zhenzhen-minmax-h3-context-ir-lowprice

Read the inputs on Comfly_minmax_h3_context_ir_lowprice ("zhenzhen-minmax-h3-context-ir-lowprice") and it looks like a video generator - images, videos, audio, a seconds dropdown, aspect ratios. It is not. This node does not generate anything. It's MiniMax's official H3 prompt-enhancement endpoint ("IR" = instruction refinement), the tool that takes a rough idea plus reference material and writes you back a much better video prompt as a plain string. The output is result_text, meant to be piped into an H3 video node.

That's an unusual and genuinely useful job. Video models are dramatically better when the prompt is dense and specific, and writing those prompts by hand is the slowest part of most people's video workflows. You feed this node "a guy walking through a neon market at night" plus a reference image, and it returns a few hundred words of scene, camera, and motion description that the actual generation node can chew on. Think of it as the "stuck writer" node.

It lives in the T8mars/Comfyui-zhenzhen pack, the "Comfly" API-wrapper family that calls the author's paid reseller service. It's a "lowprice" node, so it uses the cheaper China-facing endpoint (api.seedance.nz) and takes its key through the shared zhenzhen-seedance2-low-price-api-settings node via api_config. The pack's changelog explicitly flags it as MiniMax's official H3 prompt-enhancement node and API.

The inputs

• model - required, three variants: minmax-h3-context-ir-text (text only), -image (text plus reference images), -multimodal (anything goes).
• prompt - the idea you want enhanced, 1 to 7000 characters. Short is fine - that's the point.
• seconds - 4 to 15; tells the enhancer what length the final video should be.
• ratio - api_default, adaptive, or the standard set; shapes the scene description.
• image1–image9, video1–video3, audio1–audio3 - the reference material for the image/multimodal variants. Consistency references, not generation inputs.

Outputs

• result_text - the enhanced prompt. Wire it into the prompt input of any H3 video node.
• task_id - for the site's async task list.
• response - the raw payload, if you want the full JSON.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen
# restart ComfyUI

Or ComfyUI Manager → "Comfyui-zhenzhen". Add the settings node with your key and wire api_config in.

The honest part

This is one of the smarter nodes in the pack because it monetizes a step you were doing for free in your head - and it usually does it better. The multimodal variant genuinely reads reference images and steers the prompt around them, which is the hard part to do by hand. Caveats are standard: it's a reseller call (your idea and references leave your machine), and the enhanced prompt is only as good as the model behind it, so sanity-check before burning video credits on a bad description. Cheap trick: if you like what it writes, paste it back into other video models too - a good scene description is portable.