Improved native unified multimodal models (NeurIPS 2025): autoregressive modeling for text plus flow matching for visual generation, operating on a 3D-causal-VAE latent space with dual-path spatial(-temporal) fusion — extending unified understanding and generation to both images and video. Released as 1.5B and 7B checkpoints (with video-understanding variants) under ShowLab's HuggingFace org.

Joint ShowLab (Mike Zheng Shou) / ByteDance work, NUS-led — the scaled successor of Show-o.

Model Details

Parameters 7B

Variants

Name Parameters Notes
show-o2-1.5B 1.5B
show-o2-7B 7B

Paper

Venue NeurIPS 2025
multimodalgenerationvideoopen-weight

Related