Simplify README and clarify DMD checkpoint sources

README.md

@@ -16,170 +16,89 @@ base_model:
 - Wan-AI/Wan2.1-T2V-1.3B-Diffusers
 ---
 
-# ReSeed — single-seed model weights
+# ReSeed
 
 **A First-Step Bias Restores Lost Diversity in Diffusion Model Distillation**
 
 [Paper](https://arxiv.org/abs/2610.37110) · [Project page](https://twwinde.github.io/reseed-project-page/) · [Code](https://github.com/TWWinde/ReSeed)
 
-This branch contains four ReSeed bias adapters and the two self-trained DMD
-students needed to use the SD3 adapters. It is a **private, single-seed
-release**: every selected training run uses seed **42**. Use the
-**`main` revision** when downloading. The original `single-seed` branch is
-retained as a release snapshot.
-
-The SD3-Medium and SD3.5-Medium students were trained by the ReSeed authors.
-The FLUX and Wan students come from public releases by their respective
-authors. Public teacher models and shared components are referenced below,
-not copied into this repository.
-
-## Contents
-
-- [What ReSeed changes](#what-reseed-changes)
-- [Released checkpoints and model sources](#released-checkpoints-and-model-sources)
-- [Repository layout](#repository-layout)
-- [Environment and authentication](#environment-and-authentication)
-- [Download the release](#download-the-release)
-- [SD3-Medium and SD3.5-Medium inference](#sd3-medium-and-sd35-medium-inference)
-- [SenseFlow-FLUX inference](#senseflow-flux-inference)
-- [Wan2.1 text-to-video inference](#wan21-text-to-video-inference)
-- [Reproducibility and file verification](#reproducibility-and-file-verification)
-- [Troubleshooting](#troubleshooting)
-- [Limitations and licensing](#limitations-and-licensing)
-- [Citation](#citation)
-
-## What ReSeed changes
-
-ReSeed trains small, per-channel residual biases inside a frozen distilled
-student. At inference, the biases are active **only during the first of four
-denoising steps**. The original student weights remain unchanged, and the
-final three steps run without the biases. No teacher forward pass is needed
-when generating with the trained adapter.
-
-During adapter training, the student's first-step velocity is matched to
-the teacher's average velocity over the same time interval. This is a
-post-distillation correction; it does not repeat student distillation.
-
-The adapter is **not** a LoRA and is **not** a complete Diffusers pipeline.
-It must be loaded using the ReSeed adapter code and applied to the exact
-student it was fitted on. Do not load it with `load_lora_weights`, apply it
-to the teacher, or leave it active for all four steps.
-
-## Released checkpoints and model sources
-
-### Weights included in this branch
-
-| Model | Adapter training seed | Adapter update | Bias parameters | Matching student |
-| --- | ---: | ---: | ---: | --- |
-| SD3-Medium | 42 | **1900** | 72,192 | Our DMD, training seed 42, update 2000; included |
-| SD3.5-Medium | 42 | **2100** | 72,192 | Our DMD, training seed 42, update 2000; included |
-| SenseFlow-FLUX | 42 | **2000** | 233,472 | Public SenseFlow-FLUX; download separately |
-| Wan2.1-T2V-1.3B | 42 | **300** | 46,080 | Public Wan DMD2, iteration 1000; download separately |
-
-Adapter updates and student updates are different counters. For example,
-the SD3.5 adapter at update 2100 must be used with the included student at
-update 2000, not with an unrelated student checkpoint named `2100`.
-
-Three different seeds may appear in a workflow:
-
-1. **Student training seed:** selects our DMD student run; it is 42 here.
-2. **Adapter training seed:** selects the ReSeed fitting run; it is 42 here.
-3. **Generation seed:** selects the initial noise for an image or video.
-   This can be any supported seed, such as 0 or 1; it need not be 42.
-
-These are individual checkpoints, not averaged weights. One selected run
-does not by itself reproduce a three-training-seed aggregate in the paper.
-GAD and DP-DMD student weights are not included in this release.
-
-### Public teachers and external students
-
-| Component | Original source | Revision or checkpoint |
-| --- | --- | --- |
-| SD3-Medium teacher and shared components | [stabilityai/stable-diffusion-3-medium-diffusers](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) | `ea42f8cef0f178587cf766dc8129abd379c90671` |
-| SD3.5-Medium teacher and shared components | [stabilityai/stable-diffusion-3.5-medium](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) | `b940f670f0eda2d07fbb75229e779da1ad11eb80` |
-| FLUX teacher and shared components | [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) | `3de623fc3c33e44ffbe2bad470d0f45bccf2eb21` |
-| SenseFlow-FLUX student transformer | [domiso/SenseFlow](https://huggingface.co/domiso/SenseFlow) | Revision `1bbd81546d590acf24470b39ab28f9cd918c3927`, subfolder **`SenseFlow-FLUX`** |
-| Wan teacher and shared components | [Wan-AI/Wan2.1-T2V-1.3B-Diffusers](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers) | Original public teacher; see the video runtime setup |
-| Wan DMD2 student | [csusupergear/wan_t2v_checkpoints_release](https://huggingface.co/csusupergear/wan_t2v_checkpoints_release) | Prefix **`dmd2_t2v_vipe/0001000`**; retain its `.pth` and `.net_model.zip` files |
-
-The SenseFlow transformer is directly in `SenseFlow-FLUX`, not
-`SenseFlow-FLUX/transformer`. The Wan checkpoint uses the DFD/fastgen format;
-it is not a standalone Diffusers `transformer/` folder.
-
-The original SD3-M manifest retains the historical internal objective label
-`tdm`; its recorded implementation is `sd_distill_classic_v1`. Original
-manifests are preserved without silently changing their labels or paths.
-
-## Repository layout
-
-```text
-Wenwu2000/ReSeed (main)/
-├── README.md
-├── requirements.txt
-├── inference.py                  # Single-prompt SD3-M / SD3.5-M example
-├── release_manifest.json         # Pairings, sources, sizes and SHA-256 hashes
-├── adapters/
-│   ├── sd3m/seed_42/checkpoint-1900/
-│   ├── sd35m/seed_42/checkpoint-2100/
-│   ├── flux/seed_42/checkpoint-2000/
-│   └── wan21/seed_42/checkpoint-300/
-│       # Each adapter directory contains residual_biases.pt and manifest.json
-└── students/
-    ├── sd3m/seed_42/checkpoint-2000/
-    └── sd35m/seed_42/checkpoint-2000/
-        ├── manifest.json
-        └── student/
-            ├── config.json
-            ├── diffusion_pytorch_model-00001-of-00002.safetensors
-            ├── diffusion_pytorch_model-00002-of-00002.safetensors
-            └── diffusion_pytorch_model.safetensors.index.json
-```
+ReSeed improves diversity in a frozen four-step student by fitting small
+residual biases to the teacher's first-interval velocity. The biases are
+enabled only in the first denoising step; the original weights and final
+three steps remain unchanged.
 
