FoundationVision/LlamaGen provides autoregressive image generation models (LLama-based) with image/tokenizers and text-conditional variants, plus a vLLM-based serving option. It includes AR and VQ-VAE models and demos across ImageNet and COCODATA, with GitHub activity since 2024.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Autoregressive Model Beats Diffusion: 🦙 Llama for Scalable Image Generation. The repo contains pre-trained model weights and training/sampling PyTorch codes for autoregressive image generation using Llama-style models, plus tokenizers and examples for class- and text-conditional generation. It also references online demos and a serving framework integration.
How it works
The project applies a next-token prediction paradigm from large language models to visual generation, using image tokenizers at multiple downsample ratios and AR models ranging from 100M to 3B parameters for class-conditional generation, plus 700M parameter text-conditional models. It supports a vLLM serving framework to enable higher throughput.
Getting started
Getting started directions are spread across the repository, including getting started material in GETTING_STARTED.md and a serving readme for the vLLM setup. Demo usage shows sample scripts for running AR sampling and serving:
- "python3 autoregressive/sample/sample_c2i.py --vq-ckpt ./pretrained_models/vq_ds16_c2i.pt --gpt-ckpt ./pretrained_models/c2i_L_384.pt --gpt-model GPT-L --image-size 384"
- "python3 autoregressive/serve/sample_c2i.py --vq-ckpt ./pretrained_models/vq_ds16_c2i.pt --gpt-ckpt ./pretrained_models/c2i_XXL_384.pt --gpt-model GPT-XXL --from-fsdp --image-size 384"
- Gradio/Space demos are linked via Hugging Face Spaces.
Recent releases
The README lists update notes with dates such as 2024.06.28 and 2024.06.11, indicating release activity and model/tokenizer availability.
Traction
Repository stars: 1964
Behind the repo
Not provided in the facts block.
Caveats
License: MIT. Created 2024-06-03. Last push 2024-08-15. Issues: 73. Language: Python. It includes multiple model weights and training/sampling code files; the license and data usage terms are noted in the License section.