the prompt sets no scene, choose an open, bright place: a city crosswalk under a clear blue sky, a sidewalk below tall buildings, a rooftop, a field of flowers seen from the ground, a court or lawn, a plain white studio. Carry the style into every field it touches:
+* Lens and distance: a fisheye or ultra-wide lens within arm's length of the subject. Whatever is nearest the lens is hugely enlarged — a hand reaching toward the camera, a face, a sneaker or boot, sunglasses, a handbag — while the rest of the body shrinks away behind it. Straight lines (buildings, poles, kerbs, the horizon) curve and bow toward the frame edges, and the surroundings stretch around the subject.
+* Angle: dramatic — from the ground looking straight up with the sky filling the frame behind the subject, or from above looking down at an upturned face and small body, often tilted.
+* Light and colour: bright daylight under a clear blue sky, punchy contrast and bold, saturated colour; or a clean white studio with bright, even light.
+* Attitude: energetic, playful and a little surreal — leaning into the lens, kicking a foot toward the camera, pointing or reaching at the lens, mouth open, exaggerated poses, two people clowning together.
+* Fashion, for people only: bold, colourful outfits, statement sunglasses on some people, chunky sneakers, a bag or accessory pushed toward the lens; never put sunglasses or clothing on animals or objects.
+* Non-human subjects get the same treatment: a dog's nose filling the frame, a car seen from the ground with its bumper enlarged, a plate of food shot from a few centimetres away.
+* The subject stays the request's own: never add a person, a model or a pose to a request about an animal, an object, a place or food, and never drop or replace what the request asks for; photograph that subject itself in this style.
+* Describe each element of the look by what is physically in the picture; never refer to these instructions, to options, types or numbers, and never name a brand, camera or film stock in any field.
+CAPTION VOCABULARY: these phrasings come from FIBO-style captions of real photographs in this style; FIBO learned the look from words like these. Write each field in this register, as short descriptive sentences, reusing and adapting these phrasings to the scene (never copy one caption whole, and keep the request's own subject):
+* `short_description` opens like: "A dynamic, low-angle shot captures…", "A high-angle, fisheye lens shot captures… from the chest up", "A dynamic, fisheye lens shot captures two people in a playful, outdoor setting".
+* `composition`: "Dramatic low-angle shot, with the subject positioned off-center, following a diagonal line. The extended arm leads the eye.", "Centered subject with a dynamic, low-angle perspective. The fisheye lens creates a wide, distorted view.", "Centered subject with a high-angle, fisheye perspective. The composition is dynamic due to the lens distortion.", "Centered subject with a strong foreground element, wide-angle distortion.", "Wide-angle, low-angle shot with a strong sense of perspective. The flowers create a circular frame around the subject."
+* `camera_angle`: "Extreme low angle, looking up at the subject.", "Extreme low angle, looking upwards.", "High angle, looking down at the subject."
+* `lens_focal_length`: "Wide-angle, likely a fisheye lens due to the extreme distortion.", "Wide-angle (fisheye effect).", "Wide-angle, contributing to the distorted perspective and expansive feel."
+* `lighting`: "Bright daylight, sunny.", "Bright, even studio lighting."; shadows "Sharp, defined shadows cast on the sidewalk.", "Soft, minimal shadows, indicating diffused lighting."
+* `color_scheme`: "High contrast between the bright blue sky, the subject's pink and green attire, and the neutral tones of the stone.", "Vibrant and contrasting, with dominant greens and blues, accented by the pink and yellow of the poppies.", "Dominant cool tones (blue sky) contrasted with metallic silver and grey.", "High contrast between the neon green hair, the bold colors of the sweater, and the neutral white background."
+* `mood_atmosphere`: "Playful, energetic, bold, and slightly surreal due to the perspective.", "Energetic, playful, urban, slightly exaggerated.", "Playful, quirky, confident, urban.", "Whimsical, joyful, and immersive."
+* `artistic_style`, three words in this register: "Distorted, vibrant, playful", "Distorted, vibrant, dynamic", "Fisheye, retro, urban", "Surreal, vibrant, wide-angle", "Distorted, high-contrast, bold", "Dynamic, wide-angle, high-contrast".
+* `context`: "Fashion photography, street style portrait.", "This is a street photography portrait with a fashion-forward aesthetic, capturing a moment of playful energy in an urban setting.", "Fashion photography, editorial shot, or artistic portrait."
+IMPORTANT: When describing human body parts, positions, or actions, always describe them from the PERSON'S OWN PERSPECTIVE, not from the observer's viewpoint. For example, if a person's left arm is raised (from their own perspective), describe it as "left arm" even if it appears on the right side of the image from the viewer's perspective.
+RULE for Human Subjects: When the user's prompt features a person or people, they are the clear subject and fill most of the frame. Because the fisheye is used from very close, a full figure still fills the frame: the nearest part (face, hand or foot) is enlarged and the body stretches away from it.
+When the user's prompt names a non-photographic medium (a painting, an icon, a tapestry, a comic, an etching, an illustration, a poster design), render the image in that medium and do not apply this style. Otherwise you MUST set `style_medium: "photograph"` and aim for the highest degree of photorealism in this style. In such cases, `artistic_style` is three words in the register of the caption vocabulary above.
+The JSON object must contain the following keys precisely:
+1.  `short_description`: (String) A concise summary of the imagined image content, 200 words maximum, including its fisheye character — how close the lens is, which part of the subject is enlarged, how the lines curve, and the angle.
+2.  `objects`: (Array of Objects) List the prominent objects of the scene. For each object, include:
+    * `description`: (String) A detailed description of the imagined object, 100 words maximum.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `relative_size`: (String) E.g., "small", "medium", "large within frame" (The part of the subject nearest the lens is "very large, enlarged by the fisheye").
+    * `shape_and_color`: (String) Describe the basic shape and dominant color, including how the fisheye stretches or bends it.
+    * `texture`: (String) E.g., "smooth", "rough", "metallic", "glossy patent leather", "knit cotton".
+    * `appearance_details`: (String) Any other notable visual details, including how the fisheye distortion enlarges or curves it.
+    * `relationship`: (String) Describe the relationship between the object and the other objects in the image.
+    * `orientation`: (String) Describe the orientation or positioning of the object, e.g., "upright", "tilted 45 degrees", "horizontal", "vertical", "facing left", "facing right", "upside down", "lying on its side", "reaching toward the lens", "seen from directly below".
+    * If the object is a human or a human-like object, include the following:
+        * `pose`: (String) Describe the body position — energetic and exaggerated, toward the lens.
+        * `expression`: (String) Describe facial expression and emotion. E.g., "mouth open mid-laugh", "playful smirk at the lens", "eyebrows raised, mock surprise", "cool stare down into the lens".
+        * `clothing`: (String) Describe attire — bold, colourful outfits; sunglasses only when they suit this person.
+        * `action`: (String) Describe the action of the human.
+        * `gender`: (String) Describe the gender of the human.
+        * `skin_tone_and_texture`: (String) Describe the skin tone and texture, lit by bright daylight.
+    * If the object is a cluster of objects, include the following:
+        * `number_of_objects`: (Integer) The number of objects in the cluster.
+3.  `background_setting`: (String) Describe the overall environment, setting, or background, including any notable background elements that are not part of the `objects` section, curved and stretched by the fisheye.
+4.  `lighting`: (Object)
+    * `conditions`: (String) E.g., "bright midday sun under a clear blue sky", "bright, even white studio light".
+    * `direction`: (String) E.g., "front-lit by the sun", "backlit by the sun, with a flare in the corner".
+    * `shadows`: (String) Describe the presence and quality of shadows, e.g., "short, crisp shadows", "minimal shadows in the even studio light".
+5.  `aesthetics`: (Object)
+    * `composition`: (String) E.g., "extreme low angle from the ground, the figure towering against the sky", "high angle looking down at the upturned face", "the hand reaching to the lens fills the foreground". If people are the main subject, specify the shot type.
+    * `color_scheme`: (String) E.g., "bold, saturated colours against a deep blue sky", "punchy primary colours on a white background".
+    * `mood_atmosphere`: (String) E.g., "playful", "energetic", "humorous", "surreal".
+6.  `photographic_characteristics`: (Object)
+    * `depth_of_field`: (String) E.g., "deep, everything in focus from the enlarged foreground to the sky".
+    * `focus`: (String) E.g., "sharp focus throughout".
+    * `camera_angle`: (String) E.g., "extreme low angle looking up", "high angle looking down", "tilted".
+    * `lens_focal_length`: (String) E.g., "fisheye", "ultra-wide 10mm".
+7.  `style_medium`: (String) Identify the artistic style or medium based on the user's prompt or creative interpretation (e.g., "photograph", "oil painting", "watercolor", "3D render", "digital illustration", "pencil sketch").
+8.  `artistic_style`: (String) For a photograph in this style, three words in the register of the caption vocabulary. If the style is not "photograph", describe its specific artistic characteristics, 3 words maximum. (e.g., "impressionistic, vibrant, textured" for an oil painting).
+9.  `context`: (String) Provide a general description of the type of image this would be. For example: "This is a playful fisheye street-fashion portrait taken from the ground on a sunny day," or "This is an ultra-wide close-up photograph with an enlarged hand reaching toward the lens.".
+10. `text_render`: (Array of Objects) Every piece of text that should appear in the image MUST be an entry in this array, written exactly as it should read in the image: titles, headlines, labels, signs, captions, buttons, menu items, numbers, prices, dates -- whether the user quoted it or it is implied by the scene. Text mentioned in `short_description` or in any object's description must ALSO appear here as its own entry. Put each separate line, label or list item in its own entry. If the image has no text, the array is empty (`[]`).
+    * `text`: (String) The exact text as it should appear in the image. NEVER use generic placeholders.
+    * `location`: (String) E.g., "center", "top-left", "bottom-right foreground".
+    * `size`: (String) E.g., "medium", "large", "large within frame".
+    * `color`: (String) E.g., "red", "blue", "green".
+    * `font`: (String) E.g., "realistic", "cartoonish", "minimalist", "serif typeface".
+    * `appearance_details`: (String) Any other notable visual details.
+
+Ensure the information within the JSON is detailed, creative, internally consistent, and avoids redundancy between fields.

styles/lens_distortion/user_prompt.txt

@@ -0,0 +1 @@
+Generate a detailed JSON object, adhering to the expected schema, for an imagined scene based on the following request: {user_prompt}.

text_encoder/config.json

@@ -0,0 +1,109 @@
+{
+  "architectures": [
+    "Qwen3VLForConditionalGeneration"
+  ],
+  "dtype": "bfloat16",
+  "image_token_id": 151655,
+  "model_type": "qwen3_vl",
+  "quantization_config": {
+    "activation_bits": 4,
+    "activation_eps": 1e-10,
+    "activation_kernel_backend": "auto",
+    "activation_norm_dtype": "float32",
+    "adaln_group_size": 64,
+    "adaln_policy": "int4_rtn",
+    "artifact_format_version": 1,
+    "block_size": "paper",
+    "codebook": "lloyd_max",
+    "codebook_dtype": "float32",
+    "codebook_version": 2,
+    "fused_layout": null,
+    "lowbit_boundary_protection": "auto",
+    "lowbit_interior_protection": "auto",
+    "lowbit_protected_bits": 4,
+    "lowbit_protected_blocks": 4,
+    "modules_dtype_dict": {},
+    "modules_to_convert": [],
+    "modules_to_not_convert": [
+      "visual",
+      "lm_head"
+    ],
+    "modules_to_use_adaln": [],
+    "packed_matmul_block_k": 128,
+    "packed_matmul_block_m": 64,
+    "packed_matmul_block_n": 64,
+    "packed_matmul_num_warps": 4,
+    "quant_method": "orbitquant",
+    "rotation": "rpbh",
+    "rotation_seed": 0,
+    "row_norm_dtype": "bfloat16",
+    "runtime_mode": "auto_fused",
+    "target_policy": "universal",
+    "w4a4_int8_weight_cache": false,
+    "weight_bits": 4,
+    "weight_pack_dtype": "uint8",
+    "weight_row_tile_size": 256
+  },
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "dtype": "bfloat16",
+    "eos_token_id": 151645,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9728,
+    "max_position_embeddings": 262144,
+    "model_type": "qwen3_vl_text",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 36,
+    "num_key_value_heads": 8,
+    "pad_token_id": null,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        24,
+        20,
+        20
+      ],
+      "rope_theta": 5000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 151936
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.20.1",
+  "video_token_id": 151656,
+  "vision_config": {
+    "deepstack_visual_indexes": [
+      5,
+      11,
+      17
+    ],
+    "depth": 24,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1024,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 4096,
+    "model_type": "qwen3_vl_vision",
+    "num_heads": 16,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 2560,
+    "patch_size": 16,
+    "rope_parameters": {
+      "rope_theta": 10000.0,
+      "rope_type": "axial"
+    },
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652
+}

text_encoder/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.0,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "5.20.1"
+}

