APNext H3 Prompt Writer

MiniMax-H3 is the newest open-weights video model, and it is brutally picky about how you write for it. It wants a heavily structured prompt - integrated_multimodal_description, overall_soundscape, non_diegetic_music, shot blocks, <d> dialogue tags, exact reference-alignment lines depending on whether you're doing text-to-video or starting from frames. Hand-writing that is miserable and easy to get subtly wrong, and with a model this new, "subtly wrong" means a full generation that ignores half of what you said. Video models already reward long, detailed prompts (ask anyone who fought LTX's early prompt adherence); H3 turns that up a notch.

The APNext H3 Prompt Writer turns a short idea - or an image - into a complete, spec-compliant H3 prompt, then hands it to whatever LLM you want: cloud Claude, GPT, Gemini, Grok, Groq, or a local server via Ollama/LM Studio/vLLM. You don't need to memorize the guide; the pack ships the official MiniMax writing guide verbatim in data/h3/ and uses it as the system prompt, so the model follows the real spec rather than a paraphrase.

How it works

The official guide becomes the system prompt; your idea plus the settings below become the request. task_type picks the format - T2VA (text only), I2VA (first frame), FL2VA (first + last), L2VA (last frame) - and anything other than T2VA makes the node emit the exact reference-alignment instruction line with the right S.SS timing from duration_seconds. The wildness dial (0–100) maps to creative-latitude bands: 0–15 strictly literal, 16–40 grounded, and above 40 it injects concrete surreal events (not mood words) drawn from a seeded pool - same seed, same weirdness. dialogue_language tells the model to actually write the spoken lines in that language inside <d>[Language] …</d>, which matters for small local models that would otherwise happily write English and label it as something else.

The inputs that matter

• idea - the short prompt or image description being expanded. The core input.
• task_type + duration_seconds - what format you're targeting and how long the video is.
• wildness - 25 is a sane default; push past 40 for deliberately weird shots.
• camera_motion / camera_amplitude / camera_speed - the guide's full camera vocabulary. Medium amplitude and normal speed are omitted from the output, as the spec requires.
• include_dialogue + dialogue_language - off means no (Sx) IDs and no <d> blocks at all; Auto (match the setting) lets the model pick a language that fits the scene.
• model - auto-detect grabs the first provider with an API key set (Claude → GPT → Gemini → Grok → Groq), then the Claude Code CLI, then a running local server. claudecode: entries use your Claude Code login instead of a key.

The optional image sends the frame as vision input so the model describes it itself instead of you writing the description. model_override takes an exact provider:model string (e.g. ollama:qwen3:8b) and beats the dropdown - that's how you reach a local model the dropdown hasn't discovered. local_base_url points a single node at a different host or port.

Outputs: h3_prompt (wire this to the video model) plus integrated_multimodal_description, overall_soundscape and non_diegetic_music split out separately, and model_used so you can see what actually answered.

Install

ComfyUI Manager (search "comfyui_dagthomas") or the usual manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas
pip install -r requirements.txt

Then restart. For cloud providers set the matching key as an environment variable (ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, …) and restart ComfyUI so it picks them up. Local servers don't need a key - anything running when the page loads appears at the bottom of the model dropdown.

Common issues

• <d> tags confuse you - they're not stray markup. They mark spoken audio, and the spec requires them. Don't strip them before generating unless you want the words treated as narration.
• Local model list looks empty - the dropdown only shows what was running when the page loaded. Start the server, refresh, and it appears (no ComfyUI restart). Use a vision-capable model like ollama:qwen3-vl:8b if you're connecting images.
• openai errors - the pack pins openai>=2.54.0,<3.0.0 because v3 broke its client wiring. Don't force-upgrade it.
• Invalid model names - model strings must match what the provider expects exactly, or you get runtime API errors. That's what model_override is for.