-The complete release is approximately **18.22 GB** in decimal units. Most
-of that size is the two self-trained students. Their original inference
-files are retained without conversion or quantization. The adapters are small.
+This repository provides four ReSeed adapters and two self-trained SD3 DMD
+students. The selected adapter runs and our two student runs use training
+seed 42. Generation seeds can be chosen independently.
 
-This repository excludes optimizer state, fake-score models, teacher
-weights, text encoders, VAEs, datasets, prompt caches and generated media.
-It is an inference release, not a resumable training checkpoint archive.
-Absolute paths in original manifests are provenance records from our
-training system; users do not need to recreate those server paths.
+## Student models and their sources
 
-## Environment and authentication
+### SD3-Medium and SD3.5-Medium: trained by us
 
-Use Python 3.10 or later and install PyTorch for your CUDA environment.
-The image inference examples target a CUDA GPU with BF16 support. FLUX
-also needs enough memory for its transformer, text encoders and VAE; the
-example generator does not automatically quantize or offload them.
+We trained these four-step students in our CRRA `unified_distill` codebase
+from the corresponding Stability AI models. Both released students are
+**seed 42, checkpoint-2000**. Their full transformer weights are included:
 
-The image-side reference environment uses:
+| Student | Released weights | Teacher | Teacher sampling reference |
+| --- | --- | --- | --- |
+| SD3-Medium DMD | [checkpoint-2000](https://huggingface.co/Wenwu2000/ReSeed/tree/main/students/sd3m/seed_42/checkpoint-2000) | [SD3-Medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) | 28 steps, CFG 7 |
+| SD3.5-Medium DMD | [checkpoint-2000](https://huggingface.co/Wenwu2000/ReSeed/tree/main/students/sd35m/seed_42/checkpoint-2000) | [SD3.5-Medium](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) | 40 steps, CFG 4.5 |
 
-| Package | Version |
-| --- | --- |
-| PyTorch | 2.6.0 in the original generation environment; install for your CUDA setup |
-| Diffusers | 0.34.0 |
-| Transformers | 4.45.2 |
-| Hugging Face Hub | 0.36.2 |
+These are our experimental students, **not checkpoints from the official
+DMD2 release**. They use four-step FlowMatch Euler inference at CFG 1.
+For the underlying methods, see [DMD](https://arxiv.org/abs/2311.18828) and
+[DMD2](https://github.com/tianweiy/DMD2).
 