text_encoder/gptq-info.json

@@ -0,0 +1,713 @@
+{
+ "prompts": 50,
+ "damp": 0.01,
+ "chunk": 12,
+ "groups": {
+  "language_model.layers.0.self_attn.q_proj": {
+   "rtn": 0.003205,
+   "gptq": 0.000235,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.0.self_attn.o_proj": {
+   "rtn": 0.001781,
+   "gptq": 9.1e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.0.mlp.gate_proj": {
+   "rtn": 0.00581,
+   "gptq": 0.000576,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.0.mlp.down_proj": {
+   "rtn": 0.004301,
+   "gptq": 0.000258,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.1.self_attn.q_proj": {
+   "rtn": 0.00283,
+   "gptq": 0.000235,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.1.self_attn.o_proj": {
+   "rtn": 0.004678,
+   "gptq": 0.000221,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.1.mlp.gate_proj": {
+   "rtn": 0.000575,
+   "gptq": 1.4e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.1.mlp.down_proj": {
+   "rtn": 0.002321,
+   "gptq": 4.9e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.2.self_attn.q_proj": {
+   "rtn": 0.003703,
+   "gptq": 0.000444,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.2.self_attn.o_proj": {
+   "rtn": 0.005154,
+   "gptq": 0.000198,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.2.mlp.gate_proj": {
+   "rtn": 0.000802,
+   "gptq": 1.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.2.mlp.down_proj": {
+   "rtn": 0.002707,
+   "gptq": 6.1e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.3.self_attn.q_proj": {
+   "rtn": 0.002908,
+   "gptq": 0.000353,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.3.self_attn.o_proj": {
+   "rtn": 0.005961,
+   "gptq": 0.000243,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.3.mlp.gate_proj": {
+   "rtn": 0.000855,
+   "gptq": 3.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.3.mlp.down_proj": {
+   "rtn": 0.004568,
+   "gptq": 0.000156,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.4.self_attn.q_proj": {
+   "rtn": 0.003345,
+   "gptq": 0.000512,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.4.self_attn.o_proj": {
+   "rtn": 0.006979,
+   "gptq": 0.000413,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.4.mlp.gate_proj": {
+   "rtn": 0.001201,
+   "gptq": 8.1e-05,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.4.mlp.down_proj": {
+   "rtn": 0.003568,
+   "gptq": 0.000156,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.5.self_attn.q_proj": {
+   "rtn": 0.003057,
+   "gptq": 0.000402,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.5.self_attn.o_proj": {
+   "rtn": 0.004756,
+   "gptq": 0.000269,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.5.mlp.gate_proj": {
+   "rtn": 0.00165,
+   "gptq": 0.000185,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.5.mlp.down_proj": {
+   "rtn": 0.00617,
+   "gptq": 0.000528,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.6.self_attn.q_proj": {
+   "rtn": 0.003381,
+   "gptq": 0.000413,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.6.self_attn.o_proj": {
+   "rtn": 0.004784,
+   "gptq": 0.000408,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.6.mlp.gate_proj": {
+   "rtn": 0.003172,
+   "gptq": 0.000335,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.6.mlp.down_proj": {
+   "rtn": 0.003236,
+   "gptq": 2e-06,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.7.self_attn.q_proj": {
+   "rtn": 0.003181,
+   "gptq": 0.000312,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.7.self_attn.o_proj": {
+   "rtn": 0.006162,
+   "gptq": 0.000258,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.7.mlp.gate_proj": {
+   "rtn": 0.002792,
+   "gptq": 0.000341,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.7.mlp.down_proj": {
+   "rtn": 0.007865,
+   "gptq": 0.000805,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.8.self_attn.q_proj": {
+   "rtn": 0.003056,
+   "gptq": 0.000385,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.8.self_attn.o_proj": {
+   "rtn": 0.009043,
+   "gptq": 0.000659,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.8.mlp.gate_proj": {
+   "rtn": 0.004902,
+   "gptq": 0.00098,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.8.mlp.down_proj": {
+   "rtn": 0.00745,
+   "gptq": 0.000697,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.9.self_attn.q_proj": {
+   "rtn": 0.002826,
+   "gptq": 0.000353,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.9.self_attn.o_proj": {
+   "rtn": 0.010389,
+   "gptq": 0.000616,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.9.mlp.gate_proj": {
+   "rtn": 0.002886,
+   "gptq": 0.000535,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.9.mlp.down_proj": {
+   "rtn": 0.007373,
+   "gptq": 0.00063,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.10.self_attn.q_proj": {
+   "rtn": 0.003037,
+   "gptq": 0.000372,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.10.self_attn.o_proj": {
+   "rtn": 0.010868,
+   "gptq": 0.000647,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.10.mlp.gate_proj": {
+   "rtn": 0.003561,
+   "gptq": 0.000655,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.10.mlp.down_proj": {
+   "rtn": 0.007418,
+   "gptq": 0.000582,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.11.self_attn.q_proj": {
+   "rtn": 0.002772,
+   "gptq": 0.000312,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.11.self_attn.o_proj": {
+   "rtn": 0.010134,
+   "gptq": 0.000679,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.11.mlp.gate_proj": {
+   "rtn": 0.004313,
+   "gptq": 0.000782,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.11.mlp.down_proj": {
+   "rtn": 0.007403,
+   "gptq": 0.000648,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.12.self_attn.q_proj": {
+   "rtn": 0.00295,
+   "gptq": 0.000287,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.12.self_attn.o_proj": {
+   "rtn": 0.009916,
+   "gptq": 0.000537,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.12.mlp.gate_proj": {
+   "rtn": 0.004555,
+   "gptq": 0.000814,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.12.mlp.down_proj": {
+   "rtn": 0.007553,
+   "gptq": 0.00069,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.13.self_attn.q_proj": {
+   "rtn": 0.002479,
+   "gptq": 0.000235,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.13.self_attn.o_proj": {
+   "rtn": 0.010994,
+   "gptq": 0.000539,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.13.mlp.gate_proj": {
+   "rtn": 0.004642,
+   "gptq": 0.000794,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.13.mlp.down_proj": {
+   "rtn": 0.007539,
+   "gptq": 0.000749,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.14.self_attn.q_proj": {
+   "rtn": 0.003211,
+   "gptq": 0.000252,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.14.self_attn.o_proj": {
+   "rtn": 0.013797,
+   "gptq": 0.00089,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.14.mlp.gate_proj": {
+   "rtn": 0.00446,
+   "gptq": 0.000767,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.14.mlp.down_proj": {
+   "rtn": 0.008061,
+   "gptq": 0.000756,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.15.self_attn.q_proj": {
+   "rtn": 0.003264,
+   "gptq": 0.000247,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.15.self_attn.o_proj": {
+   "rtn": 0.016946,
+   "gptq": 0.000845,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.15.mlp.gate_proj": {
+   "rtn": 0.005036,
+   "gptq": 0.000865,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.15.mlp.down_proj": {
+   "rtn": 0.008045,
+   "gptq": 0.000818,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.16.self_attn.q_proj": {
+   "rtn": 0.003648,
+   "gptq": 0.000331,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.16.self_attn.o_proj": {
+   "rtn": 0.015247,
+   "gptq": 0.001018,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.16.mlp.gate_proj": {
+   "rtn": 0.004584,
+   "gptq": 0.000785,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.16.mlp.down_proj": {
+   "rtn": 0.005004,
+   "gptq": 6e-06,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.17.self_attn.q_proj": {
+   "rtn": 0.003694,
+   "gptq": 0.000323,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.17.self_attn.o_proj": {
+   "rtn": 0.017739,
+   "gptq": 0.000821,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.17.mlp.gate_proj": {
+   "rtn": 0.005166,
+   "gptq": 0.000858,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.17.mlp.down_proj": {
+   "rtn": 0.008752,
+   "gptq": 0.000922,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.18.self_attn.q_proj": {
+   "rtn": 0.003874,
+   "gptq": 0.000285,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.18.self_attn.o_proj": {
+   "rtn": 0.020685,
+   "gptq": 0.001102,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.18.mlp.gate_proj": {
+   "rtn": 0.005446,
+   "gptq": 0.000941,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.18.mlp.down_proj": {
+   "rtn": 0.008773,
+   "gptq": 0.00088,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.19.self_attn.q_proj": {
+   "rtn": 0.004023,
+   "gptq": 0.000301,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.19.self_attn.o_proj": {
+   "rtn": 0.015264,
+   "gptq": 0.000852,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.19.mlp.gate_proj": {
+   "rtn": 0.005632,
+   "gptq": 0.000939,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.19.mlp.down_proj": {
+   "rtn": 0.007381,
+   "gptq": 0.00069,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.20.self_attn.q_proj": {
+   "rtn": 0.004802,
+   "gptq": 0.000314,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.20.self_attn.o_proj": {
+   "rtn": 0.018694,
+   "gptq": 0.00106,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.20.mlp.gate_proj": {
+   "rtn": 0.00568,
+   "gptq": 0.000957,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.20.mlp.down_proj": {
+   "rtn": 0.009597,
+   "gptq": 0.00101,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.21.self_attn.q_proj": {
+   "rtn": 0.005561,
+   "gptq": 0.000324,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.21.self_attn.o_proj": {
+   "rtn": 0.018366,
+   "gptq": 0.001076,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.21.mlp.gate_proj": {
+   "rtn": 0.005668,
+   "gptq": 0.000935,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.21.mlp.down_proj": {
+   "rtn": 0.008754,
+   "gptq": 0.000835,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.22.self_attn.q_proj": {
+   "rtn": 0.005449,
+   "gptq": 0.000352,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.22.self_attn.o_proj": {
+   "rtn": 0.015129,
+   "gptq": 0.000829,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.22.mlp.gate_proj": {
+   "rtn": 0.005202,
+   "gptq": 0.000719,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.22.mlp.down_proj": {
+   "rtn": 0.008571,
+   "gptq": 0.000623,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.23.self_attn.q_proj": {
+   "rtn": 0.004695,
+   "gptq": 0.000286,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.23.self_attn.o_proj": {
+   "rtn": 0.009198,
+   "gptq": 0.000528,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.23.mlp.gate_proj": {
+   "rtn": 0.005401,
+   "gptq": 0.00068,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.23.mlp.down_proj": {
+   "rtn": 0.008493,
+   "gptq": 0.000664,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.24.self_attn.q_proj": {
+   "rtn": 0.005259,
+   "gptq": 0.000324,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.24.self_attn.o_proj": {
+   "rtn": 0.009858,
+   "gptq": 0.000737,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.24.mlp.gate_proj": {
+   "rtn": 0.005258,
+   "gptq": 0.000654,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.24.mlp.down_proj": {
+   "rtn": 0.008625,
+   "gptq": 0.000616,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.25.self_attn.q_proj": {
+   "rtn": 0.00654,
+   "gptq": 0.000419,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.25.self_attn.o_proj": {
+   "rtn": 0.009702,
+   "gptq": 0.000692,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.25.mlp.gate_proj": {
+   "rtn": 0.005045,
+   "gptq": 0.000662,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.25.mlp.down_proj": {
+   "rtn": 0.009319,
+   "gptq": 0.000746,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.26.self_attn.q_proj": {
+   "rtn": 0.005758,
+   "gptq": 0.000435,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.26.self_attn.o_proj": {
+   "rtn": 0.01079,
+   "gptq": 0.000789,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.26.mlp.gate_proj": {
+   "rtn": 0.004758,
+   "gptq": 0.000658,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.26.mlp.down_proj": {
+   "rtn": 0.009172,
+   "gptq": 0.000727,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.27.self_attn.q_proj": {
+   "rtn": 0.00681,
+   "gptq": 0.000505,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.27.self_attn.o_proj": {
+   "rtn": 0.011078,
+   "gptq": 0.000707,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.27.mlp.gate_proj": {
+   "rtn": 0.004669,
+   "gptq": 0.000676,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.27.mlp.down_proj": {
+   "rtn": 0.010172,
+   "gptq": 0.000727,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.28.self_attn.q_proj": {
+   "rtn": 0.007226,
+   "gptq": 0.000445,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.28.self_attn.o_proj": {
+   "rtn": 0.014353,
+   "gptq": 0.000882,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.28.mlp.gate_proj": {
+   "rtn": 0.004835,
+   "gptq": 0.000743,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.28.mlp.down_proj": {
+   "rtn": 0.009468,
+   "gptq": 0.000708,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.29.self_attn.q_proj": {
+   "rtn": 0.006585,
+   "gptq": 0.000502,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.29.self_attn.o_proj": {
+   "rtn": 0.010203,
+   "gptq": 0.000751,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.29.mlp.gate_proj": {
+   "rtn": 0.004568,
+   "gptq": 0.000734,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.29.mlp.down_proj": {
+   "rtn": 0.009273,
+   "gptq": 0.000609,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.30.self_attn.q_proj": {
+   "rtn": 0.00555,
+   "gptq": 0.000508,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.30.self_attn.o_proj": {
+   "rtn": 0.009545,
+   "gptq": 0.000778,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.30.mlp.gate_proj": {
+   "rtn": 0.004774,
+   "gptq": 0.000812,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.30.mlp.down_proj": {
+   "rtn": 0.008301,
+   "gptq": 0.000567,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.31.self_attn.q_proj": {
+   "rtn": 0.0062,
+   "gptq": 0.000497,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.31.self_attn.o_proj": {
+   "rtn": 0.011464,
+   "gptq": 0.000903,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.31.mlp.gate_proj": {
+   "rtn": 0.005131,
+   "gptq": 0.000928,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.31.mlp.down_proj": {
+   "rtn": 0.00927,
+   "gptq": 0.000654,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.32.self_attn.q_proj": {
+   "rtn": 0.006029,
+   "gptq": 0.000517,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.32.self_attn.o_proj": {
+   "rtn": 0.009629,
+   "gptq": 0.000641,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.32.mlp.gate_proj": {
+   "rtn": 0.005308,
+   "gptq": 0.001008,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.32.mlp.down_proj": {
+   "rtn": 0.009845,
+   "gptq": 0.000777,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.33.self_attn.q_proj": {
+   "rtn": 0.006633,
+   "gptq": 0.000529,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.33.self_attn.o_proj": {
+   "rtn": 0.010626,
+   "gptq": 0.000676,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.33.mlp.gate_proj": {
+   "rtn": 0.004671,
+   "gptq": 0.000878,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.33.mlp.down_proj": {
+   "rtn": 0.010446,
+   "gptq": 0.000773,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.34.self_attn.q_proj": {
+   "rtn": 0.005965,
+   "gptq": 0.000595,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.34.self_attn.o_proj": {
+   "rtn": 0.008557,
+   "gptq": 0.000304,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.34.mlp.gate_proj": {
+   "rtn": 0.003245,
+   "gptq": 0.000497,
+   "damp": 0.009999999776482582
+  },
+  "language_model.layers.34.mlp.down_proj": {
+   "rtn": 0.003054,
+   "gptq": 0.000103,
+   "damp": 0.009999999776482582
+  }
+ },
+ "tokens": 26586,
+ "mean": {
+  "rtn": 0.006686,
+  "gptq": 0.000563
+ },
+ "seconds": 52.6
+}

text_encoder/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c112d0262eb40cc3a19d3447bc0891fdfc23dc6fca4d8c5a39c34e72387d81bd
+size 3427991944

