[ICCV 2025] Grouped Speculative Decoding for Autoregressive Image Generation

Official PyTorch implementation for the ICCV 2025 paper "Grouped Speculative Decoding for Autoregressive Image Generation (GSD)"

---

Roadmap

• Release Arxiv Paper
• Release Initial Codebase
• Release Colab tutorial
• Apply to more AR models (e.g., Lumina mGPT 2)

---

Usage

1. Download Tokenizer

Our model uses the image tokenizer from Meta's Chameleon. Please download the necessary files from Meta's Chameleon homepage.

Place the downloaded files into the ckpts/chameleon/tokenizer/ directory. The final folder structure should look like this:

ckpts/
└── chameleon/
    └── tokenizer/
        ├── checklist.chk
        ├── text_tokenizer.json
        ├── vqgan.ckpt
        └── vqgan.yaml

2. Install Dependencies

We recommend using PyTorch >= 2.3.0. Install the required packages with pip:

pip install transformers==4.48.1 sentencepiece accelerate>=0.26.0 absl-py

3. Run Tests

Test our GSD model

To generate images with our GSD model, run the following command. You can change the prompt directly within the test_GSD.py file.

python test_GSD.py

Test the baseline (SJD)

To compare the speed with the SJD (ICLR 2025) baseline, run this command:

python test_SJD.py

---

Acknowledgements

This implementation is heavily based on the official repository for SJD by tyshiwo1: tyshiwo1/Accelerating-T2I-AR-with-SJD.