-Install the Hub CLI before downloading:
+The original training manifests are included beside each student. The
+SD3-Medium manifest retains the historical objective label `tdm` and records
+the implementation `sd_distill_classic_v1`; the SD3.5-Medium manifest records
+`dmd`. These records have not been relabelled. The release's exact student
+and adapter pairings are listed in [release_manifest.json](release_manifest.json).
 
-```bash
-pip install huggingface_hub==0.36.2
-hf auth login
-hf auth whoami
-```
+### Wan2.1 DMD2: public checkpoint from the DFD authors
+
+The video student is the Wan2.1-T2V-1.3B DMD2 checkpoint released by the
+[Data-Forcing Distillation authors](https://github.com/csy2077/data-forcing-distillation).
+We use **`dmd2_t2v_vipe/0001000`**, trained on the ViPE `long` set:
+
+- [Checkpoint files](https://huggingface.co/csusupergear/wan_t2v_checkpoints_release/tree/main/dmd2_t2v_vipe)
+- [Teacher: Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers)
+- [DFD / fastgen code and setup](https://github.com/csy2077/data-forcing-distillation)
+
+We did not train this DMD2 student. We trained the ReSeed adapter on top of
+it and provide only that adapter here. The separately released DFD
+post-trained model (`ours_wan_t2v_vipe/0000040`) is a different baseline,
+not the student used by this adapter.
+
+### FLUX: public SenseFlow student
+
+We use the four-step **SenseFlow-FLUX** transformer from
+[domiso/SenseFlow](https://huggingface.co/domiso/SenseFlow/tree/main/SenseFlow-FLUX),
+with shared components from [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev).
+This student was released by the SenseFlow authors, not trained by us.
+Download it separately; this repository contains its ReSeed adapter only.
+
+## Released adapters
+
+| Student | Adapter directory | Adapter updates | Bias parameters |
+| --- | --- | ---: | ---: |
+| SD3-Medium DMD | `adapters/sd3m/seed_42/checkpoint-1900` | 1900 | 72,192 |
+| SD3.5-Medium DMD | `adapters/sd35m/seed_42/checkpoint-2100` | 2100 | 72,192 |
+| SenseFlow-FLUX | `adapters/flux/seed_42/checkpoint-2000` | 2000 | 233,472 |
+| Wan2.1 DMD2 | `adapters/wan21/seed_42/checkpoint-300` | 300 | 46,080 |
 
-Your account needs access to this private repository. Separately accept
-the terms of any gated upstream model before downloading its components.
-Access to ReSeed does not automatically grant access to Stability AI or
-FLUX checkpoints. Never write an access token into source code or a README.
+Each directory contains `residual_biases.pt` and `manifest.json`. Adapter
+updates count ReSeed fitting, not student distillation: both included SD3
+students remain at update 2000. Use each adapter with its listed student.
+These are residual-bias adapters, not LoRA weights.
 
-## Download the release
+## Download and install
 
-Run the following from a working directory in which you want to keep the
-weights, upstream models and code:
+Use Python 3.10+ and a CUDA/BF16-capable PyTorch installation. Our image
+runtime used PyTorch 2.6.0, Diffusers 0.34.0 and Transformers 4.45.2.
 
 ```bash
+pip install huggingface_hub==0.36.2
 hf download Wenwu2000/ReSeed --revision main --local-dir reseed-weights
 
 git clone https://github.com/TWWinde/ReSeed.git
@@ -188,10 +107,8 @@ pip install -r reseed-weights/requirements.txt
 pip install -e ReSeed/image --no-deps
 ```
 
