Inferway Prompt Expand (MiMo)

You have a one-line idea - "old fisherman mending nets at dawn, slow push in" - and H3 wants a paragraph. InferwayPromptExpand hands that idea to Inferway's own chat model and returns something you can wire straight into the Generate node's prompt input. That's the whole job. No second model in VRAM, no Ollama server, no download.

Prompt enhancement as a node is thoroughly mainstream by now, and the interesting community builds do it locally - a small abliterated LLM on your own card, free per call and offline. This is the other trade, and it's worth being clear-eyed about it: zero setup and no VRAM, in exchange for a per-token bill and a server-side filter you can't remove. There is no abliterated version of someone else's endpoint.

What it's actually doing

It calls POST /v1/chat/completions non-streaming against inferway/mimo-v2.6-flash - Inferway's own model, not MiniMax H3 - with a system prompt that constrains the output hard: exactly one paragraph of usable prompt text, describing subject, action, scene, camera movement, lighting and atmosphere, under 120 English words or 200 Chinese characters, no preamble, no explanation, no Markdown, no quote marks around the result, and an instruction never to invent people's names or brands you didn't mention. That last line matters. The classic failure of an enhancer is subject drift: it "improves" your idea by adding narrative you never asked for, and you only notice after paying a video model to render it. This one is scoped against the obvious version of that.

Only three inputs, all of them self-explanatory:

• idea - your short prompt. Blank fails locally before any request goes out, so you can't accidentally pay for nothing.
• language - auto, zh or en. auto answers in the language of the idea, which is usually what you want.
• profile - and this one is not a formality: this profile's balance pays for the MiMo tokens.

One output, prompt, a STRING that plugs into Generate. It's not an output node, so it only runs when something downstream actually needs it.

Billing, caching, and the one warning you'll see

The call is billed per token, separately from any video generation. Re-queuing with identical inputs usually reuses ComfyUI's cache, so the same idea doesn't get paid for twice - the author left the node's caching at the default on purpose. But that's best effort, and the honest list of what breaks it is in the README: a ComfyUI restart, --cache-none, or the cache being evicted under memory pressure re-runs the expansion and bills it again. At temperature 0.7 you may not even get the same text back.

A failed expansion isn't cached either, and the sharpest edge is a timeout arriving after the service already finished and billed the tokens - re-queue then and you can pay for two expansions.

If the reply hits the token ceiling, the node keeps the text and puts 提示词被截断… / Prompt truncated… on the node itself. That's the author being straight with you: the paragraph may end mid-thought. Read it before it reaches a paid Generate node, or shorten the idea and expand again.

Installing it

Same pack as the rest of the Inferway nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/inferway/comfyui-inferway
/path/to/ComfyUI/.venv/bin/python -m pip install -r comfyui-inferway/requirements.txt

Or search Inferway H3 in ComfyUI Manager (registry package inferway-comfy) and restart. The only dependency is httpx; there's no model file to fetch and no GPU work involved. Set INFERWAY_API_KEY in the environment of the ComfyUI server process and restart, or save a key from the in-app dialog under Inferway → Manage Inferway API key, then pick it in profile.

Should you use it?

If you're already paying for H3, yes - a few tokens of rewriting to get a prompt that suits the model is cheap next to the video, and the node is honest about truncation and caching. If you're doing this constantly, put a small local LLM in the graph instead: it's free per call, it doesn't filter, and it never bills you twice because a request timed out. The pack's own error vocabulary is a good sign - bilingual sentences, a raw code for support, and never a trace of your API key - but it's still a metered call to someone else's server. Use it like one.

One last piece of context: the same pack's Generate node is the only realistic route to H3 for most people, since the open weights are licence-locked out of the US, EU, UK and Korea and enormous besides. If you're on that path, a prompt assistant costing pennies is a rounding error.