initial commit
Browse files- .gitignore +2 -0
- LICENSE +21 -0
- README.md +196 -0
- config.json +88 -0
- examples/reconstruction.ipynb +203 -0
- factorization.svg +0 -0
- load_stellar.py +119 -0
- stellar-b16.safetensors +3 -0
- stellar-b24.safetensors +3 -0
- stellar-b8.safetensors +3 -0
- stellar-h16.safetensors +3 -0
- stellar-l16.safetensors +3 -0
.gitignore
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
__pycache__/
|
| 2 |
+
*.pyc
|
LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MIT License
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2026 Microsoft
|
| 4 |
+
|
| 5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
in the Software without restriction, including without limitation the rights
|
| 8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
furnished to do so, subject to the following conditions:
|
| 11 |
+
|
| 12 |
+
The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
copies or substantial portions of the Software.
|
| 14 |
+
|
| 15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
SOFTWARE.
|
README.md
ADDED
|
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: stellar
|
| 4 |
+
pipeline_tag: image-feature-extraction
|
| 5 |
+
datasets:
|
| 6 |
+
- imagenet-1k
|
| 7 |
+
tags:
|
| 8 |
+
- vision
|
| 9 |
+
- self-supervised-learning
|
| 10 |
+
- representation-learning
|
| 11 |
+
- sparse-tokens
|
| 12 |
+
- vision-transformer
|
| 13 |
+
- image-feature-extraction
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# STELLAR — Sparse Visual Representations via Spatial–Semantic Factorization
|
| 17 |
+
|
| 18 |
+
**STELLAR** learns a **unified sparse visual representation** that supports both
|
| 19 |
+
**reconstruction** and **semantics** using as few as **16 tokens**. By factorizing
|
| 20 |
+
*"what"* (semantics) from *"where"* (spatial layout), each image is encoded as the
|
| 21 |
+
low-rank product of a **localization** matrix and a **semantics** matrix.
|
| 22 |
+
|
| 23 |
+
<p align="center">
|
| 24 |
+
<img src="factorization.svg" alt="Spatial–semantic factorization" width="720">
|
| 25 |
+
</p>
|
| 26 |
+
|
| 27 |
+
- 📄 **Paper:** [arXiv:2602.01905](https://arxiv.org/abs/2602.01905) (ICML 2026)
|
| 28 |
+
- 💻 **Code:** [github.com/microsoft/STELLAR](https://github.com/microsoft/STELLAR)
|
| 29 |
+
|
| 30 |
+
These checkpoints contain the **full set of trained STELLAR modules** (encoder, sparse
|
| 31 |
+
tokens, projections, reconstruction decoder, and clustering heads), so a single file
|
| 32 |
+
supports feature extraction, image reconstruction, and continued pretraining. All
|
| 33 |
+
models are self-supervised on **ImageNet-1K** at 224×224.
|
| 34 |
+
|
| 35 |
+
## Highlights
|
| 36 |
+
|
| 37 |
+
- **Sparse & unified** — one small set of tokens serves both high-level semantics
|
| 38 |
+
and pixel-level reconstruction.
|
| 39 |
+
- **Factorized latents** — each token captures a concept (*what*) together with a
|
| 40 |
+
spatial map of *where* it appears.
|
| 41 |
+
- **Strong on both axes** — STELLAR-H reaches **2.60 FID** (reconstruction) and
|
| 42 |
+
**79.1%** ImageNet linear-probing accuracy with just **16 tokens**.
|
| 43 |
+
|
| 44 |
+
## Available models
|
| 45 |
+
|
| 46 |
+
| Model | Backbone | Tokens | Params | Type | File |
|
| 47 |
+
| :--- | :--- | :---: | :---: | :--- | :--- |
|
| 48 |
+
| `stellar-b16` | ViT-B/16 | 16 | 88M | main | [`stellar-b16.safetensors`](stellar-b16.safetensors) |
|
| 49 |
+
| `stellar-l16` | ViT-L/16 | 16 | 307M | main | [`stellar-l16.safetensors`](stellar-l16.safetensors) |
|
| 50 |
+
| `stellar-h16` | ViT-H/14 | 16 | 636M | main | [`stellar-h16.safetensors`](stellar-h16.safetensors) |
|
| 51 |
+
| `stellar-b8` | ViT-B/16 | 8 | 88M | ablation | [`stellar-b8.safetensors`](stellar-b8.safetensors) |
|
| 52 |
+
| `stellar-b24` | ViT-B/16 | 24 | 88M | ablation | [`stellar-b24.safetensors`](stellar-b24.safetensors) |
|
| 53 |
+
|
| 54 |
+
The main models (`b16`, `l16`, `h16`) are recommended for downstream use; the 8- and
|
| 55 |
+
24-token base models are ablations on the number of sparse tokens.