-The code revision above is the public ReSeed version used by these examples.
-The weight repository itself does not install the `reseed` Python package.
-
-For SD3-Medium only, avoid downloading the other student:
+The full download is about 18.22 GB, mostly the two student transformers.
+To download only SD3-Medium, use:
 
 ```bash
 hf download Wenwu2000/ReSeed --revision main --local-dir reseed-weights \
@@ -199,26 +116,13 @@ hf download Wenwu2000/ReSeed --revision main --local-dir reseed-weights \
   "adapters/sd3m/*" "students/sd3m/*"
 ```
 
-Replace both `sd3m` patterns with `sd35m` for SD3.5-Medium. For FLUX or
-Wan, download `adapters/flux/*` or `adapters/wan21/*` and the documentation;
-their public students must still be obtained separately.
-
-Python can download the same branch without cloning a large Git repository:
-
-```python
-from huggingface_hub import snapshot_download
-
-weights = snapshot_download(
-    repo_id="Wenwu2000/ReSeed",
-    revision="main",
-    local_dir="reseed-weights",
-)
-print(weights)
-```
+Use `sd35m` instead of `sd3m` for SD3.5-Medium. For FLUX or Wan, only the
+corresponding adapter is needed from this repository. Upstream gated models
+require accepting their terms and logging in with `hf auth login`.
 
-## SD3-Medium and SD3.5-Medium inference
+## SD3 image generation
 
-### Download the matching public components
+Download the public components for the model you want to run:
 
 ```bash
 hf download stabilityai/stable-diffusion-3-medium-diffusers \
@@ -228,93 +132,38 @@ hf download stabilityai/stable-diffusion-3.5-medium \
   --revision b940f670f0eda2d07fbb75229e779da1ad11eb80 --local-dir models/sd35m
 ```
 
-The example loads the included DMD transformer in place of the public
-teacher transformer. It reuses the public text encoders, tokenizers, VAE
-and scheduler. It does not run the teacher during student inference.
-
-### Generate with ReSeed
+Generate an image with the matching DMD student and ReSeed:
 
 ```bash
 python reseed-weights/inference.py --weights reseed-weights --model sd3m \
   --base models/sd3m --prompt "a photo of a cat" --seed 0 \
-  --output samples/sd3m_reseed_seed0.png
+  --output samples/sd3m_reseed.png
 
 python reseed-weights/inference.py --weights reseed-weights --model sd35m \
   --base models/sd35m --prompt "a photo of a cat" --seed 0 \
-  --output samples/sd35m_reseed_seed0.png
+  --output samples/sd35m_reseed.png
 ```
 
-`--base` is a **local upstream snapshot directory**. `--weights` is the
-downloaded ReSeed branch root. `--model sd3m` or `sd35m` selects the correct
-student and adapter from `release_manifest.json` automatically.
+Add `--without-reseed` and choose a different output filename to generate
+the plain DMD baseline with the same prompt and initial noise.
 
-### Generate the matched DMD baseline
+The script loads our student transformer and the public model's text
+encoders, tokenizers, VAE and scheduler. It uses 1024×1024 BF16 inference,
+four Euler steps, CFG 1 and static shift 3. The pipeline receives unshifted
+sigmas `[1, .75, .5, .25]`; do not substitute the shifted boundaries from
+the training manifest. Initial noise is drawn in FP32 on CPU, then cast to
+BF16. The adapter is disabled after the first step.
 