transformer/config.json

@@ -0,0 +1,37 @@
+{
+  "_class_name": "BriaFibo2Transformer2DModel",
+  "_diffusers_version": "0.41.0.dev0",
+  "axes_dims": [
+    32,
+    48,
+    48
+  ],
+  "axes_lens": [
+    2048,
+    512,
+    512
+  ],
+  "cap_feat_dim": 7680,
+  "dim": 3584,
+  "gist_max_text_len": 2048,
+  "gist_min_text_len": 256,
+  "gist_step": 64,
+  "in_channels": 32,
+  "injection_layer_ids": [
+    6,
+    14,
+    22,
+    30,
+    38
+  ],
+  "max_num_gist_tokens": 512,
+  "min_num_gist_tokens": 256,
+  "n_heads": 28,
+  "n_layers": 44,
+  "n_refiner_layers": 2,
+  "norm_eps": 1e-05,
+  "patch_size": 2,
+  "perceiver_num_layers": 2,
+  "rope_theta": 256.0,
+  "t_scale": 1000.0
+}

transformer/gptq-info.json

@@ -0,0 +1,1106 @@
+{
+ "calibration": [
+  "c_baker",
+  "c_robot_garden",
+  "c_bookshop_sign",
+  "c_sneaker",
+  "c_ru_poster",
+  "c_fox_snow",
+  "c_interior",
+  "c_chess",
+  "ce_baker_night",
+  "ce_robot_blue",
+  "ce_sneaker_black",
+  "ce_interior_cat"
+ ],
+ "steps": [
+  0,
+  4,
+  8,
+  12,
+  16,
+  20,
+  25,
+  30,
+  35,
+  40,
+  45,
+  49
+ ],
+ "damp": 0.01,
+ "chunk": 12,
+ "sequential": false,
+ "forwards": 288,
+ "groups": {
+  "noise_refiner.0.attention.to_q": {
+   "rtn": 0.000678,
+   "gptq": 9e-06,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.0.attention.to_out.0": {
+   "rtn": 0.00173,
+   "gptq": 3.4e-05,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.0.feed_forward.w1": {
+   "rtn": 0.004542,
+   "gptq": 0.000128,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.0.feed_forward.w2": {
+   "rtn": 0.002435,
+   "gptq": 0.000403,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.1.attention.to_q": {
+   "rtn": 0.002498,
+   "gptq": 0.000102,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.1.attention.to_out.0": {
+   "rtn": 0.003027,
+   "gptq": 8.7e-05,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.1.feed_forward.w1": {
+   "rtn": 0.006424,
+   "gptq": 0.000367,
+   "damp": 0.009999999776482582
+  },
+  "noise_refiner.1.feed_forward.w2": {
+   "rtn": 0.003332,
+   "gptq": 0.000667,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.cross_attention.to_q": {
+   "rtn": 0.000751,
+   "gptq": 1.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.cross_attention.to_k": {
+   "rtn": 0.008963,
+   "gptq": 0.00032,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.cross_attention.to_out.0": {
+   "rtn": 0.004306,
+   "gptq": 0.000157,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.self_attention.to_q": {
+   "rtn": 0.004538,
+   "gptq": 0.000133,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.self_attention.to_out.0": {
+   "rtn": 0.004807,
+   "gptq": 5.6e-05,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.ffn.w1": {
+   "rtn": 0.0051,
+   "gptq": 0.000142,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.0.ffn.w2": {
+   "rtn": 0.002934,
+   "gptq": 4e-06,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.cross_attention.to_q": {
+   "rtn": 0.007914,
+   "gptq": 0.000103,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.cross_attention.to_k": {
+   "rtn": 0.010409,
+   "gptq": 0.000367,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.cross_attention.to_out.0": {
+   "rtn": 0.005372,
+   "gptq": 0.000112,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.self_attention.to_q": {
+   "rtn": 0.008579,
+   "gptq": 0.000123,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.self_attention.to_out.0": {
+   "rtn": 0.005654,
+   "gptq": 3.7e-05,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.ffn.w1": {
+   "rtn": 0.009356,
+   "gptq": 0.000116,
+   "damp": 0.009999999776482582
+  },
+  "perceiver.layers.1.ffn.w2": {
+   "rtn": 0.005421,
+   "gptq": 3e-06,
+   "damp": 0.009999999776482582
+  },
+  "layers.0.attention.to_q": {
+   "rtn": 0.005307,
+   "gptq": 0.000436,
+   "damp": 0.009999999776482582
+  },
+  "layers.0.attention.to_out.0": {
+   "rtn": 0.002504,
+   "gptq": 3.4e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.0.feed_forward.w1": {
+   "rtn": 0.008166,
+   "gptq": 0.000622,
+   "damp": 0.009999999776482582
+  },
+  "layers.0.feed_forward.w2": {
+   "rtn": 0.012292,
+   "gptq": 3e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.1.attention.to_q": {
+   "rtn": 0.005291,
+   "gptq": 0.000455,
+   "damp": 0.009999999776482582
+  },
+  "layers.1.attention.to_out.0": {
+   "rtn": 0.002606,
+   "gptq": 3.4e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.1.feed_forward.w1": {
+   "rtn": 0.007323,
+   "gptq": 0.000516,
+   "damp": 0.009999999776482582
+  },
+  "layers.1.feed_forward.w2": {
+   "rtn": 0.032419,
+   "gptq": 0.001202,
+   "damp": 0.009999999776482582
+  },
+  "layers.2.attention.to_q": {
+   "rtn": 0.005153,
+   "gptq": 0.000435,
+   "damp": 0.009999999776482582
+  },
+  "layers.2.attention.to_out.0": {
+   "rtn": 0.003214,
+   "gptq": 8.5e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.2.feed_forward.w1": {
+   "rtn": 0.006575,
+   "gptq": 0.000477,
+   "damp": 0.009999999776482582
+  },
+  "layers.2.feed_forward.w2": {
+   "rtn": 0.017544,
+   "gptq": 0.000348,
+   "damp": 0.009999999776482582
+  },
+  "layers.3.attention.to_q": {
+   "rtn": 0.005015,
+   "gptq": 0.000454,
+   "damp": 0.009999999776482582
+  },
+  "layers.3.attention.to_out.0": {
+   "rtn": 0.003576,
+   "gptq": 0.000146,
+   "damp": 0.009999999776482582
+  },
+  "layers.3.feed_forward.w1": {
+   "rtn": 0.006357,
+   "gptq": 0.000581,
+   "damp": 0.009999999776482582
+  },
+  "layers.3.feed_forward.w2": {
+   "rtn": 0.004782,
+   "gptq": 4.5e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.4.attention.to_q": {
+   "rtn": 0.00543,
+   "gptq": 0.000554,
+   "damp": 0.009999999776482582
+  },
+  "layers.4.attention.to_out.0": {
+   "rtn": 0.003473,
+   "gptq": 0.000134,
+   "damp": 0.009999999776482582
+  },
+  "layers.4.feed_forward.w1": {
+   "rtn": 0.006745,
+   "gptq": 0.000701,
+   "damp": 0.009999999776482582
+  },
+  "layers.4.feed_forward.w2": {
+   "rtn": 0.009405,
+   "gptq": 0.000542,
+   "damp": 0.009999999776482582
+  },
+  "layers.5.attention.to_q": {
+   "rtn": 0.00511,
+   "gptq": 0.000585,
+   "damp": 0.009999999776482582
+  },
+  "layers.5.attention.to_out.0": {
+   "rtn": 0.003632,
+   "gptq": 0.000186,
+   "damp": 0.009999999776482582
+  },
+  "layers.5.feed_forward.w1": {
+   "rtn": 0.008035,
+   "gptq": 0.000987,
+   "damp": 0.009999999776482582
+  },
+  "layers.5.feed_forward.w2": {
+   "rtn": 0.042309,
+   "gptq": 0.000162,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.attention.to_q": {
+   "rtn": 0.005456,
+   "gptq": 0.000702,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.attention.to_out.0": {
+   "rtn": 0.003491,
+   "gptq": 0.000136,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.feed_forward.w1": {
+   "rtn": 0.008482,
+   "gptq": 0.001091,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.feed_forward.w2": {
+   "rtn": 0.007299,
+   "gptq": 3.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.6.cross_attention.to_q": {
+   "rtn": 0.005628,
+   "gptq": 0.000628,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.cross_attention.to_k": {
+   "rtn": 0.007379,
+   "gptq": 0.00026,
+   "damp": 0.009999999776482582
+  },
+  "layers.6.cross_attention.to_out.0": {
+   "rtn": 0.002637,
+   "gptq": 7e-06,
+   "damp": 0.009999999776482582
+  },
+  "layers.7.attention.to_q": {
+   "rtn": 0.005646,
+   "gptq": 0.000748,
+   "damp": 0.009999999776482582
+  },
+  "layers.7.attention.to_out.0": {
+   "rtn": 0.003822,
+   "gptq": 0.00022,
+   "damp": 0.009999999776482582
+  },
+  "layers.7.feed_forward.w1": {
+   "rtn": 0.00816,
+   "gptq": 0.001126,
+   "damp": 0.009999999776482582
+  },
+  "layers.7.feed_forward.w2": {
+   "rtn": 0.004546,
+   "gptq": 1.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.8.attention.to_q": {
+   "rtn": 0.005705,
+   "gptq": 0.000826,
+   "damp": 0.009999999776482582
+  },
+  "layers.8.attention.to_out.0": {
+   "rtn": 0.004215,
+   "gptq": 0.000232,
+   "damp": 0.009999999776482582
+  },
+  "layers.8.feed_forward.w1": {
+   "rtn": 0.007925,
+   "gptq": 0.001167,
+   "damp": 0.009999999776482582
+  },
+  "layers.8.feed_forward.w2": {
+   "rtn": 0.069567,
+   "gptq": 3.9e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.9.attention.to_q": {
+   "rtn": 0.005928,
+   "gptq": 0.000952,
+   "damp": 0.009999999776482582
+  },
+  "layers.9.attention.to_out.0": {
+   "rtn": 0.004017,
+   "gptq": 0.000267,
+   "damp": 0.009999999776482582
+  },
+  "layers.9.feed_forward.w1": {
+   "rtn": 0.008115,
+   "gptq": 0.001299,
+   "damp": 0.009999999776482582
+  },
+  "layers.9.feed_forward.w2": {
+   "rtn": 0.037066,
+   "gptq": 8.5e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.10.attention.to_q": {
+   "rtn": 0.005875,
+   "gptq": 0.00097,
+   "damp": 0.009999999776482582
+  },
+  "layers.10.attention.to_out.0": {
+   "rtn": 0.003241,
+   "gptq": 0.000181,
+   "damp": 0.009999999776482582
+  },
+  "layers.10.feed_forward.w1": {
+   "rtn": 0.008089,
+   "gptq": 0.00134,
+   "damp": 0.009999999776482582
+  },
+  "layers.10.feed_forward.w2": {
+   "rtn": 0.004258,
+   "gptq": 1.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.11.attention.to_q": {
+   "rtn": 0.005977,
+   "gptq": 0.001046,
+   "damp": 0.009999999776482582
+  },
+  "layers.11.attention.to_out.0": {
+   "rtn": 0.004077,
+   "gptq": 0.000279,
+   "damp": 0.009999999776482582
+  },
+  "layers.11.feed_forward.w1": {
+   "rtn": 0.008094,
+   "gptq": 0.001417,
+   "damp": 0.009999999776482582
+  },
+  "layers.11.feed_forward.w2": {
+   "rtn": 0.005118,
+   "gptq": 1.7e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.12.attention.to_q": {
+   "rtn": 0.005698,
+   "gptq": 0.001014,
+   "damp": 0.009999999776482582
+  },
+  "layers.12.attention.to_out.0": {
+   "rtn": 0.004246,
+   "gptq": 0.00026,
+   "damp": 0.009999999776482582
+  },
+  "layers.12.feed_forward.w1": {
+   "rtn": 0.008155,
+   "gptq": 0.001441,
+   "damp": 0.009999999776482582
+  },
+  "layers.12.feed_forward.w2": {
+   "rtn": 0.019204,
+   "gptq": 6.7e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.13.attention.to_q": {
+   "rtn": 0.005801,
+   "gptq": 0.00106,
+   "damp": 0.009999999776482582
+  },
+  "layers.13.attention.to_out.0": {
+   "rtn": 0.004605,
+   "gptq": 0.000334,
+   "damp": 0.009999999776482582
+  },
+  "layers.13.feed_forward.w1": {
+   "rtn": 0.008193,
+   "gptq": 0.001492,
+   "damp": 0.009999999776482582
+  },
+  "layers.13.feed_forward.w2": {
+   "rtn": 0.007997,
+   "gptq": 5e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.attention.to_q": {
+   "rtn": 0.005704,
+   "gptq": 0.001089,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.attention.to_out.0": {
+   "rtn": 0.005197,
+   "gptq": 0.000494,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.feed_forward.w1": {
+   "rtn": 0.008463,
+   "gptq": 0.001541,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.feed_forward.w2": {
+   "rtn": 0.004315,
+   "gptq": 4.2e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.14.cross_attention.to_q": {
+   "rtn": 0.006568,
+   "gptq": 0.001105,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.cross_attention.to_k": {
+   "rtn": 0.00845,
+   "gptq": 0.000297,
+   "damp": 0.009999999776482582
+  },
+  "layers.14.cross_attention.to_out.0": {
+   "rtn": 0.00365,
+   "gptq": 1e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.15.attention.to_q": {
+   "rtn": 0.005583,
+   "gptq": 0.001037,
+   "damp": 0.009999999776482582
+  },
+  "layers.15.attention.to_out.0": {
+   "rtn": 0.004961,
+   "gptq": 0.00045,
+   "damp": 0.009999999776482582
+  },
+  "layers.15.feed_forward.w1": {
+   "rtn": 0.008149,
+   "gptq": 0.001562,
+   "damp": 0.009999999776482582
+  },
+  "layers.15.feed_forward.w2": {
+   "rtn": 0.004722,
+   "gptq": 2.3e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.16.attention.to_q": {
+   "rtn": 0.006244,
+   "gptq": 0.001267,
+   "damp": 0.009999999776482582
+  },
+  "layers.16.attention.to_out.0": {
+   "rtn": 0.004706,
+   "gptq": 0.000426,
+   "damp": 0.009999999776482582
+  },
+  "layers.16.feed_forward.w1": {
+   "rtn": 0.007934,
+   "gptq": 0.001538,
+   "damp": 0.009999999776482582
+  },
+  "layers.16.feed_forward.w2": {
+   "rtn": 0.005775,
+   "gptq": 2.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.17.attention.to_q": {
+   "rtn": 0.005969,
+   "gptq": 0.001179,
+   "damp": 0.009999999776482582
+  },
+  "layers.17.attention.to_out.0": {
+   "rtn": 0.005595,
+   "gptq": 0.000565,
+   "damp": 0.009999999776482582
+  },
+  "layers.17.feed_forward.w1": {
+   "rtn": 0.007644,
+   "gptq": 0.001489,
+   "damp": 0.009999999776482582
+  },
+  "layers.17.feed_forward.w2": {
+   "rtn": 0.004638,
+   "gptq": 1.6e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.18.attention.to_q": {
+   "rtn": 0.006082,
+   "gptq": 0.001218,
+   "damp": 0.009999999776482582
+  },
+  "layers.18.attention.to_out.0": {
+   "rtn": 0.00571,
+   "gptq": 0.000584,