|
| 56 |
+
|
| 57 |
+
## Usage
|
| 58 |
+
|
| 59 |
+
Install the STELLAR code and the Hub helpers:
|
| 60 |
+
|
| 61 |
+
```bash
|
| 62 |
+
pip install huggingface_hub safetensors
|
| 63 |
+
git clone https://github.com/microsoft/STELLAR && cd STELLAR
|
| 64 |
+
pip install -r requirements.txt
|
| 65 |
+
```
|
| 66 |
+
|
| 67 |
+
### Quick start
|
| 68 |
+
|
| 69 |
+
From the STELLAR code directory, use the [`load_stellar.py`](load_stellar.py) helper
|
| 70 |
+
(it downloads the weights from the Hub for you):
|
| 71 |
+
|
| 72 |
+
```python
|
| 73 |
+
import torch
|
| 74 |
+
from load_stellar import load_stellar, list_models
|
| 75 |
+
|
| 76 |
+
print(list_models()) # ['stellar-b16', 'stellar-l16', ...]
|
| 77 |
+
model = load_stellar("stellar-b16") # purpose="encode" (default)
|
| 78 |
+
|
| 79 |
+
# RGB image in [0, 1], resized to 224×224 (ImageNet normalization is applied internally)
|
| 80 |
+
image = torch.rand(1, 3, 224, 224)
|
| 81 |
+
with torch.no_grad():
|
| 82 |
+
out = model.encode(image)
|
| 83 |
+
|
| 84 |
+
out["sparse"] # (1, K, D) sparse concept tokens ("what")
|
| 85 |
+
out["spatial"] # (1, P, K) per-token spatial maps ("where")
|
| 86 |
+
out["dense"] # (1, P, D) dense per-patch features
|
| 87 |
+
out["cls"] # (1, 1, D) global image token
|
| 88 |
+
```
|
| 89 |
+
|
| 90 |
+
### Reconstruction & continued pretraining
|
| 91 |
+
|
| 92 |
+
The same checkpoint can be loaded for other purposes via the `purpose` argument. Image
|
| 93 |
+
reconstruction and continued pretraining use the decoder, which predicts
|
| 94 |
+
[MaskGIT-VQGAN](https://huggingface.co/fun-research/TiTok) tokens — pass the tokenizer
|
| 95 |
+
path as `vq_model`:
|
| 96 |
+
|
| 97 |
+
```python
|
| 98 |
+
# 1. encode -> factorized features (sparse concept tokens + spatial maps)
|
| 99 |
+
model = load_stellar("stellar-b16", purpose="reconstruct", vq_model=VQGAN_PATH)
|
| 100 |
+
features = model.encode(image) # dict: sparse (B,K,D), spatial (B,P,K), ...
|
| 101 |
+
|
| 102 |
+
# 2. decode the factorized features -> VQGAN decoder -> pixels
|
| 103 |
+
out = model.reconstruct(features) # or model.reconstruct(features["sparse"], features["spatial"])
|
| 104 |
+
pixels = out["reconstruction"] # (B, 3, H, W) RGB in [0, 1]
|
| 105 |
+
# 224x224 for /16 models, 256x256 for the /14 H model
|
| 106 |
+
# out["tokens"] : (B, P) predicted VQGAN token ids
|
| 107 |
+
# out["logits"] : (B, P, 1024) raw codebook logits
|
| 108 |
+
|
| 109 |
+
# continued pretraining (all modules, gradients enabled)
|
| 110 |
+
model = load_stellar("stellar-b16", purpose="pretrain", vq_model=VQGAN_PATH)
|
| 111 |
+
losses = model({"image": image, "labels": labels, ...})["predictions"]
|
| 112 |
+
```
|
| 113 |
+
|
| 114 |
+
`reconstruct` is the **decoder half** of STELLAR: it takes the factorized features and
|
| 115 |
+
runs low-rank dense map → ViT decoder → VQGAN decoder to return RGB pixels. See
|
| 116 |
+
[`examples/reconstruction.ipynb`](examples/reconstruction.ipynb) for an end-to-end demo
|
| 117 |
+
that loads an image and displays the reconstruction.
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
### What the model returns
|
| 121 |
+
|
| 122 |
+
| Key | Shape | Description | Typical use |
|
| 123 |
+
| :--- | :--- | :--- | :--- |
|
| 124 |
+
| `sparse` | `(B, K, D)` | sparse concept tokens | classification, retrieval |
|
| 125 |
+
| `spatial` | `(B, P, K)` | spatial map of each token | segmentation, visualization |
|
| 126 |
+
| `dense` | `(B, P, D)` | dense per-patch features | segmentation |
|
| 127 |
+
| `lowrank` | `(B, P, D)` | reassembled dense map | reconstruction |
|
| 128 |
+
| `cls` | `(B, 1, D)` | global representation | classification |
|
| 129 |
+
|
| 130 |
+
`B` = batch, `K` = number of sparse tokens, `P` = number of patches (196 for /16 at
|
| 131 |
+
224², 256 for /14), `D` = embedding dim (768 / 1024 / 1280 for B / L / H).
