Upload folder using huggingface_hub

README.md

@@ -1,272 +1,3 @@
-<div align="center">
-
-# RDVQ: Differentiable Vector Quantization for Rate-Distortion Optimization of Generative Image Compression
-
-### CVPR 2026 Oral
-
-[![arXiv](https://img.shields.io/badge/arXiv-2604.10546-orange?style=flat-square)](https://arxiv.org/abs/2604.10546)
-[![CVPR 2026](https://img.shields.io/badge/CVPR-2026-blue?style=flat-square)](https://cvpr.thecvf.com/)
-[![Python 3.10+](https://img.shields.io/badge/Python-3.10+-blue?style=flat-square\&logo=python)](https://www.python.org/)
-[![PyTorch](https://img.shields.io/badge/PyTorch-2.1+-red?style=flat-square\&logo=pytorch)](https://pytorch.org/)
-[![License](https://img.shields.io/badge/License-CC--BY--4.0-lightgrey?style=flat-square)](LICENSE)
-
-[Shiyin Jiang](https://scholar.google.com/citations?user=yf748WAAAAAJ&hl=en) ·
-[Wei Long](https://scholar.google.com/citations?user=CsVTBJoAAAAJ&hl=en) ·
-Minghao Han ·
-[Zhenghao Chen](https://scholar.google.com/citations?user=BThVCu8AAAAJ&hl=en) ·
-[Ce Zhu](http://scholar.google.com/citations?hl=en&user=C7iZbYMAAAAJ) ·
-[Shuhang Gu](https://scholar.google.com/citations?user=-kSTt40AAAAJ&hl=en)
-
-**CVL Lab @ University of Electronic Science and Technology of China**
-
-<img src="assets/framework.jpg" width="85%">
-
-</div>
-
 ---
-
-## 🔥 News
-
-* **Jun 12, 2026** — Training code and configurations are released.
-
+license: apache-2.0
 ---
-
-## 🌟 Introduction
-
-**RDVQ** is a VQ-based generative image compression framework for **efficient and controllable ultra-low-bitrate image compression**.
-
-Conventional VQ-VAE learns powerful discrete representations, but its **non-differentiable nearest-neighbor lookup** decouples representation learning from probability modeling. The entropy model can only predict the resulting code indices, while its rate feedback cannot effectively optimize the encoder. This limits true joint rate-distortion optimization.
-
-RDVQ addresses this issue with a **simple relaxed lookup mechanism**, which builds a differentiable path between encoder features, discrete code indices, and the autoregressive entropy model. As a result, the rate loss can directly guide the encoder to learn more compressible representations, transforming **VQ-VAE from a representation learning framework into a practical learned image codec**.
-
-### Key Features
-
-* **Differentiable VQ-based R-D optimization**
-  Enables joint distortion and rate minimization through relaxed lookup.
-
-* **Multi-scale shared-codebook latents**
-  Provide compact and expressive discrete representations across scales.
-
-* **Masked Transformer entropy model**
-  Estimates accurate probabilities for effective entropy coding.
-
-* **Test-time rate control**
-  Supports bitrate adjustment via prefix transmission and autoregressive completion.
-
-Despite its lightweight design, RDVQ achieves strong perceptual compression performance at ultra-low bitrates while requiring only a small fraction of the parameters used by large generative compression models.
-
----
-
-## 🚀 Performance
-
-### Model Efficiency
-
-<p align="center">
-  <img src="assets/performance.png" width="60%">
-</p>
-
-### Visual Comparison
-
-<p align="center">
-  <img src="assets/visual.jpg" width="95%">
-</p>
-
-### Rate-Distortion Curves
-
-<p align="center">
-  <img src="assets/RD_curves.jpg" width="95%">
-</p>
-
----
-
-## 🛠️ Environment Setup
-
-```bash
-conda create -n RDVQ python=3.10 -y
-conda activate RDVQ
-
-git clone https://github.com/CVL-UESTC/RDVQ.git
-cd RDVQ
-
-pip install "torch>=2.1.0" torchvision --index-url https://download.pytorch.org/whl/cu121
-pip install -r requirements.txt
-```
-
-> **Note:** The default real-bitstream path uses the causal top-k tensor-rANS codec and JIT-builds a small C++17 extension on first use. Please make sure a C++17 compiler is available when running `test_Real.sh`.
-
----
-
-## 📂 Data Preparation
-
-RDVQ expects an **ImageFolder-style** directory where all images are directly placed under one folder:
-
-```text
-/path/to/images/
-  image_0001.png
-  image_0002.jpg
-  ...
-```
-
-Nested subdirectories are not scanned by the testing scripts.
-
----
-
-## 🧪 Testing
-
-RDVQ provides two testing scripts:
-
-| Script         | Bitrate Type           | Description                                           |
-| -------------- | ---------------------- | ----------------------------------------------------- |
-| `test.sh`      | Estimated bitrate      | Reports entropy-estimated bitrate such as `cd_bpp`.   |
-| `test_Real.sh` | Real bitstream bitrate | Reports actual payload bitrate such as `cd_bpp_real`. |
-
-By default, the evaluator reports:
-
-```text
-bpp, lpips, dists, musiq, clipiqa, niqe, psnr, msssim
-```
-
-You can override the metric list with `TEST_METRICS`.
-
-### Quick Start: Estimated-Rate Evaluation
-
-```bash
-TEST_CKPT_PATH=/path/to/checkpoint \
-TEST_IMAGE_DIR=/path/to/kodak \