-Use the same prompt and generation seed, and add `--without-reseed`:
-
-```bash
-python reseed-weights/inference.py --weights reseed-weights --model sd3m \
-  --base models/sd3m --prompt "a photo of a cat" --seed 0 \
-  --without-reseed --output samples/sd3m_dmd_seed0.png
-
-python reseed-weights/inference.py --weights reseed-weights --model sd35m \
-  --base models/sd35m --prompt "a photo of a cat" --seed 0 \
-  --without-reseed --output samples/sd35m_dmd_seed0.png
-```
-
-The baseline keeps the same DMD weights and initial noise; only the adapter
-is omitted. It is not a comparison against the public many-step teacher.
-The script refuses to overwrite an existing image, so use separate output names.
-
-### Fixed sampler and adapter behavior
-
-| Setting | Both SD3 student examples |
-| --- | --- |
-| Resolution | 1024×1024 |
-| Solver | Native FlowMatch Euler |
-| Number of steps | 4 |
-| Student CFG | **1** |
-| Scheduler | Static shift 3, dynamic shifting disabled |
-| Explicit unshifted sigmas | `[1.0, 0.75, 0.5, 0.25]` |
-| Text maximum sequence length | 256 |
-| Inference dtype | BF16 |
-| Initial noise | CPU FP32 Gaussian draw, then BF16 cast |
-| ReSeed gate | Active for step 1 only; disabled for steps 2–4 |
-
-The `sigmas` argument above is the **unshifted** schedule supplied to the
-pipeline; the scheduler applies its shift. Do not replace it with the
-already shifted values from an adapter manifest. Those manifests may also
-contain BF16-rounded training boundaries.
-
-The teacher references use **28 steps / CFG 7** for SD3-Medium and **40
-steps / CFG 4.5** for SD3.5-Medium. Those guidance settings must not replace
-the student's CFG 1. The adapter manifests record the exact first interval
-and teacher microstep settings used for fitting.
-
-For multiple samples, change `--seed` and the output filename. The example
-handles one prompt/sample per invocation and resets its gate for each run.
-The commands are process-level examples, not single-GPU Slurm submissions.
-Our MUSICA jobs request a whole four-H100 node and parallelize independent
-samples across its four GPUs.
-
-## SenseFlow-FLUX inference
-
-### Obtain FLUX components and the released student
+## FLUX image generation
 
 ```bash
 hf download black-forest-labs/FLUX.1-dev \
   --revision 3de623fc3c33e44ffbe2bad470d0f45bccf2eb21 --local-dir models/flux-dev
-
 hf download domiso/SenseFlow --revision 1bbd81546d590acf24470b39ab28f9cd918c3927 \
   --include "SenseFlow-FLUX/*" --local-dir models/senseflow
 ```
 
-Use the SenseFlow transformer, not the FLUX.1-dev transformer, with this
-adapter. The public FLUX snapshot provides shared components and the scheduler.
-
-### Prepare a prompt CSV
-
-Save the following as `prompts.csv` in the working directory:
+Create `prompts.csv`:
 
 ```csv
 case_number,evaluation_seed,prompt,prompt_id,sample_id
@@ -322,11 +171,7 @@ case_number,evaluation_seed,prompt,prompt_id,sample_id
 1,1,a photo of a cat,0,1
 ```
 
-`case_number` must be unique. `evaluation_seed` is the generation noise
-seed. Rows for the same prompt share `prompt_id`; `sample_id` identifies
-the sample within that prompt. Quote the prompt field if it contains commas.
-
-### Generate SenseFlow + ReSeed
+Then run:
 
 ```bash
 cd ReSeed/image
@@ -335,165 +180,57 @@ python tools/flux_generate.py --model senseflow \
   --senseflow-transformer ../../models/senseflow/SenseFlow-FLUX \
   --first-interval-residual-bias-checkpoint ../../reseed-weights/adapters/flux/seed_42/checkpoint-2000 \
   --prompts ../../prompts.csv --expected-count 2 --samples-per-prompt 2 \
-  --evaluation-protocol reseed-single-seed-example \
+  --evaluation-protocol reseed-example \
   --steps 4 --resolution 1024 --max-sequence-length 512 \
   --output ../../samples/flux_reseed
 cd ../..
 ```
 
-Omit `--first-interval-residual-bias-checkpoint` and use a different output
-directory for the plain SenseFlow baseline. Retain the same CSV and seeds.
-The generator writes images, a generation manifest and a protocol record.
-`reseed-single-seed-example` is an example run label, not a paper benchmark.
+Omit the adapter argument for the plain SenseFlow baseline. This student
+uses its native four-step dynamically shifted Euler schedule with the
+guidance embedding disabled. Do not pass the teacher's guidance 3.5 to the
+student. The transformer is in `SenseFlow-FLUX`, not a nested `transformer/` folder.
 
-The native student uses four dynamically shifted Euler steps, approximately
-`[1, .904531, .759511, .512844, 0]`, with BF16 inference and text length 512.
-Its guidance embedding is **disabled**: do not pass `--guidance 3.5` for
-SenseFlow. FLUX.1-dev guidance 3.5 belongs to the teacher target, not this
-student. The generator checks the schedule against Diffusers and activates
-the bias only during the first step.
+## Wan video generation
 
