MiniMax H3-Image: open-weights 16 MP image model from the H3 family, with 16 references and day-one ComfyUI support

What it is, per the model card. A 33.1B-parameter single-stream diffusion transformer with the H3 architecture minus the audio branch, using the MiniMax-H3 text encoder unchanged. Generation and editing live in one checkpoint, distilled to 6 sampling steps without classifier-free guidance. Output is about 16 MP at every aspect ratio, from 4096x4096 to 6272x2688 (21:9), with no separate upscaler; MiniMax says the working range is roughly 4 to 20 MP.

References and prompts. Edits take one or more reference images, addressed in the prompt as <Picture 1>, <Picture 2> and so on, up to 16 per MiniMax's launch post. The model expects a long, detailed English description; the MiniMax API expands short prompts automatically, and a separate H3-Image-PE prompt expander (17.9 GB, NVFP4) is published for local use. Negative prompts are not supported.

Running it. The BF16 transformer is about 66 GB and the shared H3 text encoder another ~67 GB (skipped if you already have MiniMax-H3), so this is a data-center or cloud-GPU model at full precision. MiniMax lists SGLang, vLLM-Omni, diffusers and ComfyUI; ComfyUI merged native support (PR #17655) and shipped it in v0.45.0, with repackaged files on Comfy-Org's Hugging Face. Community GGUF and quantized uploads were already appearing on Hugging Face within hours. Hosted use is through the MiniMax API and Hailuo AI, with 50% off 16 MP+ generations during launch week, per MiniMax.

Why it matters for video. If you animate H3-family clips (MiniMax-H3, HyperFlow H3 Flash, Pika H3 Flex or the H3 LoRAs we have listed), stills from the same model family are a natural source for first frames, character sheets and location plates. Pruna's community P-H3-Image did a version of this a week earlier; this is MiniMax's own release.

Licence, read before downloading. The weights are under the MiniMax-H3-Image Community License, which excludes the European Union, the United Kingdom, South Korea and the United States (use there needs a separate licence from MiniMax) and requires a 'Powered by MiniMax-H3-Image' notice on products built with it. Using the hosted API or Hailuo app is a different question from self-hosting the weights; check MiniMax's terms for your region.

Signals (HOT). Official @MiniMax_AI launch (about 950 likes within the first hour when we checked), native ComfyUI support merged the same morning, and a run of early-access creator posts on X since Nov 9. Quality and speed claims are MiniMax's; we have not run it.

Our take: for agencies outside the excluded regions, this is the strongest open option so far for H3-consistent keyframes. Everyone else can try the hosted API on a handful of reference-heavy stills before deciding whether it beats their current image model for video prep.