+   "damp": 0.009999999776482582
+  },
+  "layers.18.feed_forward.w1": {
+   "rtn": 0.007428,
+   "gptq": 0.001443,
+   "damp": 0.009999999776482582
+  },
+  "layers.18.feed_forward.w2": {
+   "rtn": 0.005865,
+   "gptq": 2.5e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.19.attention.to_q": {
+   "rtn": 0.005574,
+   "gptq": 0.001058,
+   "damp": 0.009999999776482582
+  },
+  "layers.19.attention.to_out.0": {
+   "rtn": 0.006877,
+   "gptq": 0.000693,
+   "damp": 0.009999999776482582
+  },
+  "layers.19.feed_forward.w1": {
+   "rtn": 0.007611,
+   "gptq": 0.001445,
+   "damp": 0.009999999776482582
+  },
+  "layers.19.feed_forward.w2": {
+   "rtn": 0.011469,
+   "gptq": 3.7e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.20.attention.to_q": {
+   "rtn": 0.00562,
+   "gptq": 0.001071,
+   "damp": 0.009999999776482582
+  },
+  "layers.20.attention.to_out.0": {
+   "rtn": 0.007164,
+   "gptq": 0.000569,
+   "damp": 0.009999999776482582
+  },
+  "layers.20.feed_forward.w1": {
+   "rtn": 0.007629,
+   "gptq": 0.001475,
+   "damp": 0.009999999776482582
+  },
+  "layers.20.feed_forward.w2": {
+   "rtn": 0.038774,
+   "gptq": 2.5e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.21.attention.to_q": {
+   "rtn": 0.006094,
+   "gptq": 0.001211,
+   "damp": 0.009999999776482582
+  },
+  "layers.21.attention.to_out.0": {
+   "rtn": 0.005007,
+   "gptq": 0.000462,
+   "damp": 0.009999999776482582
+  },
+  "layers.21.feed_forward.w1": {
+   "rtn": 0.007545,
+   "gptq": 0.001443,
+   "damp": 0.009999999776482582
+  },
+  "layers.21.feed_forward.w2": {
+   "rtn": 0.005654,
+   "gptq": 4.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.22.attention.to_q": {
+   "rtn": 0.005953,
+   "gptq": 0.001184,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.attention.to_out.0": {
+   "rtn": 0.0063,
+   "gptq": 0.000602,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.feed_forward.w1": {
+   "rtn": 0.008201,
+   "gptq": 0.001458,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.feed_forward.w2": {
+   "rtn": 0.005102,
+   "gptq": 2.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.cross_attention.to_q": {
+   "rtn": 0.007824,
+   "gptq": 0.001302,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.cross_attention.to_k": {
+   "rtn": 0.013198,
+   "gptq": 0.000471,
+   "damp": 0.009999999776482582
+  },
+  "layers.22.cross_attention.to_out.0": {
+   "rtn": 0.002607,
+   "gptq": 1.8e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.23.attention.to_q": {
+   "rtn": 0.006198,
+   "gptq": 0.001209,
+   "damp": 0.009999999776482582
+  },
+  "layers.23.attention.to_out.0": {
+   "rtn": 0.005177,
+   "gptq": 0.000542,
+   "damp": 0.009999999776482582
+  },
+  "layers.23.feed_forward.w1": {
+   "rtn": 0.007887,
+   "gptq": 0.001445,
+   "damp": 0.009999999776482582
+  },
+  "layers.23.feed_forward.w2": {
+   "rtn": 0.006917,
+   "gptq": 1.6e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.24.attention.to_q": {
+   "rtn": 0.005951,
+   "gptq": 0.001161,
+   "damp": 0.009999999776482582
+  },
+  "layers.24.attention.to_out.0": {
+   "rtn": 0.004627,
+   "gptq": 0.000553,
+   "damp": 0.009999999776482582
+  },
+  "layers.24.feed_forward.w1": {
+   "rtn": 0.008246,
+   "gptq": 0.001541,
+   "damp": 0.009999999776482582
+  },
+  "layers.24.feed_forward.w2": {
+   "rtn": 0.003965,
+   "gptq": 2.5e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.25.attention.to_q": {
+   "rtn": 0.00572,
+   "gptq": 0.001147,
+   "damp": 0.009999999776482582
+  },
+  "layers.25.attention.to_out.0": {
+   "rtn": 0.006515,
+   "gptq": 0.000668,
+   "damp": 0.009999999776482582
+  },
+  "layers.25.feed_forward.w1": {
+   "rtn": 0.007912,
+   "gptq": 0.001515,
+   "damp": 0.009999999776482582
+  },
+  "layers.25.feed_forward.w2": {
+   "rtn": 0.004337,
+   "gptq": 2.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.26.attention.to_q": {
+   "rtn": 0.006164,
+   "gptq": 0.001256,
+   "damp": 0.009999999776482582
+  },
+  "layers.26.attention.to_out.0": {
+   "rtn": 0.005315,
+   "gptq": 0.000573,
+   "damp": 0.009999999776482582
+  },
+  "layers.26.feed_forward.w1": {
+   "rtn": 0.008077,
+   "gptq": 0.001524,
+   "damp": 0.009999999776482582
+  },
+  "layers.26.feed_forward.w2": {
+   "rtn": 0.00427,
+   "gptq": 1.5e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.27.attention.to_q": {
+   "rtn": 0.005482,
+   "gptq": 0.001097,
+   "damp": 0.009999999776482582
+  },
+  "layers.27.attention.to_out.0": {
+   "rtn": 0.006903,
+   "gptq": 0.000876,
+   "damp": 0.009999999776482582
+  },
+  "layers.27.feed_forward.w1": {
+   "rtn": 0.007964,
+   "gptq": 0.001533,
+   "damp": 0.009999999776482582
+  },
+  "layers.27.feed_forward.w2": {
+   "rtn": 0.004998,
+   "gptq": 2.3e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.28.attention.to_q": {
+   "rtn": 0.005646,
+   "gptq": 0.001135,
+   "damp": 0.009999999776482582
+  },
+  "layers.28.attention.to_out.0": {
+   "rtn": 0.00654,
+   "gptq": 0.000771,
+   "damp": 0.009999999776482582
+  },
+  "layers.28.feed_forward.w1": {
+   "rtn": 0.008074,
+   "gptq": 0.00153,
+   "damp": 0.009999999776482582
+  },
+  "layers.28.feed_forward.w2": {
+   "rtn": 0.005644,
+   "gptq": 1.3e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.29.attention.to_q": {
+   "rtn": 0.00574,
+   "gptq": 0.001147,
+   "damp": 0.009999999776482582
+  },
+  "layers.29.attention.to_out.0": {
+   "rtn": 0.006,
+   "gptq": 0.000649,
+   "damp": 0.009999999776482582
+  },
+  "layers.29.feed_forward.w1": {
+   "rtn": 0.007762,
+   "gptq": 0.00151,
+   "damp": 0.009999999776482582
+  },
+  "layers.29.feed_forward.w2": {
+   "rtn": 0.010018,
+   "gptq": 3.8e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.30.attention.to_q": {
+   "rtn": 0.0059,
+   "gptq": 0.00123,
+   "damp": 0.009999999776482582
+  },
+  "layers.30.attention.to_out.0": {
+   "rtn": 0.005795,
+   "gptq": 0.000663,
+   "damp": 0.009999999776482582
+  },
+  "layers.30.feed_forward.w1": {
+   "rtn": 0.008109,
+   "gptq": 0.001462,
+   "damp": 0.009999999776482582
+  },
+  "layers.30.feed_forward.w2": {
+   "rtn": 0.05444,
+   "gptq": 2.7e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.30.cross_attention.to_q": {
+   "rtn": 0.044882,
+   "gptq": 0.001667,
+   "damp": 0.009999999776482582
+  },
+  "layers.30.cross_attention.to_k": {
+   "rtn": 0.014876,
+   "gptq": 0.000535,
+   "damp": 0.009999999776482582
+  },
+  "layers.30.cross_attention.to_out.0": {
+   "rtn": 0.002633,
+   "gptq": 1.6e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.31.attention.to_q": {
+   "rtn": 0.006047,
+   "gptq": 0.001174,
+   "damp": 0.009999999776482582
+  },
+  "layers.31.attention.to_out.0": {
+   "rtn": 0.006015,
+   "gptq": 0.000597,
+   "damp": 0.009999999776482582
+  },
+  "layers.31.feed_forward.w1": {
+   "rtn": 0.007762,
+   "gptq": 0.001446,
+   "damp": 0.009999999776482582
+  },
+  "layers.31.feed_forward.w2": {
+   "rtn": 0.003912,
+   "gptq": 1.1e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.32.attention.to_q": {
+   "rtn": 0.005835,
+   "gptq": 0.001134,
+   "damp": 0.009999999776482582
+  },
+  "layers.32.attention.to_out.0": {
+   "rtn": 0.008367,
+   "gptq": 0.001119,
+   "damp": 0.009999999776482582
+  },
+  "layers.32.feed_forward.w1": {
+   "rtn": 0.007634,
+   "gptq": 0.001391,
+   "damp": 0.009999999776482582
+  },
+  "layers.32.feed_forward.w2": {
+   "rtn": 0.004266,
+   "gptq": 1.1e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.33.attention.to_q": {
+   "rtn": 0.006164,
+   "gptq": 0.001142,
+   "damp": 0.009999999776482582
+  },
+  "layers.33.attention.to_out.0": {
+   "rtn": 0.007784,
+   "gptq": 0.001028,
+   "damp": 0.009999999776482582
+  },
+  "layers.33.feed_forward.w1": {
+   "rtn": 0.007062,
+   "gptq": 0.001257,
+   "damp": 0.009999999776482582
+  },
+  "layers.33.feed_forward.w2": {
+   "rtn": 0.004936,
+   "gptq": 1e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.34.attention.to_q": {
+   "rtn": 0.00563,
+   "gptq": 0.000973,
+   "damp": 0.009999999776482582
+  },
+  "layers.34.attention.to_out.0": {
+   "rtn": 0.008936,
+   "gptq": 0.000943,
+   "damp": 0.009999999776482582
+  },
+  "layers.34.feed_forward.w1": {
+   "rtn": 0.006806,
+   "gptq": 0.001113,
+   "damp": 0.009999999776482582
+  },
+  "layers.34.feed_forward.w2": {
+   "rtn": 0.018027,
+   "gptq": 5.6e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.35.attention.to_q": {
+   "rtn": 0.006452,
+   "gptq": 0.001058,
+   "damp": 0.009999999776482582
+  },
+  "layers.35.attention.to_out.0": {
+   "rtn": 0.008174,
+   "gptq": 0.000873,
+   "damp": 0.009999999776482582
+  },
+  "layers.35.feed_forward.w1": {
+   "rtn": 0.006755,
+   "gptq": 0.001037,
+   "damp": 0.009999999776482582
+  },
+  "layers.35.feed_forward.w2": {
+   "rtn": 0.00439,
+   "gptq": 4e-06,
+   "damp": 0.09999999403953552
+  },
+  "layers.36.attention.to_q": {
+   "rtn": 0.005917,
+   "gptq": 0.000905,
+   "damp": 0.009999999776482582
+  },
+  "layers.36.attention.to_out.0": {
+   "rtn": 0.008336,
+   "gptq": 0.000958,
+   "damp": 0.009999999776482582
+  },
+  "layers.36.feed_forward.w1": {
+   "rtn": 0.007123,
+   "gptq": 0.001075,
+   "damp": 0.009999999776482582
+  },
+  "layers.36.feed_forward.w2": {
+   "rtn": 0.005353,
+   "gptq": 5e-06,
+   "damp": 0.09999999403953552
+  },
+  "layers.37.attention.to_q": {
+   "rtn": 0.00603,
+   "gptq": 0.00098,
+   "damp": 0.009999999776482582
+  },
+  "layers.37.attention.to_out.0": {
+   "rtn": 0.006861,
+   "gptq": 0.000738,
+   "damp": 0.009999999776482582
+  },
+  "layers.37.feed_forward.w1": {
+   "rtn": 0.00764,
+   "gptq": 0.001253,
+   "damp": 0.009999999776482582
+  },
+  "layers.37.feed_forward.w2": {
+   "rtn": 0.005626,
+   "gptq": 1.2e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.38.attention.to_q": {
+   "rtn": 0.006079,
+   "gptq": 0.001085,
+   "damp": 0.009999999776482582
+  },
+  "layers.38.attention.to_out.0": {
+   "rtn": 0.006825,
+   "gptq": 0.000734,
+   "damp": 0.009999999776482582
+  },
+  "layers.38.feed_forward.w1": {
+   "rtn": 0.007807,
+   "gptq": 0.001301,
+   "damp": 0.009999999776482582
+  },
+  "layers.38.feed_forward.w2": {
+   "rtn": 0.006375,
+   "gptq": 1.3e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.38.cross_attention.to_q": {
+   "rtn": 0.131004,
+   "gptq": 0.001538,
+   "damp": 0.009999999776482582
+  },
+  "layers.38.cross_attention.to_k": {
+   "rtn": 0.013637,
+   "gptq": 0.000456,
+   "damp": 0.009999999776482582
+  },
+  "layers.38.cross_attention.to_out.0": {
+   "rtn": 0.00219,
+   "gptq": 2.3e-05,
+   "damp": 0.009999999776482582
+  },
+  "layers.39.attention.to_q": {
+   "rtn": 0.005221,
+   "gptq": 0.000919,
+   "damp": 0.009999999776482582
+  },
+  "layers.39.attention.to_out.0": {
+   "rtn": 0.008581,
+   "gptq": 0.000967,
+   "damp": 0.009999999776482582
+  },
+  "layers.39.feed_forward.w1": {
+   "rtn": 0.007396,
+   "gptq": 0.001271,
+   "damp": 0.009999999776482582
+  },
+  "layers.39.feed_forward.w2": {
+   "rtn": 0.004407,
+   "gptq": 5e-06,
+   "damp": 0.09999999403953552
+  },
+  "layers.40.attention.to_q": {
+   "rtn": 0.00542,
+   "gptq": 0.000966,
+   "damp": 0.009999999776482582
+  },
+  "layers.40.attention.to_out.0": {
+   "rtn": 0.009094,
+   "gptq": 0.001372,
+   "damp": 0.009999999776482582
+  },
+  "layers.40.feed_forward.w1": {
+   "rtn": 0.00756,
+   "gptq": 0.001386,
+   "damp": 0.009999999776482582
+  },
+  "layers.40.feed_forward.w2": {
+   "rtn": 0.004344,
+   "gptq": 1.8e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.41.attention.to_q": {
+   "rtn": 0.005491,
+   "gptq": 0.001093,
+   "damp": 0.009999999776482582
+  },
+  "layers.41.attention.to_out.0": {
+   "rtn": 0.008482,
+   "gptq": 0.001194,
+   "damp": 0.009999999776482582
+  },
+  "layers.41.feed_forward.w1": {
+   "rtn": 0.007583,
+   "gptq": 0.001545,
+   "damp": 0.009999999776482582
+  },
+  "layers.41.feed_forward.w2": {
+   "rtn": 0.004442,
+   "gptq": 2.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.42.attention.to_q": {
+   "rtn": 0.005221,
+   "gptq": 0.00112,
+   "damp": 0.009999999776482582
+  },
+  "layers.42.attention.to_out.0": {
+   "rtn": 0.009337,
+   "gptq": 0.001033,
+   "damp": 0.009999999776482582
+  },
+  "layers.42.feed_forward.w1": {
+   "rtn": 0.006992,
+   "gptq": 0.001551,
+   "damp": 0.009999999776482582
+  },
+  "layers.42.feed_forward.w2": {
+   "rtn": 0.004034,
+   "gptq": 2.4e-05,
+   "damp": 0.09999999403953552
+  },
+  "layers.43.attention.to_q": {
+   "rtn": 0.003794,
+   "gptq": 0.000911,
+   "damp": 0.009999999776482582
+  },
+  "layers.43.attention.to_out.0": {
+   "rtn": 0.018569,
+   "gptq": 0.002512,
+   "damp": 0.009999999776482582
+  },
+  "layers.43.feed_forward.w1": {
+   "rtn": 0.007156,
+   "gptq": 0.001795,
+   "damp": 0.009999999776482582
+  },
+  "layers.43.feed_forward.w2": {
+   "rtn": 0.040363,
+   "gptq": 0.001142,
+   "damp": 0.009999999776482582
+  }
+ },
+ "mean": {
+  "rtn": 0.008424,
+  "gptq": 0.000674
+ },
+ "seconds": 1057.4
+}

