MiniMax H3 Prompt Writer ⚡

MiniMax H3 is a demanding prompter. Its input is a structured envelope - multimodal description, soundscape, music line, keyframe instruction - and the community consensus on these omni-modal video models is that the prompt matters more than on almost anything you've used. This node is the shortcut: type a raw idea, and a local LLM writes the three core fields for you, then wraps them in H3's exact format. No cloud API, no key, nothing leaves your machine.

How it works

The node has two text paths, and it's worth knowing which one you're on.

Path one - dropdown: text_encoder lists a local LLM (the dropdown is populated from your models/text_encoders folder; the architecture is detected from the file). It needs a generation-capable encoder - the tooltip names Gemma-3 12B or a Qwen3-VL - which you'll likely already have if you're running H3, since the model itself uses a big Gemma encoder. The encoder takes your idea, fills in concrete detail where you were vague, and produces the three structured fields.

Path two - clip: the optional clip input overrides the dropdown entirely. That's the smart move when you already have the CLIP loaded for the video model - connect it and the node avoids loading a second multi-gigabyte text encoder just to write a prompt. The dropdown is ignored when a clip is connected, and the tooltip is explicit that reusing your loaded CLIP avoids a reload.

Either way, the node assembles the exact MiniMax envelope around the generated fields, including the keyframe instruction line - same assembly as the sibling MiniMax H3 Prompt Format node.

The inputs that matter

• idea - your raw idea (the default is a lighthouse keeper, a fine starting point). The tooltip's contract: everything you state here is kept; anything vague is filled in with concrete detail. So be specific about what must survive.
• mode - t2va / i2va / fl2va / l2va - text-only, from a first frame, between two frames, or landing on a last frame.
• length - frame preset (124/243/362/481), which sets reference-alignment time and the latest valid shot cut.
• seed - makes a rewrite reproducible.
• temperature - 0.7 default; 0 generates deterministically.
• include_music - off writes non_diegetic_music: N/A, for clips that should carry no score.

The two outputs

• prompt - the assembled, schema-correct H3 prompt. This is the one you wire into CLIPTextEncode.
• raw_output - what the LLM produced before formatting. Wire this to a text viewer if you want to see what the model actually generated, or keep it around for debugging.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/MiniMax H3. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Common issues

The obvious trap is pointing text_encoder at a file that isn't generation-capable - an embedding-only or small model will produce weak fields. If your output prompt reads like it came from a toddler, your encoder is too small; plug in a proper Gemma-3 12B (or a GGUF quant of it, since this pack reads those) or connect your existing clip. Second: if the dropdown is empty, your text_encoders folder is empty - drop a file in and re-add the node so the list refreshes. And remember that H3 itself is a big, license-gated model: the prompt writer is local, but the model you're prompting needs its own hardware and licensing (the H3 community license excludes the US, EU, UK and Korea).