|
| 132 |
+
|
| 133 |
+
### Loading the weights manually
|
| 134 |
+
|
| 135 |
+
```python
|
| 136 |
+
import json, torch
|
| 137 |
+
from huggingface_hub import hf_hub_download
|
| 138 |
+
from safetensors.torch import load_file
|
| 139 |
+
from src.models.stellar_model import STELLARModel
|
| 140 |
+
|
| 141 |
+
repo = "microsoft/STELLAR"
|
| 142 |
+
cfg = json.load(open(hf_hub_download(repo, "config.json")))["models"]["stellar-b16"]
|
| 143 |
+
state = load_file(hf_hub_download(repo, cfg["weights"]))
|
| 144 |
+
|
| 145 |
+
model = STELLARModel(
|
| 146 |
+
num_sparse_tokens=cfg["num_sparse_tokens"],
|
| 147 |
+
num_decoder_layers=cfg["num_decoder_layers"],
|
| 148 |
+
spatial_temp=cfg["spatial_temp"],
|
| 149 |
+
vit_pretrained=cfg["backbone"],
|
| 150 |
+
do_recon=False, do_clustering=False, vq_model=None,
|
| 151 |
+
)
|
| 152 |
+
model.load_state_dict(state, strict=False) # encoder-only build ignores decoder/heads
|
| 153 |
+
model.eval()
|
| 154 |
+
features = model.encode(torch.rand(1, 3, 224, 224))
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
> **Tip:** download with `huggingface_hub` (as above) rather than `git clone` so that
|
| 158 |
+
> downloads are registered on the Hub — `git clone` is not counted in download stats.
|
| 159 |
+
|
| 160 |
+
## Model details
|
| 161 |
+
|
| 162 |
+
- **Architecture:** ViT encoder (MAE-initialized) + learned sparse latent queries with
|
| 163 |
+
spatial–semantic factorization.
|
| 164 |
+
- **Pretraining data:** ImageNet-1K (self-supervised; labels not used).
|
| 165 |
+
- **Input:** RGB images in `[0, 1]`, resized to 224×224 (bicubic). ImageNet mean/std
|
| 166 |
+
normalization is applied **inside** the model — pass raw `[0, 1]` images.
|
| 167 |
+
- **Weights:** the complete set of trained STELLAR modules (encoder, sparse tokens,
|
| 168 |
+
projections, reconstruction decoder, and clustering heads), stored in `safetensors`.
|
| 169 |
+
Only the third-party MaskGIT-VQGAN tokenizer is excluded — it is downloaded separately
|
| 170 |
+
(from [TiTok](https://huggingface.co/fun-research/TiTok)) and passed via `vq_model`.
|
| 171 |
+
- **Framework:** PyTorch.
|
| 172 |
+
|
| 173 |
+
## Intended uses & limitations
|
| 174 |
+
|
| 175 |
+
- **Intended use:** extracting compact sparse/dense visual features for downstream
|
| 176 |
+
recognition, segmentation, retrieval, reconstruction, and analysis.
|
| 177 |
+
- **Limitations:** pretrained on ImageNet-1K at 224×224, so features reflect that
|
| 178 |
+
distribution; performance on very different domains (e.g. medical, satellite) may
|
| 179 |
+
require fine-tuning. The models are research artifacts and are not safety-tested for
|
| 180 |
+
production decision-making.
|
| 181 |
+
|
| 182 |
+
## Citation
|
| 183 |
+
|
| 184 |
+
```bibtex
|
| 185 |
+
@inproceedings{zhao2026stellar,
|
| 186 |
+
title = {Learning Sparse Visual Representations via Spatial-Semantic Factorization},
|
| 187 |
+
author = {Zhao, Theodore Zhengde and Kiblawi, Sid and Yang, Jianwei and Usuyama, Naoto and Tan, Reuben and Codella, Noel C and Naumann, Tristan and Poon, Hoifung and Wei, Mu},
|
| 188 |
+
booktitle = {International Conference on Machine Learning (ICML)},
|
| 189 |
+
year = {2026},
|
| 190 |
+
url = {https://arxiv.org/abs/2602.01905},
|
| 191 |
+
}
|
| 192 |
+
```
|
| 193 |
+
|
| 194 |
+
## License
|
| 195 |
+
|
| 196 |
+
Released under the [MIT License](LICENSE).
