MiniMax-H3 Guide Prompt (any LLM)

Every guide-based writer in this pack works by putting MiniMax's writing guide in the system prompt and your idea in the user prompt. This node is just that plumbing, exposed: it builds system_prompt and user_prompt from the guide and hands them back as plain strings. No model runs, no VRAM is touched, no time is spent. It's the "bring your own LLM" escape hatch for the whole MiniMax-H3 prompt pipeline.

Why does that exist? Because the writer nodes are locked to their own GGUF runner, and you might already have a perfectly good language model in the loop - an API, an Ollama server, a remote box, or ComfyUI's own text generation. This node covers all five tasks, Ref2VA included, and produces the exact prompts the writers would have built. Wire them into whatever LLM node you already use and you get the same H3-shaped output from a model you already pay for or run elsewhere.

What it outputs

• system_prompt - the guide, as a system message.
• user_prompt - your prompt plus (for the frame tasks) the reference material and the alignment line.
• prompt - both of them in one string, because most LLM nodes take exactly one. Including ComfyUI's own Generate Text, which since 0.30 runs a language model in ComfyUI's own process off a model loaded by CLIPLoader. That's the shortest route to this pack's output with no GGUF downloaded at all - Load CLIP → Generate Text → done.

Inputs that matter

• task - all five: T2VA, I2VA, FL2VA, L2VA, Ref2VA. Ref2VA pulls the longer full-reference guide and its six-section output contract.
• reference_material - what the reference frames or assets show, one per line. Required for Ref2VA; for the other frame tasks it's what anchors the rewrite to your images.
• resolution / duration - folded into the user prompt so the model writes to your aspect ratio and length.
• format - plain (default) or chatml. On plain the two prompts are joined with a blank line and the LLM node applies the model's own chat template, landing the whole guide in the user turn. On chatml the turns are written out yourself, so a Qwen text encoder takes the guide as a real system message. It's Qwen-shaped by construction - on Gemma or anything else, stay on plain.
• auto_download - fetch the guide from MiniMax on first use (it's not bundled with the pack, for licence-territory reasons; ~16 KB / ~24 KB once).

Making Generate Text actually work

Three settings decide whether it does:

• max_length ≈ 2048. Its default of 512 is the output budget, and six Ref2VA sections don't fit in 512 tokens. The pack's own writers use 2048 (max_new_tokens), and a measured Ref2VA rewrite stops on its own at roughly 580 tokens - inside 2048 with room to spare, and cut mid-section at 512.
• thinking off. The guide asks for fields and nothing else; reasoning just spends the budget on prose you'd have to strip.
• use_default_template on, format on plain.

The number to keep in your head: a full guide-based system prompt is ~10,240 tokens of context before the model writes a word. That's what decides whether a given encoder can take this at all - a 4B Qwen3-VL handled a Ref2VA rewrite in about 38 seconds, but the context budget is the thing to check first.

Install

Same pack, same story - ComfyUI Manager (search "MiniMax-H3") or:

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

This node installs nothing extra and downloads nothing on its own; the only fetch is the guide, on first use, under the auto_download switch.

Where people get burned

• Skipping max_length. Default 512 output tokens silently chops the answer mid-section. Set it to ~2048.
• chatml on a non-Qwen model. It's built around Qwen's <|im_start|> template. On Gemma or anything else the turns won't parse - leave format on plain and let the model's own template do the wrapping.
• No model = no output. This node returns strings; it's not a generator. If nothing's wired to its outputs, nothing happens - which is the point, but it surprises people who expect a rewrite.