Show-o

Why we included this project

Show-o is a research codebase from NUS Show Lab and ByteDance that folds image and text understanding together with image generation into one transformer, so a single model can caption a photo, answer visual questions, and synthesize new images from a prompt without shuttling between separate systems. Text is decoded autoregressively while image tokens are produced with discrete diffusion, and the Show-o2 follow-up extends the same approach to video and mixed-modality generation at 1.5B and 7B scale. If you are experimenting with native unified multimodal models rather than stitching a vision encoder onto a standalone diffusion pipeline, the repo gives you training and inference code, pretrained weights, and a live Hugging Face demo to try before you build anything. It also works as a reference for how causal and full attention can coexist for different token types. Expect research-grade code with configs and ablation tables, not a plug-and-play product; that is exactly what makes it a decent starting point for digging into the architecture.