MiniMax H3: The New King of AI Video Generation

Text-to-video, image-to-video, reference-to-video — with consistency that finally holds up to scrutiny.

Every six months someone declares AI video "fixed." MiniMax H3 is the first one where the hype mostly checks out. It's the model dominating video generation discussions right now — and for once, the reasons are technical, not just marketing.

What Makes H3 Different

H3 isn't just another text-to-video model. It's built around structured multimodal prompting — the model expects your prompt as a formal structure, not a paragraph. That structure is what gives it its edge:

  • Shot-level direction — camera moves, cuts, and timing described per shot, with chronological scene descriptions.
  • Explicit soundscape — ambient audio, physical action sounds, and dialogue separated from non-diegetic music.
  • Reference-driven modes — image-to-video, first/last frame anchoring, and full reference-to-video where subjects, pictures, videos, and audio are each defined and retained.
  • Identity consistency — the reference mode's explicit subject definitions hold a character's face, clothing, and voice across cuts far better than prior models.

The Modes That Matter

H3 supports several task types, and picking the right one is half the battle:

  • Text-to-video (T2V) — pure prompt to video. Best for original scenes and cinematic abstracts.
  • Image-to-video (I2V) — anchor a first frame, the model animates from it. The workhorse mode.
  • First + last frame (FL2V) — define both endpoints and H3 interpolates the motion between them.
  • Last-frame only (L2V) — converge the entire clip toward a target image. Great for reveals and payoff shots.
  • Reference-to-video (R2V) — the full mode: reference subjects, storyboard pictures, structural videos, and audio all feed the generation with retention analysis.

Prompting H3 Is a Skill

A raw prompt like "a man jumps off a cliff into a lake, slow motion" gets you a mediocre clip. H3's structured format expects three blocks — an integrated multimodal description with shot markers, an overall soundscape, and non-diegetic music — each written to its own spec:

integrated_multimodal_description: [Shot 1] Cinematic, live-action, wide shot of a cliff top at golden hour, camera: slow push-in, 0.2x amplitude... subject: a man in his 30s, bearded, wearing a burnt-orange jacket... [Shot 2] At 00:03.500, the camera cuts to a low angle... ... overall_soundscape: Wind over open water, distant birds, the rustle of fabric as the man runs... non_diegetic_music: Ambient score, 70 BPM, swelling strings entering at 00:02.000...

Writing that by hand, every time, is painful. This is exactly the problem we set out to solve.

Introducing ComfyUI-H3O: Our H3 Node Pack

We built ComfyUI-H3O — a free, open-source ComfyUI node pack for MiniMax H3, now live on our GitHub. Two nodes:

  • H3PromptEnhancer — a vision-LLM prompt enhancer that reformats a raw request into H3's exact structured prompt format. It actually sees your reference images and source video frames, performs a two-pass identity-anchoring analysis, and outputs the ready-to-wire h3_prompt. Supports all five task types, style transfer, same-subject merging, and runs on OpenRouter or local Ollama.
  • H3AspectRatioDetector — feed any image and it emits the exact aspect-ratio option string H3's Resolution Selector expects, orientation-aware. No more manual resolution picking.

Install it in one line:

cd ComfyUI/custom_nodes git clone https://github.com/Indras-Mirror/ComfyUI-H3O.git

Custom LoRAs & Training

Beyond prompting, the next frontier is custom H3 training. We build custom LoRAs for video models — consistent characters, motion styles, brand-specific visuals — and we develop advanced custom ComfyUI nodes. If you need a model to hold a face, a mood, or a house style across clips, that's exactly what we do.

We build custom ComfyUI nodes and workflows for clients, including full custom-node development targeting the H3 ecosystem. Whether you're a studio needing consistent brand motion or a creator wanting a repeatable style, we can build it.

Where H3 Goes Next

MiniMax H3's structured-prompt architecture is a genuine leap — but it's also a hint at where all video models are heading. The teams that master structured prompts, reference retention, and the tooling around them are the teams that will ship the most consistent AI video. That's why we're building in this ecosystem now, and why the nodes are open source.

Want to see what H3 can do for your project? Get in touch — or try the node pack yourself and tell us what to build next.