Infinity

Why we included this project

Infinity is a CVPR 2025 Oral project that takes autoregressive image generation down to the bit level. Instead of predicting the next token from a large vocabulary, its tokenizer works with an infinite vocabulary and the model predicts individual bits, which the authors argue scales better for high-resolution synthesis. The repo ships 2B and 8B checkpoints that produce 1024x1024 photorealistic images from text prompts, and the reported GenEval and ImageReward scores beat SD3-Medium while generation runs in about 0.8 seconds without extra optimization. Training and inference code, a web demo, and Hugging Face weights are all included, so you can either run the models directly or study the method. If you are evaluating autoregressive alternatives like VAR against diffusion pipelines, this is a useful reference point.