-TEST_DATASET=kodak \
-bash test.sh
-```
-
-### Quick Start: Real-Bitstream Evaluation
-
-```bash
-TEST_CKPT_PATH=/path/to/checkpoint \
-TEST_IMAGE_DIR=/path/to/kodak \
-TEST_DATASET=kodak \
-bash test_Real.sh
-```
-
-### DIV2K / CLIC Evaluation with FID and KID
-
-For DIV2K and CLIC, please provide `FID_REF_ROOT`:
-
-```bash
-TEST_CKPT_PATH=/path/to/checkpoint \
-TEST_IMAGE_DIR=/path/to/DIV2K_valid_HR \
-TEST_DATASET=div2k \
-FID_REF_ROOT=/path/to/fid_refs \
-bash test.sh
-```
-
-For CLIC:
-
-```bash
-TEST_CKPT_PATH=/path/to/checkpoint \
-TEST_IMAGE_DIR=/path/to/CLIC_valid \
-TEST_DATASET=clic \
-FID_REF_ROOT=/path/to/fid_refs \
-bash test.sh
-```
-
-The evaluator uses:
-
-```text
-<FID_REF_ROOT>/<TEST_DATASET>_256teles
-```
-
-If the reference directory is missing or empty, it will be generated automatically from the original images and reused in later runs.
-
-### Test-Time Rate Control
-
-`test_Real.sh` supports test-time rate control through prefix transmission and autoregressive completion by changing `TEST_TRANSFER_SLICES`:
-
-```bash
-TEST_CKPT_PATH=/path/to/checkpoint \
-TEST_IMAGE_DIR=/path/to/kodak \
-TEST_DATASET=kodak \
-TEST_TRANSFER_SLICES=4 \
-bash test_Real.sh
-```
-
-
-### Useful Debug Options
-
-```bash
-TEST_MAX_IMAGES=1
-TEST_METRICS=bpp,psnr,msssim
-DISABLE_FID=1
-SAVE_IMAGES=0
-```
-
-### Main Environment Variables
-
-| Variable               | Required       | Description                                                          |
-| ---------------------- | -------------- | -------------------------------------------------------------------- |
-| `TEST_CKPT_PATH`       | Yes            | Path to the checkpoint.                                              |
-| `TEST_IMAGE_DIR`       | Yes            | Image folder for evaluation.                                         |
-| `TEST_DATASET`         | Yes            | Dataset label: `kodak`, `div2k`, or `clic`.                          |
-| `FID_REF_ROOT`         | For DIV2K/CLIC | Root directory for FID/KID reference tiles.                          |
-| `FID_REF_DIR`          | Optional       | Manually specified reference tile directory.                         |
-| `TEST_METRICS`         | Optional       | Evaluation metric list.                                              |
-| `TEST_TRANSFER_SLICES` | Optional       | Number of transmitted latent slices for real-bitstream rate control. |
-| `TEST_TOPK`            | Optional       | Top-k/escape entropy width for `test_Real.sh`. Default: `1024`.      |
-| `TEST_MAX_IMAGES`      | Optional       | Maximum number of images for debugging.                              |
-| `DISABLE_FID`          | Optional       | Disable FID/KID computation.                                         |
-| `SAVE_IMAGES`          | Optional       | Whether to save reconstructed images.                                |
-
-### Output Structure
-
-```text
-<checkpoint_stem>/forward/<dataset_name>/
-<checkpoint_stem>/Real/transfer_slices_<N>/<dataset_name>/
-```
-
----
-
-## 🏋️ Training
-
-Training follows a multi-stage pipeline. Please refer to [TRAINING_STAGES.md](TRAINING_STAGES.md) for the full recipe.
-
-### Stage 1: Tokenizer Training
-
-```bash
-bash scripts/tokenizer/train_vq.sh \
-  --data-path /path/to/train/images \
-  --image-size 256 \
-  --vq-model VQ-16-32-64_quant_once \
-  --dataset openimage \
-  --global-batch-size 32 \
-  --results-dir ./results/s1_tokenizer \
-  --codebook-size 4096 \
-  --codebook-embed-dim 32 \
-  --entropy-loss-ratio 0.0 \
-  --lr 1e-4 \
-  --disc-lr 1e-4 \
-  --wo-attn
-```
-
----
-
-## 📖 Citation
-
-If you find RDVQ helpful for your research, please cite:
-
-```bibtex
-@inproceedings{jiang2026rdvq,
-  title     = {Differentiable Vector Quantization for Rate-Distortion Optimization
-               of Generative Image Compression},
-  author    = {Jiang, Shiyin and Long, Wei and Han, Minghao and Chen, Zhenghao
-               and Zhu, Ce and Gu, Shuhang},
-  booktitle = {CVPR},
-  year      = {2026},
-}
-```
-
----
-
-## 💬 Contact
-
-For questions or feedback, please contact:
-
-**Shiyin Jiang**
-📧 **[shiyin.jsy@gmail.com](mailto:shiyin.jsy@gmail.com)**

lam_0.8_tau_0.01.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3390ca00b793b3cae2eb0965fdcd046f686f6efadff06e4f39af72b1e0285c21
+size 1007856746

lam_1.2_tau_0.01.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31c09b2b403966d3baf0ef509c53ff6e6e4027f8a3205c1e5dfcee657d46a986
+size 1007856746

lam_12_tau_0.1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c8b29e5310f2c096e046f7e415e14919135424949006a1fa3198ee6517c9861
+size 1007855640

lam_4.8_tau_0.1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:827aaab038626bb306fc30fac37634eece418dd02bd4617f88e65d8d19e92294
+size 1007856193

lam_7.2_tau_0.1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd5339ec47443bdbe996c7d524927cc6ef4a0ef7f32a07889dc2ba60b5d13ad8
+size 1007856193