Add pipeline tag, library name, project page, and link to code

README.md

@@ -1,4 +1,108 @@
 ---
 license: mit
+pipeline_tag: unconditional-image-generation
+library_name: pytorch
 ---
-https://arxiv.org/abs/2503.10568
+
+# Autoregressive Image Generation with Randomized Parallel Decoding
+
+[Haopeng Li](https://github.com/hp-l33)<sup>1</sup>, Jinyue Yang<sup>2</sup>, [Guoqi Li](https://casialiguoqi.github.io)<sup>2,📧</sup>, [Huan Wang](https://huanwang.tech)<sup>1,📧</sup>
+
+<sup>1</sup> Westlake University,
+<sup>2</sup> Institute of Automation, Chinese Academy of Sciences
+
+[![arXiv](https://img.shields.io/badge/arXiv-2503.10568-A42C25?style=flat&logo=arXiv)](https://arxiv.org/abs/2503.10568) [![Project](https://img.shields.io/badge/Project-Page-green?style=flat&logo=Google%20chrome&logoColor=green)](https://hp-l33.github.io/projects/arpg) [![HuggingFace](https://img.shields.io/badge/HuggingFace-Model-blue?style=flat&logo=HuggingFace)](https://huggingface.co/hp-l33/ARPG)
+
+## News
+* **2025-03-14**: The paper and code are released!
+
+## Introduction
+We introduce a novel autoregressive image generation framework named **ARPG**. This framework is capable of conducting **BERT-style masked modeling** by employing a **GPT-style causal architecture**. Consequently, it is able to generate images in parallel following a random token order and also provides support for the KV cache. 
+* 💪 **ARPG** achieves an FID of **1.94**
+* 🚀 **ARPG** delivers throughput **26 times faster** than [LlamaGen](https://github.com/FoundationVision/LlamaGen)—nearly matching [VAR](https://github.com/FoundationVision/VAR)
+* ♻️ **ARPG** reducing memory consumption by over **75%** compared to [VAR](https://github.com/FoundationVision/VAR).
+* 🔍 **ARPG** supports **zero-shot inference** (e.g., inpainting and outpainting).
+* 🛠️ **ARPG** can be easily extended to **controllable generation**.
+
+## Model Zoo
+We provide the model weights pre-trained on ImageNet-1K 256*256.
+| Model | Param | CFG | Step | FID | IS | Weight |
+| :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| ARPG-L | 320 M | 5.0 | 64 | 2.43 | 294 | [arpg_300m.pt](https://huggingface.co/hp-l33/ARPG/blob/main/arpg_300m.pt) |
+| ARPG-XL | 719 M | 6.0 | 64 | 2.10 | 331 | [arpg_700m.pt](https://huggingface.co/hp-l33/ARPG/blob/main/arpg_700m.pt) |
+| ARPG-XXL | 1.3 B | 7.5 | 64 | 1.94 | 340 | [arpg_1b.pt](https://huggingface.co/hp-l33/ARPG/blob/main/arpg_1b.pt) |
+
+## Getting Started
+### Preparation
+To accelerate the training process, we use the ImageNet dataset that has been pre-encoded into tokens, following the approach of [LlamaGen](https://github.com/FoundationVision/LlamaGen). You can directly download the pre-processed [dataset](https://huggingface.co/ziqipang/RandAR/blob/main/imagenet-llamagen-adm-256_codes.tar) provided by [RandAR](https://github.com/ziqipang/RandAR).
+
+### Training
+Taking ARPG-L as an example, the script for training using 8 A800-80GB GPUs is as follows:
+```shell
+torchrun \
+--nnodes=1 --nproc_per_node=8 train_c2i.py \
+--gpt-model ARPG-L \
+--code-path YOUR_DATASET_PATH \
+--epochs 400 \
+--global-batch-size 1024 \
+--lr 4e-4
+```
+Note that the learning rate is configured to be 1e-4 per 256 batch size. That is, if you set the batch size to 768, the lr should be adjusted to 3e-4.
+
+### Evaluation
+1. Prepare ADM evaluation script.
+```shell
+git clone https://github.com/openai/guided-diffusion.git
+
+wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/imagenet/256/VIRTUAL_imagenet256_labeled.npz
+```
+2. Download the [pre-trained weights](https://huggingface.co/FoundationVision/LlamaGen/resolve/main/vq_ds16_c2i.pt) of [LlamaGen](https://github.com/FoundationVision/LlamaGen)'s tokenizer.
+
+3. Reproduce the experimental results of ARPG:
+```shell
+# ARPG-L. The FID should be close to 2.43.
+# PS: cfg-scale=5 outperforms the paper's 4.5 setting.
+torchrun \
+--nnodes=1 --nproc_per_node=8 sample_c2i_ddp.py \
+--gpt-model ARPG-L \
+--gpt-ckpt arpg_300m.pt \
+--vq-ckpt vq_ds16_c2i.pt \
+--cfg-scale 5.0 \
+--step 64
+```
+```shell
+# ARPG-XL. The FID should be close to 2.10.
+torchrun \
+--nnodes=1 --nproc_per_node=8 sample_c2i_ddp.py \
+--gpt-model ARPG-XL \
+--gpt-ckpt arpg_700m.pt \
+--vq-ckpt vq_ds16_c2i.pt \
+--cfg-scale 6.0 \
+--step 64
+```
+```shell
+# ARPG-XXL. The FID should be close to 1.94.
+torchrun \
+--nnodes=1 --nproc_per_node=8 sample_c2i_ddp.py \
+--gpt-model ARPG-XXL \
+--gpt-ckpt arpg_1b.pt \
+--vq-ckpt vq_ds16_c2i.pt \
+--cfg-scale 7.5 \
+--step 64
+```
+Note that the unlisted parameters (such as temperature, top-k, etc.) are all the default values set in `sample_c2i_ddp.py`.
+
+## Citation
+If this work is helpful for your research, please give it a star or cite it:
+```bibtex
+@article{li2025autoregressive,
+    title={Autoregressive Image Generation with Randomized Parallel Decoding},
+    author={Haopeng Li and Jinyue Yang and Guoqi Li and Huan Wang},
+    journal={arXiv preprint arXiv:2503.10568},
+    year={2025}
+}
+```
+
+## Acknowledgement
+
+Thanks to [LlamaGen](https://github.com/FoundationVision/LlamaGen) for its open-source codebase. Appreciate [RandAR](https://github.com/ziqipang/RandAR) and [RAR](https://github.com/bytedance/1d-tokenizer/blob/main/README_RAR.md) for inspiring this work, and also thank [ControlAR](https://github.com/hustvl/ControlAR).