Nougan MiniMax H3 · Image Edit 🎬

MiniMax H3 is the 33B omni-modal video model that landed on HuggingFace in August 2026 - text, image, video, and audio in one context. ComfyUI's core support gives you MiniMaxH3ImageToVideo, which builds the conditioning but leaves the actual sampling and decoding to you. NouganH3ImageEdit is a drop-in superset of that core node: it accepts the same prompt/width/height/frame interface, exposes the conditioning and latent for advanced wiring, and runs the full sampling-and-decode pipeline so you get a finished image out the other side.

How it works

Under the hood it's a thin but real wrapper around the core H3 node. It resolves target dimensions - explicit width/height (rounded to multiples of 32), or auto from your input image scaled to a megapixels target, or a 1344×768 default when there's no image. It builds the conditioning through the core node, samples with comfy.sample.sample using your sampler/scheduler/cfg/denoise settings, decodes through the VAE, and - only in edit mode - resizes the result back to your source image's original dimensions with the resize_back_method of your choice.

frame_length maps to the core node's length, and since H3 is fundamentally a video model, the node decodes a clip and picks out the frame at batch_index. Feed it one image and it edits; leave the image empty and it generates fresh.

The inputs that matter

• model, clip, vae - the H3 components from your loaders.
• prompt - your edit instruction.
• image - the first frame. Connect it to edit, leave empty to generate.
• width / height - 0 = auto. When auto and an image is present, megapixels (default 1.0) sets the auto-size target.
• frame_length - the core node's length; 1 is a single-frame edit.
• seed, steps, cfg, sampler_name, scheduler, denoise - standard sampling controls. Note the default cfg of 1.0 and that the negative is just an empty-string encoding - at cfg 1 it's effectively unused, which is normal for this class of model.
• last_frame - optional second boundary frame for clip-style edits.

Outputs: image (the finished frame, resized back in edit mode), native_res (at the sampled resolution, before resize-back), positive conditioning, latent, and width/height ints - so you can do something custom with the internals without re-running the whole pipeline.

Installing

Manager search Nougan, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

The hard requirement is a ComfyUI version with core H3 support - the node raises Core node 'MiniMaxH3ImageToVideo' not found. Update ComfyUI… if it isn't there, which is the most common failure mode by far. That's not a pack bug; that's a ComfyUI version check. And the same MiniMax H3 community-license geography caveat applies that applies to every H3 node: the weights are licence-locked out of the US, EU, UK, and South Korea, so check what you're allowed to run before you set up. One node, full edit pipeline, fewer moving parts - that's the sell.