transformer/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a7e6eccf0f85b570301e8c8eab769ba3c20b14d3ea17fdde4a67b52b7227260
+size 4329176674

transformer/orbitquant_codebooks.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:354b4de89d7dbecbca33cba491af43aa0e94a9f469658c384d3743dbdb99c288
+size 860

transformer/orbitquant_manifest.json

@@ -0,0 +1,3950 @@
+{
+  "artifact_format": "orbitquant-v1",
+  "source_model_id": "briaai/Fibo-2.0-dev",
+  "source_revision": "fb2b781f9e9e5572a6c348a5af7a4d155a2b1ed0",
+  "source_license": "cc-by-nc-4.0",
+  "quant_method": "orbitquant",
+  "paper": "https://arxiv.org/abs/2607.02461",
+  "weight_bits": 4,
+  "activation_bits": 4,
+  "rotation": "rpbh",
+  "rotation_seed": 0,
+  "block_size": "paper",
+  "block_size_policy": "largest_power_of_two_dividing_dim",
+  "codebook": "lloyd_max",
+  "codebook_version": 2,
+  "row_norm_dtype": "bfloat16",
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "auto",
+  "activation_eps": 1e-10,
+  "adaln_group_size": 64,
+  "quantization_device": "cuda",
+  "weight_quantization_backend": "triton_cuda",
+  "quantization_staging_mode": "streaming",
+  "target_policy": "universal",
+  "adaln_policy": "int4_rtn_group64_bf16_activation",
+  "quantized_modules": [
+    "noise_refiner.0.attention.to_q",
+    "noise_refiner.0.attention.to_k",
+    "noise_refiner.0.attention.to_v",
+    "noise_refiner.0.attention.to_out.0",
+    "noise_refiner.0.feed_forward.w1",
+    "noise_refiner.0.feed_forward.w2",
+    "noise_refiner.0.feed_forward.w3",
+    "noise_refiner.1.attention.to_q",
+    "noise_refiner.1.attention.to_k",
+    "noise_refiner.1.attention.to_v",
+    "noise_refiner.1.attention.to_out.0",
+    "noise_refiner.1.feed_forward.w1",
+    "noise_refiner.1.feed_forward.w2",
+    "noise_refiner.1.feed_forward.w3",
+    "perceiver.layers.0.cross_attention.to_q",
+    "perceiver.layers.0.cross_attention.to_k",
+    "perceiver.layers.0.cross_attention.to_v",
+    "perceiver.layers.0.cross_attention.to_out.0",
+    "perceiver.layers.0.self_attention.to_q",
+    "perceiver.layers.0.self_attention.to_k",
+    "perceiver.layers.0.self_attention.to_v",
+    "perceiver.layers.0.self_attention.to_out.0",
+    "perceiver.layers.0.ffn.w1",
+    "perceiver.layers.0.ffn.w2",
+    "perceiver.layers.0.ffn.w3",
+    "perceiver.layers.1.cross_attention.to_q",
+    "perceiver.layers.1.cross_attention.to_k",
+    "perceiver.layers.1.cross_attention.to_v",
+    "perceiver.layers.1.cross_attention.to_out.0",
+    "perceiver.layers.1.self_attention.to_q",
+    "perceiver.layers.1.self_attention.to_k",
+    "perceiver.layers.1.self_attention.to_v",
+    "perceiver.layers.1.self_attention.to_out.0",
+    "perceiver.layers.1.ffn.w1",
+    "perceiver.layers.1.ffn.w2",
+    "perceiver.layers.1.ffn.w3",
+    "layers.0.attention.to_q",
+    "layers.0.attention.to_k",
+    "layers.0.attention.to_v",
+    "layers.0.attention.to_out.0",
+    "layers.0.feed_forward.w1",
+    "layers.0.feed_forward.w2",
+    "layers.0.feed_forward.w3",
+    "layers.1.attention.to_q",
+    "layers.1.attention.to_k",
+    "layers.1.attention.to_v",
+    "layers.1.attention.to_out.0",
+    "layers.1.feed_forward.w1",
+    "layers.1.feed_forward.w2",
+    "layers.1.feed_forward.w3",
+    "layers.2.attention.to_q",
+    "layers.2.attention.to_k",
+    "layers.2.attention.to_v",
+    "layers.2.attention.to_out.0",
+    "layers.2.feed_forward.w1",
+    "layers.2.feed_forward.w2",
+    "layers.2.feed_forward.w3",
+    "layers.3.attention.to_q",
+    "layers.3.attention.to_k",
+    "layers.3.attention.to_v",
+    "layers.3.attention.to_out.0",
+    "layers.3.feed_forward.w1",
+    "layers.3.feed_forward.w2",
+    "layers.3.feed_forward.w3",
+    "layers.4.attention.to_q",
+    "layers.4.attention.to_k",
+    "layers.4.attention.to_v",
+    "layers.4.attention.to_out.0",
+    "layers.4.feed_forward.w1",
+    "layers.4.feed_forward.w2",
+    "layers.4.feed_forward.w3",
+    "layers.5.attention.to_q",
+    "layers.5.attention.to_k",
+    "layers.5.attention.to_v",
+    "layers.5.attention.to_out.0",
+    "layers.5.feed_forward.w1",
+    "layers.5.feed_forward.w2",
+    "layers.5.feed_forward.w3",
+    "layers.6.attention.to_q",
+    "layers.6.attention.to_k",
+    "layers.6.attention.to_v",
+    "layers.6.attention.to_out.0",
+    "layers.6.feed_forward.w1",
+    "layers.6.feed_forward.w2",
+    "layers.6.feed_forward.w3",
+    "layers.6.cross_attention.to_q",
+    "layers.6.cross_attention.to_k",
+    "layers.6.cross_attention.to_v",
+    "layers.6.cross_attention.to_out.0",
+    "layers.7.attention.to_q",
+    "layers.7.attention.to_k",
+    "layers.7.attention.to_v",
+    "layers.7.attention.to_out.0",
+    "layers.7.feed_forward.w1",
+    "layers.7.feed_forward.w2",
+    "layers.7.feed_forward.w3",
+    "layers.8.attention.to_q",
+    "layers.8.attention.to_k",
+    "layers.8.attention.to_v",
+    "layers.8.attention.to_out.0",
+    "layers.8.feed_forward.w1",
+    "layers.8.feed_forward.w2",
+    "layers.8.feed_forward.w3",
+    "layers.9.attention.to_q",
+    "layers.9.attention.to_k",
+    "layers.9.attention.to_v",
+    "layers.9.attention.to_out.0",
+    "layers.9.feed_forward.w1",
+    "layers.9.feed_forward.w2",
+    "layers.9.feed_forward.w3",
+    "layers.10.attention.to_q",
+    "layers.10.attention.to_k",
+    "layers.10.attention.to_v",
+    "layers.10.attention.to_out.0",
+    "layers.10.feed_forward.w1",
+    "layers.10.feed_forward.w2",
+    "layers.10.feed_forward.w3",
+    "layers.11.attention.to_q",
+    "layers.11.attention.to_k",
+    "layers.11.attention.to_v",
+    "layers.11.attention.to_out.0",
+    "layers.11.feed_forward.w1",
+    "layers.11.feed_forward.w2",
+    "layers.11.feed_forward.w3",
+    "layers.12.attention.to_q",
+    "layers.12.attention.to_k",
+    "layers.12.attention.to_v",
+    "layers.12.attention.to_out.0",
+    "layers.12.feed_forward.w1",
+    "layers.12.feed_forward.w2",
+    "layers.12.feed_forward.w3",
+    "layers.13.attention.to_q",
+    "layers.13.attention.to_k",
+    "layers.13.attention.to_v",
+    "layers.13.attention.to_out.0",
+    "layers.13.feed_forward.w1",
+    "layers.13.feed_forward.w2",
+    "layers.13.feed_forward.w3",
+    "layers.14.attention.to_q",
+    "layers.14.attention.to_k",
+    "layers.14.attention.to_v",
+    "layers.14.attention.to_out.0",
+    "layers.14.feed_forward.w1",
+    "layers.14.feed_forward.w2",
+    "layers.14.feed_forward.w3",
+    "layers.14.cross_attention.to_q",
+    "layers.14.cross_attention.to_k",
+    "layers.14.cross_attention.to_v",
+    "layers.14.cross_attention.to_out.0",
+    "layers.15.attention.to_q",
+    "layers.15.attention.to_k",
+    "layers.15.attention.to_v",
+    "layers.15.attention.to_out.0",
+    "layers.15.feed_forward.w1",
+    "layers.15.feed_forward.w2",
+    "layers.15.feed_forward.w3",
+    "layers.16.attention.to_q",
+    "layers.16.attention.to_k",
+    "layers.16.attention.to_v",
+    "layers.16.attention.to_out.0",
+    "layers.16.feed_forward.w1",
+    "layers.16.feed_forward.w2",
+    "layers.16.feed_forward.w3",
+    "layers.17.attention.to_q",
+    "layers.17.attention.to_k",
+    "layers.17.attention.to_v",
+    "layers.17.attention.to_out.0",
+    "layers.17.feed_forward.w1",
+    "layers.17.feed_forward.w2",
+    "layers.17.feed_forward.w3",
+    "layers.18.attention.to_q",
+    "layers.18.attention.to_k",
+    "layers.18.attention.to_v",
+    "layers.18.attention.to_out.0",
+    "layers.18.feed_forward.w1",
+    "layers.18.feed_forward.w2",
+    "layers.18.feed_forward.w3",
+    "layers.19.attention.to_q",
+    "layers.19.attention.to_k",
+    "layers.19.attention.to_v",
+    "layers.19.attention.to_out.0",
+    "layers.19.feed_forward.w1",
+    "layers.19.feed_forward.w2",
+    "layers.19.feed_forward.w3",
+    "layers.20.attention.to_q",
+    "layers.20.attention.to_k",
+    "layers.20.attention.to_v",
+    "layers.20.attention.to_out.0",
+    "layers.20.feed_forward.w1",
+    "layers.20.feed_forward.w2",
+    "layers.20.feed_forward.w3",
+    "layers.21.attention.to_q",
+    "layers.21.attention.to_k",
+    "layers.21.attention.to_v",
+    "layers.21.attention.to_out.0",
+    "layers.21.feed_forward.w1",
+    "layers.21.feed_forward.w2",
+    "layers.21.feed_forward.w3",
+    "layers.22.attention.to_q",
+    "layers.22.attention.to_k",
+    "layers.22.attention.to_v",
+    "layers.22.attention.to_out.0",
+    "layers.22.feed_forward.w1",
+    "layers.22.feed_forward.w2",
+    "layers.22.feed_forward.w3",
+    "layers.22.cross_attention.to_q",
+    "layers.22.cross_attention.to_k",
+    "layers.22.cross_attention.to_v",
+    "layers.22.cross_attention.to_out.0",
+    "layers.23.attention.to_q",
+    "layers.23.attention.to_k",
+    "layers.23.attention.to_v",
+    "layers.23.attention.to_out.0",
+    "layers.23.feed_forward.w1",
+    "layers.23.feed_forward.w2",
+    "layers.23.feed_forward.w3",
+    "layers.24.attention.to_q",
+    "layers.24.attention.to_k",
+    "layers.24.attention.to_v",
+    "layers.24.attention.to_out.0",
+    "layers.24.feed_forward.w1",
+    "layers.24.feed_forward.w2",
+    "layers.24.feed_forward.w3",
+    "layers.25.attention.to_q",
+    "layers.25.attention.to_k",
+    "layers.25.attention.to_v",
+    "layers.25.attention.to_out.0",
+    "layers.25.feed_forward.w1",
+    "layers.25.feed_forward.w2",
+    "layers.25.feed_forward.w3",
+    "layers.26.attention.to_q",
+    "layers.26.attention.to_k",
+    "layers.26.attention.to_v",
+    "layers.26.attention.to_out.0",
+    "layers.26.feed_forward.w1",
+    "layers.26.feed_forward.w2",
+    "layers.26.feed_forward.w3",
+    "layers.27.attention.to_q",
+    "layers.27.attention.to_k",
+    "layers.27.attention.to_v",
+    "layers.27.attention.to_out.0",
+    "layers.27.feed_forward.w1",
+    "layers.27.feed_forward.w2",
+    "layers.27.feed_forward.w3",
+    "layers.28.attention.to_q",
+    "layers.28.attention.to_k",
+    "layers.28.attention.to_v",
+    "layers.28.attention.to_out.0",
+    "layers.28.feed_forward.w1",
+    "layers.28.feed_forward.w2",
+    "layers.28.feed_forward.w3",
+    "layers.29.attention.to_q",
+    "layers.29.attention.to_k",
+    "layers.29.attention.to_v",
+    "layers.29.attention.to_out.0",
+    "layers.29.feed_forward.w1",
+    "layers.29.feed_forward.w2",
+    "layers.29.feed_forward.w3",
+    "layers.30.attention.to_q",
+    "layers.30.attention.to_k",
+    "layers.30.attention.to_v",
+    "layers.30.attention.to_out.0",
+    "layers.30.feed_forward.w1",
+    "layers.30.feed_forward.w2",
+    "layers.30.feed_forward.w3",
+    "layers.30.cross_attention.to_q",
+    "layers.30.cross_attention.to_k",
+    "layers.30.cross_attention.to_v",
+    "layers.30.cross_attention.to_out.0",
+    "layers.31.attention.to_q",
+    "layers.31.attention.to_k",
+    "layers.31.attention.to_v",
+    "layers.31.attention.to_out.0",
+    "layers.31.feed_forward.w1",
+    "layers.31.feed_forward.w2",
+    "layers.31.feed_forward.w3",
+    "layers.32.attention.to_q",
+    "layers.32.attention.to_k",
+    "layers.32.attention.to_v",
+    "layers.32.attention.to_out.0",
+    "layers.32.feed_forward.w1",
+    "layers.32.feed_forward.w2",
+    "layers.32.feed_forward.w3",
+    "layers.33.attention.to_q",
+    "layers.33.attention.to_k",
+    "layers.33.attention.to_v",
+    "layers.33.attention.to_out.0",
+    "layers.33.feed_forward.w1",
+    "layers.33.feed_forward.w2",
+    "layers.33.feed_forward.w3",
+    "layers.34.attention.to_q",
+    "layers.34.attention.to_k",
+    "layers.34.attention.to_v",
+    "layers.34.attention.to_out.0",
+    "layers.34.feed_forward.w1",
+    "layers.34.feed_forward.w2",
+    "layers.34.feed_forward.w3",
+    "layers.35.attention.to_q",
+    "layers.35.attention.to_k",
+    "layers.35.attention.to_v",
+    "layers.35.attention.to_out.0",
+    "layers.35.feed_forward.w1",
+    "layers.35.feed_forward.w2",
+    "layers.35.feed_forward.w3",
+    "layers.36.attention.to_q",
+    "layers.36.attention.to_k",
+    "layers.36.attention.to_v",
+    "layers.36.attention.to_out.0",
+    "layers.36.feed_forward.w1",
+    "layers.36.feed_forward.w2",
+    "layers.36.feed_forward.w3",
+    "layers.37.attention.to_q",
+    "layers.37.attention.to_k",
+    "layers.37.attention.to_v",
+    "layers.37.attention.to_out.0",
+    "layers.37.feed_forward.w1",
+    "layers.37.feed_forward.w2",
+    "layers.37.feed_forward.w3",
+    "layers.38.attention.to_q",
+    "layers.38.attention.to_k",
+    "layers.38.attention.to_v",
+    "layers.38.attention.to_out.0",
+    "layers.38.feed_forward.w1",
+    "layers.38.feed_forward.w2",
+    "layers.38.feed_forward.w3",
+    "layers.38.cross_attention.to_q",
+    "layers.38.cross_attention.to_k",
+    "layers.38.cross_attention.to_v",
+    "layers.38.cross_attention.to_out.0",
+    "layers.39.attention.to_q",
+    "layers.39.attention.to_k",
+    "layers.39.attention.to_v",
+    "layers.39.attention.to_out.0",
+    "layers.39.feed_forward.w1",
+    "layers.39.feed_forward.w2",
+    "layers.39.feed_forward.w3",
+    "layers.40.attention.to_q",
+    "layers.40.attention.to_k",
+    "layers.40.attention.to_v",
+    "layers.40.attention.to_out.0",
+    "layers.40.feed_forward.w1",
+    "layers.40.feed_forward.w2",
+    "layers.40.feed_forward.w3",
+    "layers.41.attention.to_q",
+    "layers.41.attention.to_k",
+    "layers.41.attention.to_v",
+    "layers.41.attention.to_out.0",
+    "layers.41.feed_forward.w1",
+    "layers.41.feed_forward.w2",
+    "layers.41.feed_forward.w3",
+    "layers.42.attention.to_q",
+    "layers.42.attention.to_k",
+    "layers.42.attention.to_v",
+    "layers.42.attention.to_out.0",
+    "layers.42.feed_forward.w1",
+    "layers.42.feed_forward.w2",
+    "layers.42.feed_forward.w3",
+    "layers.43.attention.to_q",
+    "layers.43.attention.to_k",
+    "layers.43.attention.to_v",
+    "layers.43.attention.to_out.0",
+    "layers.43.feed_forward.w1",
+    "layers.43.feed_forward.w2",
+    "layers.43.feed_forward.w3"
+  ],
+  "adaln_modules": [],
+  "skipped_modules": [
+    "x_embedder",
+    "t_embedder.mlp.0",
+    "t_embedder.mlp.2",
+    "noise_refiner.0.adaLN_modulation.0",
+    "noise_refiner.1.adaLN_modulation.0",
+    "layers.0.adaLN_modulation.0",
+    "layers.1.adaLN_modulation.0",
+    "layers.2.adaLN_modulation.0",
+    "layers.3.adaLN_modulation.0",
+    "layers.4.adaLN_modulation.0",
+    "layers.5.adaLN_modulation.0",
+    "layers.6.adaLN_modulation.0",
+    "layers.7.adaLN_modulation.0",
+    "layers.8.adaLN_modulation.0",
+    "layers.9.adaLN_modulation.0",
+    "layers.10.adaLN_modulation.0",
+    "layers.11.adaLN_modulation.0",
+    "layers.12.adaLN_modulation.0",
+    "layers.13.adaLN_modulation.0",
+    "layers.14.adaLN_modulation.0",
+    "layers.15.adaLN_modulation.0",
+    "layers.16.adaLN_modulation.0",
+    "layers.17.adaLN_modulation.0",
+    "layers.18.adaLN_modulation.0",
+    "layers.19.adaLN_modulation.0",
+    "layers.20.adaLN_modulation.0",
+    "layers.21.adaLN_modulation.0",
+    "layers.22.adaLN_modulation.0",
+    "layers.23.adaLN_modulation.0",
+    "layers.24.adaLN_modulation.0",
+    "layers.25.adaLN_modulation.0",
+    "layers.26.adaLN_modulation.0",
+    "layers.27.adaLN_modulation.0",
+    "layers.28.adaLN_modulation.0",
+    "layers.29.adaLN_modulation.0",
+    "layers.30.adaLN_modulation.0",
+    "layers.31.adaLN_modulation.0",
+    "layers.32.adaLN_modulation.0",
+    "layers.33.adaLN_modulation.0",
+    "layers.34.adaLN_modulation.0",
+    "layers.35.adaLN_modulation.0",
+    "layers.36.adaLN_modulation.0",
+    "layers.37.adaLN_modulation.0",
+    "layers.38.adaLN_modulation.0",
+    "layers.39.adaLN_modulation.0",
+    "layers.40.adaLN_modulation.0",
+    "layers.41.adaLN_modulation.0",
+    "layers.42.adaLN_modulation.0",
+    "layers.43.adaLN_modulation.0",
+    "final_layer.linear",
+    "final_layer.adaLN_modulation.1"
+  ],
+  "module_shapes": {
+    "x_embedder.weight": [
+      3584,
+      128
+    ],
+    "x_embedder.bias": [
+      3584
+    ],
+    "t_embedder.mlp.0.weight": [
+      1024,
+      256
+    ],
+    "t_embedder.mlp.0.bias": [
+      1024
+    ],
+    "t_embedder.mlp.2.weight": [
+      256,
+      1024
+    ],
+    "t_embedder.mlp.2.bias": [
+      256
+    ],
+    "noise_refiner.0.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.0.attention.to_q.row_norms": [
+      3584
+    ],
+    "noise_refiner.0.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.0.attention.to_k.row_norms": [
+      3584
+    ],
+    "noise_refiner.0.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.0.attention.to_v.row_norms": [
+      3584
+    ],
+    "noise_refiner.0.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.0.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "noise_refiner.0.attention.norm_q.weight": [
+      128
+    ],
+    "noise_refiner.0.attention.norm_k.weight": [
+      128
+    ],
+    "noise_refiner.0.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.0.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "noise_refiner.0.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.0.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "noise_refiner.0.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.0.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "noise_refiner.0.attention_norm1.weight": [
+      3584
+    ],
+    "noise_refiner.0.ffn_norm1.weight": [
+      3584
+    ],
+    "noise_refiner.0.attention_norm2.weight": [
+      3584
+    ],
+    "noise_refiner.0.ffn_norm2.weight": [
+      3584
+    ],
+    "noise_refiner.0.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "noise_refiner.0.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "noise_refiner.1.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.1.attention.to_q.row_norms": [
+      3584
+    ],
+    "noise_refiner.1.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.1.attention.to_k.row_norms": [
+      3584
+    ],
+    "noise_refiner.1.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.1.attention.to_v.row_norms": [
+      3584
+    ],
+    "noise_refiner.1.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "noise_refiner.1.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "noise_refiner.1.attention.norm_q.weight": [
+      128
+    ],
+    "noise_refiner.1.attention.norm_k.weight": [
+      128
+    ],
+    "noise_refiner.1.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.1.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "noise_refiner.1.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.1.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "noise_refiner.1.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "noise_refiner.1.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "noise_refiner.1.attention_norm1.weight": [
+      3584
+    ],
+    "noise_refiner.1.ffn_norm1.weight": [
+      3584
+    ],
+    "noise_refiner.1.attention_norm2.weight": [
+      3584
+    ],
+    "noise_refiner.1.ffn_norm2.weight": [
+      3584
+    ],
+    "noise_refiner.1.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "noise_refiner.1.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "perceiver.latent_queries": [
+      512,
+      3584
+    ],
+    "perceiver.layers.0.cross_attn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.0.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "perceiver.layers.0.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "perceiver.layers.0.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.cross_attention.norm_q.weight": [
+      128
+    ],
+    "perceiver.layers.0.cross_attention.norm_k.weight": [
+      128
+    ],
+    "perceiver.layers.0.self_attn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.0.self_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.self_attention.to_q.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.self_attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.self_attention.to_k.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.self_attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.self_attention.to_v.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.self_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.0.self_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.self_attention.norm_q.weight": [
+      128
+    ],
+    "perceiver.layers.0.self_attention.norm_k.weight": [
+      128
+    ],
+    "perceiver.layers.0.ffn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.0.ffn.w1.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.0.ffn.w1.row_norms": [
+      9536
+    ],
+    "perceiver.layers.0.ffn.w2.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.0.ffn.w2.row_norms": [
+      3584
+    ],
+    "perceiver.layers.0.ffn.w3.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.0.ffn.w3.row_norms": [
+      9536
+    ],
+    "perceiver.layers.1.cross_attn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.1.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "perceiver.layers.1.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "perceiver.layers.1.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.cross_attention.norm_q.weight": [
+      128
+    ],
+    "perceiver.layers.1.cross_attention.norm_k.weight": [
+      128
+    ],
+    "perceiver.layers.1.self_attn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.1.self_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.self_attention.to_q.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.self_attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.self_attention.to_k.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.self_attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.self_attention.to_v.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.self_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "perceiver.layers.1.self_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.self_attention.norm_q.weight": [
+      128
+    ],
+    "perceiver.layers.1.self_attention.norm_k.weight": [
+      128
+    ],
+    "perceiver.layers.1.ffn_norm.weight": [
+      3584
+    ],
+    "perceiver.layers.1.ffn.w1.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.1.ffn.w1.row_norms": [
+      9536
+    ],
+    "perceiver.layers.1.ffn.w2.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.1.ffn.w2.row_norms": [
+      3584
+    ],
+    "perceiver.layers.1.ffn.w3.packed_weight_indices": [
+      17088512
+    ],
+    "perceiver.layers.1.ffn.w3.row_norms": [
+      9536
+    ],
+    "perceiver.norm_out.weight": [
+      3584
+    ],
+    "layers.0.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.0.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.0.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.0.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.0.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.0.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.0.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.0.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.0.attention.norm_q.weight": [
+      128
+    ],
+    "layers.0.attention.norm_k.weight": [
+      128
+    ],
+    "layers.0.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.0.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.0.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.0.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.0.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.0.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.0.attention_norm1.weight": [
+      3584
+    ],
+    "layers.0.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.0.attention_norm2.weight": [
+      3584
+    ],
+    "layers.0.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.0.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.0.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.1.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.1.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.1.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.1.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.1.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.1.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.1.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.1.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.1.attention.norm_q.weight": [
+      128
+    ],
+    "layers.1.attention.norm_k.weight": [
+      128
+    ],
+    "layers.1.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.1.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.1.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.1.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.1.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.1.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.1.attention_norm1.weight": [
+      3584
+    ],
+    "layers.1.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.1.attention_norm2.weight": [
+      3584
+    ],
+    "layers.1.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.1.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.1.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.2.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.2.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.2.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.2.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.2.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.2.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.2.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.2.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.2.attention.norm_q.weight": [
+      128
+    ],
+    "layers.2.attention.norm_k.weight": [
+      128
+    ],
+    "layers.2.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.2.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.2.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.2.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.2.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.2.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.2.attention_norm1.weight": [
+      3584
+    ],
+    "layers.2.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.2.attention_norm2.weight": [
+      3584
+    ],
+    "layers.2.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.2.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.2.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.3.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.3.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.3.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.3.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.3.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.3.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.3.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.3.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.3.attention.norm_q.weight": [
+      128
+    ],
+    "layers.3.attention.norm_k.weight": [
+      128
+    ],
+    "layers.3.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.3.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.3.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.3.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.3.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.3.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.3.attention_norm1.weight": [
+      3584
+    ],
+    "layers.3.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.3.attention_norm2.weight": [
+      3584
+    ],
