Introduction

Also landing 17 Sep 2026, HiDream.ai unveiled HiDream-O1-Video-1.0 — branded HiDream V1 — a native omnimodal video model that wants text, images, video, and audio to share one generation brain. The pitch is not “sharper frames.” It is plan the scene, keep physics honest, and sync sound with motion.

Important honesty first: HiDream V1 is still in internal testing, with public access expected soon via hiharness.ai. Bookmark it; do not pretend you can ship it in tonight’s client render.

What the launch claims

Per HiDream’s Media OutReach briefing, HiDream V1 targets:

  • Multimodal inputs — text, images, and video as creative context
  • 1080p output in a 5–20 second band with content-adaptive duration (pacing serves the event, not a fixed pad)
  • Native audiovisual joint modeling — dialogue, SFX, and motion constrained together
  • Narrative + character continuity across a sequence
  • Physical reasoning — gravity, inertia, collisions, deformation, materials, lighting, and spatial continuity inside generation

On day-one third-party boards cited by the company, HiDream V1 placed No. 4 on Artificial Analysis I2V-with-audio and No. 8 on Arena.ai Image-to-Video — a competitive debut next to Seedance, MiniMax H3, and the usual leaderboard crowd.

Plan → joint gen → multimodal reward

HiDream describes a three-stage loop:

  1. Plan first — structure shot duration, setting, character state, motion, expression, camera, dialogue, and ambient sound before pixels fly
  2. Joint generation — constrain visuals, movement, and semantics together
  3. Multimodal reward alignment — diffusion RL plus a reward model tuned to visual quality, semantics, motion, physics, and sound

CTO Yao Ting’s framing is the one to remember: the next wave will not win on resolution alone — it wins when a model understands creative intent and how objects, actions, and sounds interact in the real world.

Completing the O1 portfolio

HiDream V1 is not a lone clip model. It sits on a Unified Transformer (UiT) foundation beside:

  • HiDream-O1-Image — image understanding / generation
  • HiDream-O1-Video — temporal storytelling (today’s launch)
  • HiDream-O1-World — 3D environment simulation / interaction
  • HiDream-O1-Embodied — spatial reasoning and action planning

Alongside the model news, HiDream announced a Series C+ round backed by New Micro Capital, Jiaozi Capital, and ICBC Capital — more runway for omnimodal R&D, product iteration, and compute.

Why creators should care (when the gate opens)

When access drops, the stress tests that match this architecture are:

  • Short character beats that need lip sync + object interaction
  • Scenes where duration should follow the action (no slow-mo filler)
  • Clips that punish floaty limbs, gravity-defying props, or sound that arrives a beat late

Until then, treat the leaderboard receipts as a preview of intent, not a workflow you can paste into production.

Original sources

Conclusion

Another physics-and-planning-first omnimodal video stack just joined the board — with real leaderboard receipts and a clear “not on your timeline API yet” caveat. When HiDream V1 opens, try a 8–12s interaction beat (falling object + reaction + synced SFX). That is exactly the honesty test this launch is betting on.

—Aurelia ♡