|
config.json
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "stellar",
|
| 3 |
+
"library_name": "stellar",
|
| 4 |
+
"code": "https://github.com/microsoft/STELLAR",
|
| 5 |
+
"default": "stellar-b16",
|
| 6 |
+
"models": {
|
| 7 |
+
"stellar-b16": {
|
| 8 |
+
"weights": "stellar-b16.safetensors",
|
| 9 |
+
"category": "main",
|
| 10 |
+
"backbone": "facebook/vit-mae-base",
|
| 11 |
+
"image_size": 224,
|
| 12 |
+
"patch_size": 16,
|
| 13 |
+
"embed_dim": 768,
|
| 14 |
+
"num_sparse_tokens": 16,
|
| 15 |
+
"num_decoder_layers": 6,
|
| 16 |
+
"spatial_temp": 0.06,
|
| 17 |
+
"do_recon": true,
|
| 18 |
+
"do_clustering": true,
|
| 19 |
+
"do_cls": true,
|
| 20 |
+
"num_clusters": 16384,
|
| 21 |
+
"recon_type": "vq"
|
| 22 |
+
},
|
| 23 |
+
"stellar-l16": {
|
| 24 |
+
"weights": "stellar-l16.safetensors",
|
| 25 |
+
"category": "main",
|
| 26 |
+
"backbone": "facebook/vit-mae-large",
|
| 27 |
+
"image_size": 224,
|
| 28 |
+
"patch_size": 16,
|
| 29 |
+
"embed_dim": 1024,
|
| 30 |
+
"num_sparse_tokens": 16,
|
| 31 |
+
"num_decoder_layers": 6,
|
| 32 |
+
"spatial_temp": 0.06,
|
| 33 |
+
"do_recon": true,
|
| 34 |
+
"do_clustering": true,
|
| 35 |
+
"do_cls": true,
|
| 36 |
+
"num_clusters": 16384,
|
| 37 |
+
"recon_type": "vq"
|
| 38 |
+
},
|
| 39 |
+
"stellar-h16": {
|
| 40 |
+
"weights": "stellar-h16.safetensors",
|
| 41 |
+
"category": "main",
|
| 42 |
+
"backbone": "facebook/vit-mae-huge",
|
| 43 |
+
"image_size": 224,
|
| 44 |
+
"patch_size": 14,
|
| 45 |
+
"embed_dim": 1280,
|
| 46 |
+
"num_sparse_tokens": 16,
|
| 47 |
+
"num_decoder_layers": 6,
|
| 48 |
+
"spatial_temp": 0.06,
|
| 49 |
+
"do_recon": true,
|
| 50 |
+
"do_clustering": true,
|
| 51 |
+
"do_cls": true,
|
| 52 |
+
"num_clusters": 16384,
|
| 53 |
+
"recon_type": "vq"
|
| 54 |
+
},
|
| 55 |
+
"stellar-b8": {
|
| 56 |
+
"weights": "stellar-b8.safetensors",
|
| 57 |
+
"category": "ablation",
|
| 58 |
+
"backbone": "facebook/vit-mae-base",
|
| 59 |
+
"image_size": 224,
|
| 60 |
+
"patch_size": 16,
|
| 61 |
+
"embed_dim": 768,
|
| 62 |
+
"num_sparse_tokens": 8,
|
| 63 |
+
"num_decoder_layers": 6,
|
| 64 |
+
"spatial_temp": 0.06,
|
| 65 |
+
"do_recon": true,
|
| 66 |
+
"do_clustering": true,
|
| 67 |
+
"do_cls": true,
|
| 68 |
+
"num_clusters": 16384,
|
| 69 |
+
"recon_type": "vq"
|
| 70 |
+
},
|
| 71 |
+
"stellar-b24": {
|
| 72 |
+
"weights": "stellar-b24.safetensors",
|
| 73 |
+
"category": "ablation",
|
| 74 |
+
"backbone": "facebook/vit-mae-base",
|
| 75 |
+
"image_size": 224,
|
| 76 |
+
"patch_size": 16,
|
| 77 |
+
"embed_dim": 768,
|
| 78 |
+
"num_sparse_tokens": 24,
|
| 79 |
+
"num_decoder_layers": 6,
|
| 80 |
+
"spatial_temp": 0.06,
|
| 81 |
+
"do_recon": true,
|
| 82 |
+
"do_clustering": true,
|
| 83 |
+
"do_cls": true,
|
| 84 |
+
"num_clusters": 16384,
|
| 85 |
+
"recon_type": "vq"
|
| 86 |
+
}
|
| 87 |
+
}
|
| 88 |
+
}
|
examples/reconstruction.ipynb
ADDED
|
@@ -0,0 +1,203 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cells": [
|
| 3 |
+
{
|
| 4 |
+
"cell_type": "markdown",
|
| 5 |
+
"id": "a0703884",
|
| 6 |
+
"metadata": {},
|
| 7 |
+
"source": [
|
| 8 |
+
"# STELLAR image reconstruction\n",
|
| 9 |
+
"\n",
|
| 10 |
+
"This notebook runs the **full reconstruction pipeline** end to end:\n",
|
| 11 |
+
"\n",
|
| 12 |
+
"```\n",
|
| 13 |
+
"image → encoder → sparse + spatial tokens → low-rank dense map\n",
|
| 14 |
+
" → ViT decoder → VQGAN decoder → RGB pixels\n",
|
| 15 |
+
"```\n",
|
| 16 |
+
"\n",
|
| 17 |
+
"The released STELLAR checkpoints predict [MaskGIT-VQGAN](https://huggingface.co/fun-research/TiTok)\n",
|
| 18 |
+
"tokens, so you also need the VQGAN tokenizer to decode tokens back to pixels.\n",
|
| 19 |
+
"\n",
|
| 20 |
+
"**Requirements:** `torch`, `safetensors`, `huggingface_hub`, `pillow`, `matplotlib`, and the\n",
|
| 21 |
+
"STELLAR model code on your `PYTHONPATH` (so `from src.models.stellar_model import STELLARModel` works)."