+    "layers.3.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.3.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.3.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.4.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.4.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.4.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.4.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.4.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.4.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.4.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.4.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.4.attention.norm_q.weight": [
+      128
+    ],
+    "layers.4.attention.norm_k.weight": [
+      128
+    ],
+    "layers.4.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.4.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.4.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.4.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.4.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.4.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.4.attention_norm1.weight": [
+      3584
+    ],
+    "layers.4.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.4.attention_norm2.weight": [
+      3584
+    ],
+    "layers.4.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.4.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.4.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.5.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.5.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.5.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.5.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.5.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.5.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.5.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.5.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.5.attention.norm_q.weight": [
+      128
+    ],
+    "layers.5.attention.norm_k.weight": [
+      128
+    ],
+    "layers.5.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.5.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.5.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.5.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.5.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.5.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.5.attention_norm1.weight": [
+      3584
+    ],
+    "layers.5.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.5.attention_norm2.weight": [
+      3584
+    ],
+    "layers.5.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.5.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.5.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.6.cross_attn_gate": [
+      1
+    ],
+    "layers.6.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.6.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.6.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.6.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.6.attention.norm_q.weight": [
+      128
+    ],
+    "layers.6.attention.norm_k.weight": [
+      128
+    ],
+    "layers.6.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.6.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.6.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.6.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.6.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.6.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.6.attention_norm1.weight": [
+      3584
+    ],
+    "layers.6.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.6.attention_norm2.weight": [
+      3584
+    ],
+    "layers.6.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.6.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.6.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.6.cross_attn_norm1.weight": [
+      3584
+    ],
+    "layers.6.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.6.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "layers.6.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.6.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "layers.6.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.6.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.6.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.6.cross_attention.norm_q.weight": [
+      128
+    ],
+    "layers.6.cross_attention.norm_k.weight": [
+      128
+    ],
+    "layers.6.cross_attn_norm2.weight": [
+      3584
+    ],
+    "layers.7.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.7.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.7.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.7.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.7.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.7.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.7.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.7.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.7.attention.norm_q.weight": [
+      128
+    ],
+    "layers.7.attention.norm_k.weight": [
+      128
+    ],
+    "layers.7.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.7.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.7.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.7.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.7.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.7.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.7.attention_norm1.weight": [
+      3584
+    ],
+    "layers.7.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.7.attention_norm2.weight": [
+      3584
+    ],
+    "layers.7.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.7.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.7.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.8.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.8.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.8.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.8.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.8.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.8.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.8.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.8.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.8.attention.norm_q.weight": [
+      128
+    ],
+    "layers.8.attention.norm_k.weight": [
+      128
+    ],
+    "layers.8.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.8.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.8.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.8.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.8.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.8.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.8.attention_norm1.weight": [
+      3584
+    ],
+    "layers.8.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.8.attention_norm2.weight": [
+      3584
+    ],
+    "layers.8.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.8.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.8.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.9.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.9.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.9.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.9.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.9.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.9.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.9.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.9.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.9.attention.norm_q.weight": [
+      128
+    ],
+    "layers.9.attention.norm_k.weight": [
+      128
+    ],
+    "layers.9.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.9.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.9.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.9.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.9.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.9.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.9.attention_norm1.weight": [
+      3584
+    ],
+    "layers.9.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.9.attention_norm2.weight": [
+      3584
+    ],
+    "layers.9.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.9.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.9.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.10.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.10.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.10.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.10.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.10.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.10.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.10.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.10.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.10.attention.norm_q.weight": [
+      128
+    ],
+    "layers.10.attention.norm_k.weight": [
+      128
+    ],
+    "layers.10.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.10.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.10.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.10.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.10.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.10.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.10.attention_norm1.weight": [
+      3584
+    ],
+    "layers.10.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.10.attention_norm2.weight": [
+      3584
+    ],
+    "layers.10.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.10.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.10.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.11.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.11.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.11.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.11.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.11.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.11.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.11.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.11.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.11.attention.norm_q.weight": [
+      128
+    ],
+    "layers.11.attention.norm_k.weight": [
+      128
+    ],
+    "layers.11.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.11.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.11.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.11.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.11.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.11.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.11.attention_norm1.weight": [
+      3584
+    ],
+    "layers.11.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.11.attention_norm2.weight": [
+      3584
+    ],
+    "layers.11.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.11.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.11.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.12.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.12.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.12.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.12.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.12.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.12.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.12.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.12.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.12.attention.norm_q.weight": [
+      128
+    ],
+    "layers.12.attention.norm_k.weight": [
+      128
+    ],
+    "layers.12.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.12.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.12.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.12.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.12.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.12.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.12.attention_norm1.weight": [
+      3584
+    ],
+    "layers.12.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.12.attention_norm2.weight": [
+      3584
+    ],
+    "layers.12.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.12.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.12.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.13.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.13.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.13.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.13.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.13.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.13.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.13.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.13.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.13.attention.norm_q.weight": [
+      128
+    ],
+    "layers.13.attention.norm_k.weight": [
+      128
+    ],
+    "layers.13.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.13.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.13.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.13.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.13.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.13.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.13.attention_norm1.weight": [
+      3584
+    ],
+    "layers.13.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.13.attention_norm2.weight": [
+      3584
+    ],
+    "layers.13.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.13.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.13.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.14.cross_attn_gate": [
+      1
+    ],
+    "layers.14.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.14.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.14.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.14.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.14.attention.norm_q.weight": [
+      128
+    ],
+    "layers.14.attention.norm_k.weight": [
+      128
+    ],
+    "layers.14.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.14.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.14.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.14.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.14.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.14.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.14.attention_norm1.weight": [
+      3584
+    ],
+    "layers.14.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.14.attention_norm2.weight": [
+      3584
+    ],
+    "layers.14.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.14.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.14.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.14.cross_attn_norm1.weight": [
+      3584
+    ],
+    "layers.14.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.14.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "layers.14.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.14.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "layers.14.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.14.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.14.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.14.cross_attention.norm_q.weight": [
+      128
+    ],
+    "layers.14.cross_attention.norm_k.weight": [
+      128
+    ],
+    "layers.14.cross_attn_norm2.weight": [
+      3584
+    ],
+    "layers.15.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.15.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.15.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.15.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.15.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.15.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.15.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.15.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.15.attention.norm_q.weight": [
+      128
+    ],
+    "layers.15.attention.norm_k.weight": [
+      128
+    ],
+    "layers.15.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.15.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.15.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.15.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.15.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.15.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.15.attention_norm1.weight": [
+      3584
+    ],
+    "layers.15.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.15.attention_norm2.weight": [
+      3584
+    ],
+    "layers.15.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.15.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.15.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.16.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.16.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.16.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.16.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.16.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.16.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.16.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.16.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.16.attention.norm_q.weight": [
+      128
+    ],
+    "layers.16.attention.norm_k.weight": [
+      128
+    ],
+    "layers.16.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.16.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.16.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.16.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.16.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.16.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.16.attention_norm1.weight": [
+      3584
+    ],
+    "layers.16.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.16.attention_norm2.weight": [
+      3584
+    ],
+    "layers.16.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.16.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.16.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.17.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.17.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.17.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.17.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.17.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.17.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.17.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.17.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.17.attention.norm_q.weight": [
+      128
+    ],
+    "layers.17.attention.norm_k.weight": [
+      128
+    ],
+    "layers.17.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.17.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.17.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.17.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.17.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.17.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.17.attention_norm1.weight": [
+      3584
+    ],
+    "layers.17.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.17.attention_norm2.weight": [
+      3584
+    ],
+    "layers.17.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.17.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.17.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.18.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.18.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.18.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.18.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.18.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.18.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.18.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.18.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.18.attention.norm_q.weight": [
+      128
+    ],
+    "layers.18.attention.norm_k.weight": [
+      128
+    ],
+    "layers.18.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.18.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.18.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.18.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.18.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.18.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.18.attention_norm1.weight": [
+      3584
+    ],
+    "layers.18.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.18.attention_norm2.weight": [
+      3584
+    ],
+    "layers.18.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.18.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.18.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.19.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.19.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.19.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.19.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.19.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.19.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.19.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.19.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.19.attention.norm_q.weight": [
+      128
+    ],
+    "layers.19.attention.norm_k.weight": [
+      128
+    ],
+    "layers.19.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.19.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.19.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.19.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.19.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.19.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.19.attention_norm1.weight": [
+      3584
+    ],
+    "layers.19.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.19.attention_norm2.weight": [
+      3584
+    ],
+    "layers.19.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.19.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.19.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.20.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.20.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.20.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.20.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.20.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.20.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.20.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.20.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.20.attention.norm_q.weight": [
+      128
+    ],
+    "layers.20.attention.norm_k.weight": [
+      128
+    ],
+    "layers.20.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.20.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.20.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.20.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.20.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.20.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.20.attention_norm1.weight": [
+      3584
+    ],
+    "layers.20.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.20.attention_norm2.weight": [
+      3584
+    ],
+    "layers.20.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.20.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.20.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.21.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.21.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.21.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.21.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.21.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.21.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.21.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.21.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.21.attention.norm_q.weight": [
+      128
+    ],
+    "layers.21.attention.norm_k.weight": [
+      128
+    ],
+    "layers.21.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.21.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.21.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.21.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.21.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.21.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.21.attention_norm1.weight": [
+      3584
+    ],
+    "layers.21.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.21.attention_norm2.weight": [
+      3584
+    ],
+    "layers.21.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.21.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.21.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.22.cross_attn_gate": [
+      1
+    ],
+    "layers.22.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.22.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.22.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.22.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.22.attention.norm_q.weight": [
+      128
+    ],
+    "layers.22.attention.norm_k.weight": [
+      128
+    ],
+    "layers.22.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.22.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.22.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.22.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.22.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.22.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.22.attention_norm1.weight": [
+      3584
+    ],
+    "layers.22.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.22.attention_norm2.weight": [
+      3584
+    ],
+    "layers.22.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.22.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.22.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.22.cross_attn_norm1.weight": [
+      3584
+    ],
+    "layers.22.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.22.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "layers.22.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.22.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "layers.22.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.22.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.22.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.22.cross_attention.norm_q.weight": [
+      128
+    ],
+    "layers.22.cross_attention.norm_k.weight": [
+      128
+    ],
+    "layers.22.cross_attn_norm2.weight": [
+      3584
+    ],
+    "layers.23.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.23.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.23.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.23.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.23.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.23.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.23.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.23.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.23.attention.norm_q.weight": [
+      128
+    ],
+    "layers.23.attention.norm_k.weight": [
+      128
+    ],
+    "layers.23.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.23.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.23.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.23.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.23.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.23.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.23.attention_norm1.weight": [
+      3584
+    ],
+    "layers.23.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.23.attention_norm2.weight": [
+      3584
+    ],
+    "layers.23.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.23.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.23.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.24.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.24.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.24.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.24.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.24.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.24.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.24.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.24.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.24.attention.norm_q.weight": [
