Infinity-2B

Infinity-2B is ByteDance's 2-billion-parameter text-to-image model built on bitwise autoregressive modeling rather than diffusion. A bitwise self-correction technique counters the cumulative error that plagues autoregressive image generation, and the 2B model renders a 1024x1024 image in about 0.8 seconds, several times faster than same-size diffusion models, with 8B and 20B siblings scaling quality upward.

Training code, inference code, demos, and checkpoints are open, and the work was accepted as a CVPR 2025 oral, which gives the approach unusual academic weight for the autoregressive image camp. Caveats: check the model card's license terms before commercial use, and while fast, its aesthetic quality bar is set by the training data and prompt engineering like any open image model.