|
| 22 |
+
]
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"cell_type": "code",
|
| 26 |
+
"execution_count": null,
|
| 27 |
+
"id": "3f852740",
|
| 28 |
+
"metadata": {},
|
| 29 |
+
"outputs": [],
|
| 30 |
+
"source": [
|
| 31 |
+
"import os, sys\n",
|
| 32 |
+
"# load_stellar.py and the `src/` model code live in the repo root (one level up).\n",
|
| 33 |
+
"sys.path.insert(0, os.path.abspath(\"..\"))\n",
|
| 34 |
+
"\n",
|
| 35 |
+
"import torch\n",
|
| 36 |
+
"import torch.nn.functional as F\n",
|
| 37 |
+
"from huggingface_hub import hf_hub_download\n",
|
| 38 |
+
"from load_stellar import load_stellar, list_models\n",
|
| 39 |
+
"\n",
|
| 40 |
+
"print(\"Available models:\", list_models())\n",
|
| 41 |
+
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\""
|
| 42 |
+
]
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"cell_type": "markdown",
|
| 46 |
+
"id": "fa4219aa",
|
| 47 |
+
"metadata": {},
|
| 48 |
+
"source": [
|
| 49 |
+
"## 1. Download the MaskGIT-VQGAN tokenizer\n",
|
| 50 |
+
"\n",
|
| 51 |
+
"STELLAR reuses the MaskGIT-VQGAN tokenizer from TiTok. We only need its decoder to turn\n",
|
| 52 |
+
"predicted tokens back into pixels."
|
| 53 |
+
]
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"cell_type": "code",
|
| 57 |
+
"execution_count": null,
|
| 58 |
+
"id": "8a6807dc",
|
| 59 |
+
"metadata": {},
|
| 60 |
+
"outputs": [],
|
| 61 |
+
"source": [
|
| 62 |
+
"vq_path = hf_hub_download(\n",
|
| 63 |
+
" repo_id=\"fun-research/TiTok\",\n",
|
| 64 |
+
" filename=\"maskgit-vqgan-imagenet-f16-256.bin\",\n",
|
| 65 |
+
")\n",
|
| 66 |
+
"print(\"VQGAN tokenizer:\", vq_path)"
|
| 67 |
+
]
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"cell_type": "markdown",
|
| 71 |
+
"id": "d3ebc639",
|
| 72 |
+
"metadata": {},
|
| 73 |
+
"source": [
|
| 74 |
+
"## 2. Load STELLAR for reconstruction\n",
|
| 75 |
+
"\n",
|
| 76 |
+
"`purpose=\"reconstruct\"` builds the encoder **and** decoder and attaches the VQGAN tokenizer."
|
| 77 |
+
]
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"cell_type": "code",
|
| 81 |
+
"execution_count": null,
|
| 82 |
+
"id": "8b2e344f",
|
| 83 |
+
"metadata": {},
|
| 84 |
+
"outputs": [],
|
| 85 |
+
"source": [
|
| 86 |
+
"model = load_stellar(\"stellar-b16\", purpose=\"reconstruct\", vq_model=vq_path, device=device)\n",
|
| 87 |
+
"model.eval();"
|
| 88 |
+
]
|
| 89 |
+
},
|
| 90 |
+
{
|
| 91 |
+
"cell_type": "markdown",
|
| 92 |
+
"id": "0c143298",
|
| 93 |
+
"metadata": {},
|
| 94 |
+
"source": [
|
| 95 |
+
"## 3. Load and preprocess an image\n",
|
| 96 |
+
"\n",
|
| 97 |
+
"Pass **raw** images in `[0, 1]`, resized to `224×224` with bicubic interpolation.\n",
|
| 98 |
+
"ImageNet mean/std normalization is applied **inside** the model — do not normalize yourself."
