MiniMax H3 Prompt (Qwen) Multimodal

Image-to-video prompting has a problem: H3's official prompt format demands shot-by-shot descriptions of a video that starts from your frame, and describing a picture you're staring at in text is exactly the kind of task people are bad at. This node is the multimodal sibling of the plain MiniMax H3 Prompt (Qwen) node - you hand it one to four reference images, or a video, and a vision-capable Qwen writes the H3 prompt while actually looking at them.

Same pack, same idea, one crucial difference: it's the node that sees.

How it works

It loads a vision GGUF through ComfyUI-llama-cpp_vlm and sends the reference material as base64 images in the chat request. Your image1–image4 inputs are collected in order, and a video input is sampled down to at most max_frames evenly-spaced frames using linear interpolation - so a 100-frame clip becomes 16 representative stills, not a firehose.

The mode input defaults to Auto, and that's the smart bit. With one reference image it runs a detailed single-image analysis - subject, scene, composition, lighting, style - then writes the prompt from your description. With two or more it switches to multi-image story mode: it analyzes each image in order, infers the narrative connections between them (temporal progression, action, scene changes, character continuity), and weaves them into a coherent multi-shot structure. Want explicit behavior instead? Set mode to I2VA (first frame), FL2VA (first + last frame), L2VA (last frame), or Ref2VA (full reference) and it follows that alignment instruction precisely.

The inputs that matter

• image1–image4, video - the references. All optional; anything connected gets analyzed.
• mmproj - this is the one people forget. If you feed any image or video, you must pick a mmproj projector file and a real chat_handler (not "None"). The node hard-errors if you don't, and it's the most common mistake on this node.
• max_size (default 512) - the longest edge reference frames are downscaled to before encoding. Lower it if you're feeding many images and blowing the context window; it directly trades visual detail against token usage.
• max_frames (default 16) - how many stills a reference video becomes.

style, duration, aspect_ratio and include_audio do the same job they do on the plain node - they shape the context H3's format is written against. Outputs are positive_prompt and negative_prompt, wired into your H3 generation node exactly like the text-only version.

Installing it

Identical to the rest of the pack - the install is shared, so if you've done any other node in it you're done:

cd ComfyUI/custom_nodes
git clone https://github.com/wangminxing2019/ComfyUI-MiniMaxH3-QwenPrompt.git
git clone https://github.com/JamePeng/ComfyUI-llama-cpp_vlm.git

Then the difference: this node needs a multimodal GGUF, not just any Qwen. That means a vision Qwen checkpoint with its mmproj projector file, both in ComfyUI/models/LLM, and a chat_handler matching the family (Qwen3.5 for a Qwen3.5-VL-class model, and so on). The README's verified setup uses a local Qwen3.5 GGUF with the Qwen3.5 handler.

Gotchas worth knowing

• Projector mismatch - the mmproj has to match your model family. Mixing a Qwen3-VL projector with a Qwen3.5 base produces garbage or errors, and the dropdown won't stop you.
• Token budget - four reference images plus a sampled video is a lot of vision tokens at the default max_size. If you get context overflows, drop max_size or max_frames before you touch n_ctx.
• NSFW handling - per H3's own guide, the node instructs the model to describe explicit reference content exactly as it appears, without adding clothes. The model may still refuse depending on how you quantized it - abliterated builds behave differently, same as any Qwen instruct model.