← Back to feed

HiDream.ai launched its video generation model on September 17, 2026, with two benchmark results on day one: fourth on Artificial Analysis’s Image-to-Video (With Audio) leaderboard, and eighth on Arena.ai’s Image-to-Video head-to-head ratings. The model generates 1080p clips from 5 to 20 seconds with audio natively synchronized to video, not bolted on after generation.

HiDream is a Beijing-based company. Its earlier HiDream-O1-Image 1.5 sits at number two globally on Artificial Analysis’s text-to-image leaderboard and number one among Chinese-developed image models. The company had already shipped O1-World (3D environment simulation) and O1-Embodied (spatial reasoning) before the video model arrived, completing the four-model O1 portfolio.

Native Omnimodal Architecture

The distinguishing claim is architectural: HiDream V1 does not generate video and then synthesize matching audio. Text, video, and audio tokens share the same native token space and constrain each other during generation. Visual motion influences sound timing; dialogue and ambient sound shape the emotional tone of scenes. The company calls this “native omnimodal” to distinguish it from models that run separate generation stacks and attempt post-hoc alignment.

The generation pipeline has three stages. In the planning stage, the model structures the full user prompt before any pixels are generated, mapping out shot duration, setting, character states, facial expressions, camera motion, dialogue, and ambient sound. The joint generation stage then runs visuals, motion, and semantics simultaneously under a unified model, with audio woven into the same pass. A multimodal reward signal alignment step follows, optimizing across visual quality, continuity, and physical plausibility.

Character consistency is treated as a sequence-level constraint, not a per-frame optimization. Once a character’s state, emotional arc, and physical context are established in the planning stage, each element updates continuously through the sequence. This approach targets the coherence problems common in current video models — where characters change appearance between cuts, or actions violate simple physics.

Benchmark Position

Artificial Analysis Image-to-Video (With Audio) leaderboard — debut ranking:

• #4 globally across visual quality, prompt adherence, motion quality, narrative coherence, and audiovisual coordination

Arena.ai Image-to-Video (human preference votes):

• #8 globally

The gap between the two rankings indicates that HiDream V1’s native audio synchronization is its primary differentiator. On the AA leaderboard — which weights audiovisual coordination as a distinct scoring dimension — it places four spots higher (#4 vs #8) than on Arena’s pure human preference ranking, which does not isolate audio quality as a category.

Chinese-developed models hold the top positions on both leaderboards. Seedance (ByteDance) and MiniMax H3 are the current leaders. HiDream V1 adds a third Chinese model to the upper tier.

Availability and Funding

HiDream V1 is in internal testing and has not been released publicly. No pricing has been announced.

Alongside the launch, HiDream disclosed the closing of a Series C+ financing round. Investors are New Micro Capital (co-established by Shanghai New Micro Technology Group and Shanghai Science and Technology Venture Capital Group), Jiaozi Capital (state-backed equity platform under Chengdu Jiaozi Financial Holding Group), and ICBC Capital. Round size was not disclosed.

O1 Portfolio

HiDream frames V1 as one component of a four-model omnimodal system built on a shared architecture:

• O1-Image: image understanding and generation (#2 global on AA T2I; #1 Chinese model)
• O1-Video: video generation (this model — #4 AA Image-to-Video with Audio at launch)
• O1-World: 3D environment simulation and real-time interaction (#1 on WBench interactive world model benchmark)
• O1-Embodied: spatial reasoning, action planning, physical feedback (first place in two RoboColiseum categories: disturbance adaptation and spatial reasoning)

The company’s architecture argument: a shared token space across modalities produces cross-modal understanding that isolated stacks cannot replicate. Whether that claim holds against end-to-end multimodal competitors will require third-party evaluation once the model is publicly accessible.