|
| 99 |
+
]
|
| 100 |
+
},
|
| 101 |
+
{
|
| 102 |
+
"cell_type": "code",
|
| 103 |
+
"execution_count": null,
|
| 104 |
+
"id": "9deaef69",
|
| 105 |
+
"metadata": {},
|
| 106 |
+
"outputs": [],
|
| 107 |
+
"source": [
|
| 108 |
+
"from PIL import Image\n",
|
| 109 |
+
"import numpy as np\n",
|
| 110 |
+
"\n",
|
| 111 |
+
"# Replace with your own image path.\n",
|
| 112 |
+
"img_path = hf_hub_download(repo_id=\"huggingface/documentation-images\",\n",
|
| 113 |
+
" filename=\"beignets-task-guide.png\", repo_type=\"dataset\")\n",
|
| 114 |
+
"pil = Image.open(img_path).convert(\"RGB\")\n",
|
| 115 |
+
"\n",
|
| 116 |
+
"image = torch.from_numpy(np.asarray(pil)).permute(2, 0, 1).float() / 255.0 # (3, H, W) in [0,1]\n",
|
| 117 |
+
"image = F.interpolate(image[None], size=(224, 224), mode=\"bicubic\", align_corners=False).clamp(0, 1)\n",
|
| 118 |
+
"image = image.to(device)\n",
|
| 119 |
+
"print(\"input:\", tuple(image.shape), float(image.min()), float(image.max()))"
|
| 120 |
+
]
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"cell_type": "markdown",
|
| 124 |
+
"id": "f845698f",
|
| 125 |
+
"metadata": {},
|
| 126 |
+
"source": [
|
| 127 |
+
"## 4. Encode, then reconstruct\n",
|
| 128 |
+
"\n",
|
| 129 |
+
"First `encode` the image into its factorized features (sparse concept tokens + spatial\n",
|
| 130 |
+
"maps). `reconstruct` is the decoder half of STELLAR: it takes those factorized features\n",
|
| 131 |
+
"\n",
|
| 132 |
+
"and decodes them all the way to pixels. It returns the RGB pixels (`reconstruction`), thepredicted VQGAN token ids (`tokens`), and the raw codebook logits (`logits`)."
|
| 133 |
+
]
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"cell_type": "code",
|
| 137 |
+
"execution_count": null,
|
| 138 |
+
"id": "6e986c51",
|
| 139 |
+
"metadata": {},
|
| 140 |
+
"outputs": [],
|
| 141 |
+
"source": [
|
| 142 |
+
"with torch.no_grad():\n",
|
| 143 |
+
" features = model.encode(image) # factorized features\n",
|
| 144 |
+
" out = model.reconstruct(features) # or model.reconstruct(features[\"sparse\"], features[\"spatial\"])\n",
|
| 145 |
+
"\n",
|
| 146 |
+
"print(\"factorized features:\")\n",
|
| 147 |
+
"\n",
|
| 148 |
+
"for k in (\"sparse\", \"spatial\"): print(f\" {k:14s}\", tuple(v.shape))\n",
|
| 149 |
+
"\n",
|
| 150 |
+
" print(f\" {k:8s}\", tuple(features[k].shape))for k, v in out.items():\n",
|
| 151 |
+
"print(\"reconstruction outputs:\")"
|
| 152 |
+
]
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"cell_type": "markdown",
|
| 156 |
+
"id": "e74f4664",
|
| 157 |
+
"metadata": {},
|
| 158 |
+
"source": [
|
| 159 |
+
"## 5. Display original vs. reconstruction"
|
| 160 |
+
]
|
| 161 |
+
},
|
| 162 |
+
{
|
| 163 |
+
"cell_type": "code",
|
| 164 |
+
"execution_count": null,
|
| 165 |
+
"id": "d552f2c2",
|
| 166 |
+
"metadata": {},
|
| 167 |
+
"outputs": [],
|
| 168 |
+
"source": [
|
| 169 |
+
"import matplotlib.pyplot as plt\n",
|
| 170 |
+
"\n",
|
| 171 |
+
"orig = image[0].permute(1, 2, 0).cpu().numpy()\n",
|
| 172 |
+
"recon = out[\"reconstruction\"][0].permute(1, 2, 0).cpu().numpy()\n",
|
| 173 |
+
"\n",
|
| 174 |
+
"fig, ax = plt.subplots(1, 2, figsize=(8, 4))\n",
|
| 175 |
+
"ax[0].imshow(orig); ax[0].set_title(\"original\"); ax[0].axis(\"off\")\n",
|
| 176 |
+
"ax[1].imshow(recon); ax[1].set_title(\"STELLAR reconstruction\"); ax[1].axis(\"off\")\n",
|
| 177 |
+
"plt.tight_layout(); plt.show()"
|
| 178 |
+
]
|
| 179 |
+
},
|
| 180 |
+
{
|
| 181 |
+
"cell_type": "markdown",
|
| 182 |
+
"id": "10c7f3d1",
|
| 183 |
+
"metadata": {},
|
| 184 |
+
"source": [
|
| 185 |
+
"## Notes\n",
|
| 186 |
+
"\n",
|
| 187 |
+
"- Output resolution: **224×224** for the `/16` models (`stellar-b16`, `stellar-l16`, and the\n",
|
| 188 |
+
" `b8`/`b24` ablations) and **256×256** for the `/14` `stellar-h16` model.\n",
|
| 189 |
+
"- For larger / higher-fidelity reconstructions, swap in `stellar-l16` or `stellar-h16`.\n",
|
| 190 |
+
"- `reconstruct` takes the factorized features (not the image), so you can edit / analyze the\n",
|
| 191 |
+
" sparse tokens before decoding. It is gradient-transparent — wrap it in `torch.no_grad()`\n",
|
| 192 |
+
" for inference, or keep gradients enabled (e.g. for feature-inversion experiments)."
