Show-o

One single transformer unifying multimodal understanding and generation (ICLR 2025): Show-o mixes autoregressive text modeling with discrete-diffusion image generation in one model — one of the first credible academic demonstrations of the native unified multimodal architecture that frontier labs later pursued. 2.0K★, first in the Show-o series.

From Mike Zheng Shou's ShowLab at NUS, joint with ByteDance, NUS-led. The initial model builds on a small pretrained LM backbone (Phi-1.5 class) with a unified discrete tokenizer; the successor Show-o2 scales the recipe natively.