LlamaGen

Why we included this project

This research codebase argues that image generation can work much like language modeling: predict visual tokens one after another with a plain Llama-style architecture, no vision-specific inductive biases needed. LlamaGen ships the pieces to judge that claim yourself, two image tokenizers at downsample ratios 8 and 16, seven class-conditional generators from 100M to 3B parameters, and two text-conditional 700M models, along with PyTorch training and sampling code and a live Hugging Face demo. The released weights let you sample images directly, and inference is supported through vLLM, so you can compare its behavior with diffusion baselines without re-training. It is a research implementation rather than a turnkey production service, but that framing is part of the value: the paper, the weights, and the code together make it easy to study how autoregressive scaling holds up against diffusion.