|
| 193 |
+
]
|
| 194 |
+
}
|
| 195 |
+
],
|
| 196 |
+
"metadata": {
|
| 197 |
+
"language_info": {
|
| 198 |
+
"name": "python"
|
| 199 |
+
}
|
| 200 |
+
},
|
| 201 |
+
"nbformat": 4,
|
| 202 |
+
"nbformat_minor": 5
|
| 203 |
+
}
|
factorization.svg
ADDED
|
|
load_stellar.py
ADDED
|
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Convenience loader for the STELLAR models hosted on the Hugging Face Hub.
|
| 2 |
+
|
| 3 |
+
Requires the STELLAR code (https://github.com/microsoft/STELLAR) to be installed
|
| 4 |
+
and importable, plus ``huggingface_hub`` and ``safetensors``::
|
| 5 |
+
|
| 6 |
+
pip install huggingface_hub safetensors
|
| 7 |
+
# and install the STELLAR repo so that `src` is importable
|
| 8 |
+
|
| 9 |
+
The released checkpoints contain every trained STELLAR module, so a single file
|
| 10 |
+
serves three purposes via the ``purpose`` argument:
|
| 11 |
+
|
| 12 |
+
* ``"encode"`` – feature extraction (encoder only). No extra files needed.
|
| 13 |
+
* ``"reconstruct"`` – image reconstruction (encoder + decoder). Needs the
|
| 14 |
+
MaskGIT-VQGAN tokenizer (``vq_model=<path>``).
|
| 15 |
+
* ``"pretrain"`` – continued pretraining (all modules). Needs ``vq_model``.
|
| 16 |
+
|
| 17 |
+
Usage::
|
| 18 |
+
|
| 19 |
+
from load_stellar import load_stellar, list_models
|
| 20 |
+
|
| 21 |
+
print(list_models()) # available checkpoints
|
| 22 |
+
model = load_stellar("stellar-b16") # purpose="encode" (default)
|
| 23 |
+
|
| 24 |
+
import torch
|
| 25 |
+
img = torch.rand(1, 3, 224, 224) # values in [0, 1]
|
| 26 |
+
with torch.no_grad():
|
| 27 |
+
out = model.encode(img)
|
| 28 |
+
out["sparse"] # (1, K, D) sparse concept tokens
|
| 29 |
+
out["dense"] # (1, P, D) dense patch features
|
| 30 |
+
out["spatial"] # (1, P, K) per-token spatial maps
|
| 31 |
+
"""
|
| 32 |
+
import json
|
| 33 |
+
|
| 34 |
+
import torch
|
| 35 |
+
from huggingface_hub import hf_hub_download
|
| 36 |
+
from safetensors.torch import load_file
|
| 37 |
+
|
| 38 |
+
from src.models.stellar_model import STELLARModel
|
| 39 |
+
|
| 40 |
+
REPO_ID = "microsoft/STELLAR"
|
| 41 |
+
PURPOSES = ("encode", "reconstruct", "pretrain")
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def _load_index(repo_id=REPO_ID):
|
| 45 |
+
# Downloading config.json is what registers a download on the Hub, so always
|
| 46 |
+
# fetch it through hf_hub_download (do not read a local copy).
|
| 47 |
+
return json.load(open(hf_hub_download(repo_id, "config.json")))
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def list_models(repo_id=REPO_ID):
|
| 51 |
+
"""Return the list of available model names."""
|
| 52 |
+
return list(_load_index(repo_id)["models"].keys())
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
def load_stellar(model="stellar-b16", purpose="encode", vq_model=None,
|
| 56 |
+
repo_id=REPO_ID, device="cpu"):
|
| 57 |
+
"""Download a STELLAR checkpoint from the Hub and build a ready-to-use model.
|
| 58 |
+
|
| 59 |
+
Args:
|
| 60 |
+
model: one of ``list_models()`` (e.g. ``"stellar-b16"``).
