Show-o

Show-o is a unified 1.3B-parameter transformer, introduced in August 2024 (arXiv 2408.12528), that handles both multimodal understanding and image generation inside a single model. Instead of going fully autoregressive, it mixes autoregressive modeling for text with discrete diffusion modeling for image tokens, so one network can answer visual questions, generate images from text, do text-guided inpainting and extrapolation, and produce mixed-modality outputs.

The authors report benchmark results competitive with or better than single-purpose models of similar or larger size, and note that diffusion-based image generation needs roughly 20 times fewer sampling steps than autoregressive alternatives. Inference code, training code and model weights were released by the authors. Check the model card for the exact license before commercial use.