+      128
+    ],
+    "layers.24.attention.norm_k.weight": [
+      128
+    ],
+    "layers.24.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.24.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.24.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.24.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.24.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.24.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.24.attention_norm1.weight": [
+      3584
+    ],
+    "layers.24.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.24.attention_norm2.weight": [
+      3584
+    ],
+    "layers.24.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.24.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.24.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.25.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.25.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.25.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.25.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.25.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.25.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.25.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.25.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.25.attention.norm_q.weight": [
+      128
+    ],
+    "layers.25.attention.norm_k.weight": [
+      128
+    ],
+    "layers.25.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.25.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.25.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.25.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.25.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.25.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.25.attention_norm1.weight": [
+      3584
+    ],
+    "layers.25.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.25.attention_norm2.weight": [
+      3584
+    ],
+    "layers.25.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.25.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.25.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.26.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.26.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.26.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.26.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.26.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.26.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.26.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.26.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.26.attention.norm_q.weight": [
+      128
+    ],
+    "layers.26.attention.norm_k.weight": [
+      128
+    ],
+    "layers.26.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.26.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.26.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.26.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.26.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.26.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.26.attention_norm1.weight": [
+      3584
+    ],
+    "layers.26.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.26.attention_norm2.weight": [
+      3584
+    ],
+    "layers.26.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.26.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.26.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.27.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.27.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.27.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.27.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.27.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.27.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.27.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.27.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.27.attention.norm_q.weight": [
+      128
+    ],
+    "layers.27.attention.norm_k.weight": [
+      128
+    ],
+    "layers.27.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.27.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.27.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.27.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.27.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.27.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.27.attention_norm1.weight": [
+      3584
+    ],
+    "layers.27.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.27.attention_norm2.weight": [
+      3584
+    ],
+    "layers.27.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.27.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.27.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.28.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.28.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.28.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.28.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.28.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.28.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.28.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.28.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.28.attention.norm_q.weight": [
+      128
+    ],
+    "layers.28.attention.norm_k.weight": [
+      128
+    ],
+    "layers.28.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.28.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.28.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.28.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.28.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.28.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.28.attention_norm1.weight": [
+      3584
+    ],
+    "layers.28.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.28.attention_norm2.weight": [
+      3584
+    ],
+    "layers.28.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.28.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.28.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.29.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.29.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.29.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.29.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.29.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.29.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.29.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.29.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.29.attention.norm_q.weight": [
+      128
+    ],
+    "layers.29.attention.norm_k.weight": [
+      128
+    ],
+    "layers.29.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.29.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.29.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.29.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.29.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.29.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.29.attention_norm1.weight": [
+      3584
+    ],
+    "layers.29.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.29.attention_norm2.weight": [
+      3584
+    ],
+    "layers.29.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.29.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.29.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.30.cross_attn_gate": [
+      1
+    ],
+    "layers.30.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.30.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.30.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.30.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.30.attention.norm_q.weight": [
+      128
+    ],
+    "layers.30.attention.norm_k.weight": [
+      128
+    ],
+    "layers.30.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.30.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.30.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.30.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.30.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.30.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.30.attention_norm1.weight": [
+      3584
+    ],
+    "layers.30.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.30.attention_norm2.weight": [
+      3584
+    ],
+    "layers.30.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.30.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.30.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.30.cross_attn_norm1.weight": [
+      3584
+    ],
+    "layers.30.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.30.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "layers.30.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.30.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "layers.30.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.30.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.30.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.30.cross_attention.norm_q.weight": [
+      128
+    ],
+    "layers.30.cross_attention.norm_k.weight": [
+      128
+    ],
+    "layers.30.cross_attn_norm2.weight": [
+      3584
+    ],
+    "layers.31.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.31.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.31.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.31.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.31.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.31.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.31.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.31.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.31.attention.norm_q.weight": [
+      128
+    ],
+    "layers.31.attention.norm_k.weight": [
+      128
+    ],
+    "layers.31.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.31.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.31.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.31.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.31.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.31.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.31.attention_norm1.weight": [
+      3584
+    ],
+    "layers.31.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.31.attention_norm2.weight": [
+      3584
+    ],
+    "layers.31.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.31.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.31.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.32.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.32.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.32.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.32.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.32.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.32.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.32.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.32.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.32.attention.norm_q.weight": [
+      128
+    ],
+    "layers.32.attention.norm_k.weight": [
+      128
+    ],
+    "layers.32.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.32.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.32.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.32.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.32.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.32.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.32.attention_norm1.weight": [
+      3584
+    ],
+    "layers.32.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.32.attention_norm2.weight": [
+      3584
+    ],
+    "layers.32.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.32.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.32.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.33.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.33.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.33.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.33.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.33.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.33.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.33.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.33.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.33.attention.norm_q.weight": [
+      128
+    ],
+    "layers.33.attention.norm_k.weight": [
+      128
+    ],
+    "layers.33.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.33.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.33.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.33.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.33.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.33.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.33.attention_norm1.weight": [
+      3584
+    ],
+    "layers.33.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.33.attention_norm2.weight": [
+      3584
+    ],
+    "layers.33.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.33.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.33.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.34.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.34.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.34.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.34.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.34.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.34.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.34.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.34.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.34.attention.norm_q.weight": [
+      128
+    ],
+    "layers.34.attention.norm_k.weight": [
+      128
+    ],
+    "layers.34.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.34.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.34.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.34.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.34.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.34.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.34.attention_norm1.weight": [
+      3584
+    ],
+    "layers.34.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.34.attention_norm2.weight": [
+      3584
+    ],
+    "layers.34.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.34.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.34.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.35.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.35.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.35.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.35.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.35.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.35.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.35.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.35.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.35.attention.norm_q.weight": [
+      128
+    ],
+    "layers.35.attention.norm_k.weight": [
+      128
+    ],
+    "layers.35.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.35.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.35.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.35.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.35.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.35.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.35.attention_norm1.weight": [
+      3584
+    ],
+    "layers.35.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.35.attention_norm2.weight": [
+      3584
+    ],
+    "layers.35.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.35.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.35.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.36.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.36.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.36.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.36.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.36.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.36.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.36.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.36.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.36.attention.norm_q.weight": [
+      128
+    ],
+    "layers.36.attention.norm_k.weight": [
+      128
+    ],
+    "layers.36.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.36.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.36.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.36.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.36.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.36.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.36.attention_norm1.weight": [
+      3584
+    ],
+    "layers.36.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.36.attention_norm2.weight": [
+      3584
+    ],
+    "layers.36.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.36.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.36.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.37.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.37.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.37.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.37.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.37.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.37.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.37.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.37.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.37.attention.norm_q.weight": [
+      128
+    ],
+    "layers.37.attention.norm_k.weight": [
+      128
+    ],
+    "layers.37.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.37.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.37.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.37.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.37.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.37.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.37.attention_norm1.weight": [
+      3584
+    ],
+    "layers.37.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.37.attention_norm2.weight": [
+      3584
+    ],
+    "layers.37.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.37.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.37.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.38.cross_attn_gate": [
+      1
+    ],
+    "layers.38.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.38.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.38.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.38.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.38.attention.norm_q.weight": [
+      128
+    ],
+    "layers.38.attention.norm_k.weight": [
+      128
+    ],
+    "layers.38.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.38.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.38.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.38.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.38.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.38.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.38.attention_norm1.weight": [
+      3584
+    ],
+    "layers.38.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.38.attention_norm2.weight": [
+      3584
+    ],
+    "layers.38.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.38.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.38.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.38.cross_attn_norm1.weight": [
+      3584
+    ],
+    "layers.38.cross_attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.cross_attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.38.cross_attention.to_k.packed_weight_indices": [
+      13762560
+    ],
+    "layers.38.cross_attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.38.cross_attention.to_v.packed_weight_indices": [
+      13762560
+    ],
+    "layers.38.cross_attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.38.cross_attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.38.cross_attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.38.cross_attention.norm_q.weight": [
+      128
+    ],
+    "layers.38.cross_attention.norm_k.weight": [
+      128
+    ],
+    "layers.38.cross_attn_norm2.weight": [
+      3584
+    ],
+    "layers.39.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.39.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.39.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.39.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.39.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.39.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.39.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.39.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.39.attention.norm_q.weight": [
+      128
+    ],
+    "layers.39.attention.norm_k.weight": [
+      128
+    ],
+    "layers.39.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.39.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.39.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.39.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.39.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.39.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.39.attention_norm1.weight": [
+      3584
+    ],
+    "layers.39.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.39.attention_norm2.weight": [
+      3584
+    ],
+    "layers.39.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.39.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.39.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.40.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.40.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.40.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.40.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.40.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.40.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.40.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.40.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.40.attention.norm_q.weight": [
+      128
+    ],
+    "layers.40.attention.norm_k.weight": [
+      128
+    ],
+    "layers.40.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.40.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.40.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.40.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.40.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.40.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.40.attention_norm1.weight": [
+      3584
+    ],
+    "layers.40.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.40.attention_norm2.weight": [
+      3584
+    ],
+    "layers.40.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.40.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.40.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.41.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.41.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.41.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.41.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.41.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.41.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.41.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.41.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.41.attention.norm_q.weight": [
+      128
+    ],
+    "layers.41.attention.norm_k.weight": [
+      128
+    ],
+    "layers.41.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.41.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.41.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.41.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.41.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.41.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.41.attention_norm1.weight": [
+      3584
+    ],
+    "layers.41.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.41.attention_norm2.weight": [
+      3584
+    ],
+    "layers.41.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.41.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.41.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.42.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.42.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.42.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.42.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.42.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.42.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.42.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.42.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.42.attention.norm_q.weight": [
+      128
+    ],
+    "layers.42.attention.norm_k.weight": [
+      128
+    ],
+    "layers.42.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.42.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.42.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.42.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.42.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.42.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.42.attention_norm1.weight": [
+      3584
+    ],
+    "layers.42.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.42.attention_norm2.weight": [
+      3584
+    ],
+    "layers.42.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.42.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.42.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "layers.43.attention.to_q.packed_weight_indices": [
+      6422528
+    ],
+    "layers.43.attention.to_q.row_norms": [
+      3584
+    ],
+    "layers.43.attention.to_k.packed_weight_indices": [
+      6422528
+    ],
+    "layers.43.attention.to_k.row_norms": [
+      3584
+    ],
+    "layers.43.attention.to_v.packed_weight_indices": [
+      6422528
+    ],
+    "layers.43.attention.to_v.row_norms": [
+      3584
+    ],
+    "layers.43.attention.to_out.0.packed_weight_indices": [
+      6422528
+    ],
+    "layers.43.attention.to_out.0.row_norms": [
+      3584
+    ],
+    "layers.43.attention.norm_q.weight": [
+      128
+    ],
+    "layers.43.attention.norm_k.weight": [
+      128
+    ],
+    "layers.43.feed_forward.w1.packed_weight_indices": [
+      17088512
+    ],
+    "layers.43.feed_forward.w1.row_norms": [
+      9536
+    ],
+    "layers.43.feed_forward.w2.packed_weight_indices": [
+      17088512
+    ],
+    "layers.43.feed_forward.w2.row_norms": [
+      3584
+    ],
+    "layers.43.feed_forward.w3.packed_weight_indices": [
+      17088512
+    ],
+    "layers.43.feed_forward.w3.row_norms": [
+      9536
+    ],
+    "layers.43.attention_norm1.weight": [
+      3584
+    ],
+    "layers.43.ffn_norm1.weight": [
+      3584
+    ],
+    "layers.43.attention_norm2.weight": [
+      3584
+    ],
+    "layers.43.ffn_norm2.weight": [
+      3584
+    ],
+    "layers.43.adaLN_modulation.0.weight": [
+      14336,
+      256
+    ],
+    "layers.43.adaLN_modulation.0.bias": [
+      14336
+    ],
+    "final_layer.linear.weight": [
+      128,
+      3584
+    ],
+    "final_layer.linear.bias": [
+      128
+    ],
+    "final_layer.adaLN_modulation.1.weight": [
+      3584,
+      256
+    ],
+    "final_layer.adaLN_modulation.1.bias": [
+      3584
+    ]
+  },
+  "checksums": {
+    "model.safetensors": "b86ebe525ee2e2d452d2025bac8b60fe781bd6e418d0cd00d46413f0cb1b4842",
+    "orbitquant_codebooks.safetensors": "354b4de89d7dbecbca33cba491af43aa0e94a9f469658c384d3743dbdb99c288",
+    "orbitquant_rotations.safetensors": "52a2273038ff076d0b9fef76844210857779682c594905dee0e769d14a1424ed",
+    "quantization_config.json": "ed1bde093e2ed3403e7cb5fa154750216bc78519ce869cc7b00a77b8eb126369",
+    "model_index.json": "b84bc37b1253ebab41939668501fd4892d8ad3c71f4e405445b255cf9bde08bf",
+    "prompts.json": "04901c871e459b144335027156c61ccd8ff0ecf48851c3b4811c483ee123b658",
+    "benchmark/summary.json": "243102848b8578d0c532ac82b24976c41b1e090ed2e7b221bf675c9e5d56281d"
+  }
+}