|
| 61 |
+
purpose: ``"encode"`` (features), ``"reconstruct"`` (decoder) or
|
| 62 |
+
``"pretrain"`` (all modules).
|
| 63 |
+
vq_model: path to the MaskGIT-VQGAN tokenizer
|
| 64 |
+
(``maskgit-vqgan-imagenet-f16-256.bin``). Required for
|
| 65 |
+
``reconstruct``/``pretrain`` when the model decodes into VQ tokens.
|
| 66 |
+
repo_id: Hugging Face repo id (default ``microsoft/STELLAR``).
|
| 67 |
+
device: device to move the model to.
|
| 68 |
+
"""
|
| 69 |
+
if purpose not in PURPOSES:
|
| 70 |
+
raise ValueError(f"purpose must be one of {PURPOSES}, got {purpose!r}")
|
| 71 |
+
|
| 72 |
+
index = _load_index(repo_id)
|
| 73 |
+
if model not in index["models"]:
|
| 74 |
+
raise ValueError(f"Unknown model '{model}'. Available: {list(index['models'])}")
|
| 75 |
+
cfg = index["models"][model]
|
| 76 |
+
|
| 77 |
+
weights_path = hf_hub_download(repo_id, cfg["weights"])
|
| 78 |
+
state_dict = load_file(weights_path)
|
| 79 |
+
|
| 80 |
+
needs_vq = cfg.get("recon_type") == "vq"
|
| 81 |
+
if purpose in ("reconstruct", "pretrain") and needs_vq and vq_model is None:
|
| 82 |
+
raise ValueError(
|
| 83 |
+
f"'{model}' reconstructs into VQ tokens; pass vq_model=<path to "
|
| 84 |
+
"maskgit-vqgan-imagenet-f16-256.bin> (download it from "
|
| 85 |
+
"https://huggingface.co/fun-research/TiTok).")
|
| 86 |
+
|
| 87 |
+
common = dict(
|
| 88 |
+
num_sparse_tokens=cfg["num_sparse_tokens"],
|
| 89 |
+
num_decoder_layers=cfg["num_decoder_layers"],
|
| 90 |
+
spatial_temp=cfg["spatial_temp"],
|
| 91 |
+
do_cls=cfg["do_cls"],
|
| 92 |
+
vit_pretrained=cfg["backbone"],
|
| 93 |
+
)
|
| 94 |
+
if purpose == "encode":
|
| 95 |
+
net = STELLARModel(**common, do_recon=False, do_clustering=False, vq_model=None)
|
| 96 |
+
elif purpose == "reconstruct":
|
| 97 |
+
net = STELLARModel(**common, do_recon=True, do_clustering=False, vq_model=vq_model)
|
| 98 |
+
else: # pretrain
|
| 99 |
+
net = STELLARModel(**common, do_recon=True, do_clustering=True, vq_model=vq_model)
|
| 100 |
+
|
| 101 |
+
missing, _ = net.load_state_dict(state_dict, strict=False)
|
| 102 |
+
# `tokenizer.*` come from `vq_model`, never from our weights. Anything else
|
| 103 |
+
# the built model still needs is a genuine error.
|
| 104 |
+
essential_missing = [k for k in missing if not k.startswith("tokenizer.")]
|
| 105 |
+
if essential_missing:
|
| 106 |
+
raise RuntimeError(f"Missing essential weights: {essential_missing[:8]}")
|
| 107 |
+
|
| 108 |
+
net.to(device)
|
| 109 |
+
if purpose == "encode":
|
| 110 |
+
net.eval()
|
| 111 |
+
return net
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
if __name__ == "__main__":
|
| 115 |
+
print("Available models:", list_models())
|
| 116 |
+
m = load_stellar("stellar-b16")
|
| 117 |
+
with torch.no_grad():
|
| 118 |
+
out = m.encode(torch.rand(1, 3, 224, 224))
|
| 119 |
+
print({k: tuple(v.shape) for k, v in out.items() if torch.is_tensor(v)})
|
stellar-b16.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:061489f86230a9609c5794018dadad3feaecbff9b31df1e4cab702f9c4414500
|
| 3 |
+
size 483090112
|
stellar-b24.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6859aead6faa6095d51d0c60821f1f903c4409ab64fd887527a4906e05993295
|
| 3 |
+
size 483114688
|
stellar-b8.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d9b81262070b5b362c93b8ad24020f6c3de51a6fcce82411cf30ad9b82a1daf0
|
| 3 |
+
size 483065536
|
stellar-h16.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d930de83e0cdbb902a70dbd5a954acf7e69591c5759bca59820677c8ff46ff09
|
| 3 |
+
size 2685054144
|
stellar-l16.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7428079417b5ec89b609d5ef52c3230ef97423dfa62c73264d05e3d98c9944dd
|
| 3 |
+
size 1363369280
|