HiDream launches native omni-modal video model, debuts top-10 on both leaderboards

新智元 · wechat · 2026-09-22

HiDream.ai released HiDream-O1-Video-1.0, a native omni-modal video model that debuted at #4 on Artificial Analysis' image-to-video leaderboard (with audio) and #8 on the Arena blind test, joining MiniMax, ByteDance Seedance and Alibaba Wan in the top tier.

Focus on physical plausibility: demos cover Doppler effects, free fall, water balls in space, and material-specific collision behavior; native 5–20s generation with duration driven by narrative, not fixed prompts
A multimodal intent-understanding module converts colloquial prompts into structured shot planning (scene, characters, lighting, sound design)
Joint text-video-audio modeling for native A/V sync, including Chinese lip sync and emotional shifts
Post-training uses Diffusion RL with multimodal rewards, backed by three papers: DMSampler (ICML 2026, cheaper sampling), DiffusionCompass (ECCV 2026, reward-hacking resistance), EvoID (CVPR 2026, identity preservation)
Part of HiDream's UiT foundation spanning image, interactive world model, embodied model and video. Note: promotional framing — physical capability claims rest on official demos

Original post →

HiDream launches native omni-modal video model, debuts top-10 on both leaderboards

More from Multimodal