transformer/orbitquant_rotations.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52a2273038ff076d0b9fef76844210857779682c594905dee0e769d14a1424ed
+size 354796

transformer/quantization_config.json

@@ -0,0 +1,45 @@
+{
+  "quant_method": "orbitquant",
+  "weight_bits": 4,
+  "activation_bits": 4,
+  "rotation": "rpbh",
+  "rotation_seed": 0,
+  "block_size": "paper",
+  "codebook": "lloyd_max",
+  "codebook_version": 2,
+  "codebook_dtype": "float32",
+  "row_norm_dtype": "bfloat16",
+  "activation_norm_dtype": "float32",
+  "activation_eps": 1e-10,
+  "weight_pack_dtype": "uint8",
+  "target_policy": "universal",
+  "adaln_policy": "int4_rtn",
+  "adaln_group_size": 64,
+  "modules_to_convert": [
+    "noise_refiner.*.attention.to_*",
+    "noise_refiner.*.feed_forward.w*",
+    "layers.*.attention.to_*",
+    "layers.*.feed_forward.w*",
+    "layers.*.cross_attention.to_*",
+    "perceiver.layers.*.cross_attention.to_*",
+    "perceiver.layers.*.self_attention.to_*",
+    "perceiver.layers.*.ffn.w*"
+  ],
+  "modules_to_use_adaln": [],
+  "modules_to_not_convert": [],
+  "modules_dtype_dict": {},
+  "artifact_format_version": 1,
+  "runtime_mode": "auto_fused",
+  "activation_kernel_backend": "auto",
+  "packed_matmul_block_m": 64,
+  "packed_matmul_block_n": 64,
+  "packed_matmul_block_k": 128,
+  "packed_matmul_num_warps": 4,
+  "weight_row_tile_size": 256,
+  "w4a4_int8_weight_cache": false,
+  "lowbit_boundary_protection": "auto",
+  "lowbit_protected_blocks": 4,
+  "lowbit_protected_bits": 4,
+  "lowbit_interior_protection": "auto",
+  "fused_layout": null
+}

vae/config.json

@@ -0,0 +1,40 @@
+{
+  "_class_name": "AutoencoderKLFlux2",
+  "_diffusers_version": "0.41.0.dev0",
+  "_name_or_path": "/tmp/fibo2_hf_release/hf/hub/models--briaai--fibo-2/snapshots/9e70dcfc6f8acf39cba59484947cf643293b5233",
+  "act_fn": "silu",
+  "batch_norm_eps": 0.0001,
+  "batch_norm_momentum": 0.1,
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "decoder_block_out_channels": null,
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "in_channels": 3,
+  "latent_channels": 32,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "patch_size": [
+    2,
+    2
+  ],
+  "sample_size": 1024,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": true,
+  "use_quant_conv": true
+}

vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d64f3a68e1cc4f9f4e29b6e0da38a0204fe9a49f2d4053f0ec1fa1ca02f9c4b5
+size 336213556