-## Wan2.1 text-to-video inference
-
-### Set up DFD / fastgen
-
-The Wan adapter uses the DMD2 student in the
-[DFD checkpoint release](https://huggingface.co/csusupergear/wan_t2v_checkpoints_release)
-and runs through [DFD / fastgen](https://github.com/csy2077/data-forcing-distillation).
-Follow the [ReSeed video installation guide](https://github.com/TWWinde/ReSeed/blob/eed9a03bd85a0af8a7dc36e8da5b48d2a31f58ca/video/README.md)
-to install the ReSeed module and apply its inference patch to a configured
-DFD checkout. The image package alone is not sufficient for video generation.
-Use DFD's environment and public Wan component setup; do not assume that
-the image requirements file provides the complete video environment.
-
-### Download the public DMD2 checkpoint
+First follow the [ReSeed video setup](https://github.com/TWWinde/ReSeed/blob/eed9a03bd85a0af8a7dc36e8da5b48d2a31f58ca/video/README.md)
+to install the video code and apply the ReSeed inference patch to DFD.
+The video runtime has separate dependencies from the image package.
 
 ```bash
 hf download csusupergear/wan_t2v_checkpoints_release \
   --include "dmd2_t2v_vipe/0001000*" --local-dir wan-checkpoints
 ```
 
-Keep the original `.pth` and `.net_model.zip` filenames. DFD receives the
-checkpoint **prefix** `dmd2_t2v_vipe/0001000`, rather than a renamed weight
-file or a Diffusers pipeline directory.
-
-### Generate matched DMD2 and DMD2 + ReSeed videos
-
-Create a UTF-8 text file with one prompt per line, then run the following
-from the configured, patched DFD checkout:
+Keep the `.pth` and `.net_model.zip` files under their original names.
+From the configured DFD checkout, use a text file with one prompt per line:
 
 ```bash
 export DFD_CKPT_ROOT=/absolute/path/to/wan-checkpoints
 export DFD_RESEED_ADAPTER=/absolute/path/to/reseed-weights/adapters/wan21/seed_42/checkpoint-300
 export DFD_PROMPT_FILE=/absolute/path/to/video-prompts.txt
-
 python evaluation/generate_multiseed.py --config evaluation/config.yaml \
   --models dmd2 dmd2_reseed --seeds 0 1 \
   --output-root evaluation/reseed-example
 ```
 
-The driver passes `FASTGEN_RESEED_ADAPTER` to the patched inference process
-for the ReSeed arm. Setting `DFD_RESEED_ADAPTER` without installing that
-patch does not add ReSeed support to an unmodified inference script.
-
-| Setting | Wan DMD2 / DMD2 + ReSeed |
-| --- | --- |
-| Student solver | Native four-step ODE sampler |
-| Student time boundaries | `[.999, .937, .833, .624, 0]` |
-| Student CFG | None; no student CFG branch is evaluated |
-| Video size | 832×480 |
-| Frames / FPS | 81 / 16 |
-| Inference dtype | BF16 |
-| ReSeed gate | First step only |
-| Teacher reference | 50 steps, CFG 5, shift 5 |
-
-Do not substitute the public DFD post-trained checkpoint for DMD2: the
-included ReSeed adapter was fitted on **DMD2 iteration 1000**. The DFD
-post-trained model is a different comparison method, not this adapter's base.
-
-## Reproducibility and file verification
-
-`release_manifest.json` records each model's teacher source, student origin,
-selected adapter update, matching student path, training seed provenance,
-original checkpoint location and per-file SHA-256 hashes. Original
-checkpoint manifests remain beside their weights.
-The release manifest retains `revision: single-seed`, identifying the
-original release branch. The same checkpoint files are also available on
-`main`; their hashes and model pairings are unchanged.
-
-After a **full** download, verify all inference files listed in the manifest:
-
-```python
-import hashlib
-import json
-from pathlib import Path
-
-root = Path("reseed-weights")
-manifest = json.loads((root / "release_manifest.json").read_text())
-for entry in manifest["files"]:
-    path = root / entry["path"]
-    assert path.is_file(), f"Missing: {path}"
-    assert path.stat().st_size == entry["bytes"], f"Wrong size: {path}"
-    digest = hashlib.sha256()
-    with path.open("rb") as handle:
-        for block in iter(lambda: handle.read(8 * 1024 * 1024), b""):
-            digest.update(block)
-    assert digest.hexdigest() == entry["sha256"], f"Hash mismatch: {path}"
-print("All listed inference files verified.")
-```
+This generates matched DMD2 and DMD2 + ReSeed videos. The student uses four
+ODE steps with boundaries `[.999, .937, .833, .624, 0]`, no student CFG,
+81 frames at 832×480, 16 FPS and BF16. The teacher reference uses 50 UniPC
+steps, CFG 5 and shift 5. ReSeed is active only in the first student step.
+
+## Files and licenses
 
-For a selective download, verify only the manifest entries you deliberately
-downloaded; the full check above will correctly report the omitted models
-as missing. Documentation can be updated independently of the weight
-manifest. Hashes identify the released files, not generated outputs.
-
-For comparisons, keep the prompt, initial noise, scheduler, number of steps,
-resolution, dtype and guidance unchanged, and toggle only ReSeed. A matching
-integer seed is not sufficient if different runtimes draw noise differently.
-The supplied image example explicitly draws FP32 noise on CPU before the
-model-dtype cast. Do not normalize or interpolate that noise unless this is
-an explicitly separate experiment.
-
-Exact pixel equality across hardware, attention kernels or library versions
-is not guaranteed. BF16/FP32 changes can also change generated samples.
-Use the reference environment and pinned public snapshots when checking a
-reported comparison. The example scripts are inference entry points; they
-do not by themselves run COCO quality scoring, diversity evaluation or VBench.
-
-## Troubleshooting
-
-| Symptom | Check |
-| --- | --- |
-| `401`, `403` or a private repository appears missing | Log in with an authorized account; check private-repo access and upstream gated-model approval separately. |
-| Missing weights in an old local download | Refresh the snapshot with `--revision main`; check any selective-download patterns. |
-| `No module named reseed` | Install `ReSeed/image` in the same Python environment used for inference. |
-| A LoRA loader rejects `residual_biases.pt` | Use the ReSeed residual-bias loader, not PEFT or `load_lora_weights`. |
-| Missing shard or incomplete student | Download both safetensor shards, their index and `config.json`; do not copy a single shard alone. |
-| SD3 sampler contract error | Use the matching pinned base, static shift 3, the supplied unshifted sigmas and student CFG 1. |
-| Unexpected FLUX guidance error | Use `--model senseflow`, the `SenseFlow-FLUX` folder, and no guidance embedding. |
-| Wan adapter is ignored | Install the ReSeed inference patch and use the `dmd2_reseed` arm. |
-| Wan checkpoint loader fails | Preserve the public `.pth` / `.net_model.zip` pair and pass the correct checkpoint prefix. |
-| CUDA out of memory | Check which model is loaded and available GPU memory. The example does not silently change resolution, precision, quantization or offload settings. |
-| Output image already exists | Select a new output filename; the SD3 example intentionally refuses overwrites. |
-
-## Limitations and licensing
-
-These adapters are tied to their listed students and native sampling
-settings. Applying one to another backbone, another distillation run, a
-different step count or a different resolution is outside this release's
-documented setup. A diversity gain on an evaluation bank does not guarantee
-that every prompt, quality metric or generated sample improves.
-
-This branch provides one selected run per model, not training data, complete
-training state or all experimental variants. It does not include weights
-for GAD, DP-DMD, SDXL or MiniMax experiments.
-
-The ReSeed implementation is MIT-licensed in the code repository. That does
-**not** replace the licenses of model weights or dependencies. Consult each
-linked upstream model's terms and the code repository's
-[NOTICE](https://github.com/TWWinde/ReSeed/blob/eed9a03bd85a0af8a7dc36e8da5b48d2a31f58ca/NOTICE.md).
-Keeping this repository private does not grant additional redistribution or
-commercial-use rights. This documentation update does not make the repository public.
+The release contains inference weights, not optimizer state, fake-score
+models, public teacher weights, datasets or generated samples. Text encoders
+and VAEs come from the linked base models. File hashes, pinned source
+revisions and checkpoint pairings are recorded in `release_manifest.json`;
+its branch and visibility fields describe the original private release.
+The repository is now public, with current instructions on `main` and the
+original `single-seed` branch retained as an archive.
+
+The ReSeed code is MIT-licensed. Model weights retain their upstream
+licenses and access conditions; see [NOTICE](https://github.com/TWWinde/ReSeed/blob/main/NOTICE.md).
 
 ## Citation
 
@@ -508,6 +245,3 @@ commercial-use rights. This documentation update does not make the repository pu
   url={https://arxiv.org/abs/2610.37110}
 }
 ```
-
-Please also acknowledge the teacher models, external distilled students
-and DFD / fastgen software when using them in a new study.