MiniMax H3 Apply Reference Plan (Plan v2)

Here's the part that trips people up when they first meet this pack: none of the Plan v2 nodes actually generate video. They prepare a prompt and collect media. The generation happens in ComfyUI's native MiniMax H3 implementation - the MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo nodes that ship with ComfyUI itself. Apply Reference Plan is the bridge between the two worlds, and it's where your carefully-built plan finally touches a real checkpoint.

It does three things. First, it verifies that the h3_prompt and plan_context you're handing it actually match - it won't let you feed a prompt from one branch and a plan from another. Second, it routes every stored reference (images, videos, audio) in native order and builds the conditioning. Third, it constructs the joint audio/video latent that MiniMax's sampler expects.

What you connect

• clip - the MiniMax H3 Qwen3-VL conditioning CLIP from the official workflow. This is the text-encoder CLIP, not an image CLIP.
• vae - the H3 video VAE used for keyframe and reference encoding.
• h3_prompt - from the same node that gave you plan_context (Prompt Merge, the enhancer, Apply Structured Prose, Prompt Review, or Inline Override). Full-prompt manual edits are only accepted if they came through the override or review gate - it's the pack's way of stopping you from hand-editing structure the compiler owns.
• plan_context - the matching compiled plan.
• width / height - target size, multiples of 32. Default is 1344×768. Use the same native size as the sampler workflow.
• ref_image_size - Ref2VA only. match is faster and lower-token; max preserves up to a 2048px short edge for stronger identity fidelity but costs far more reference tokens.
• audio_vae - optional, but required when your plan contains standalone or video-paired reference audio. Forgetting this is a classic "why is there no sound" moment.

Outputs

• positive - native H3 conditioning for Basic Guider or your sampler path.
• latent - the joint audio/video latent. For a Foley plan this is where the magic happens: the source video is VAE-encoded and masked 0 (preserve) while the empty audio stream is masked 1 (generate).
• applied_prompt - the exact prompt that was verified and sent, so you can confirm nothing got swapped.
• adapter_report - mode, checkpoint family, native implementation, target size, and applied routes. Worth a glance on your first run.

Install and prerequisites

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

No Python dependencies from this pack, but you need the H3 weights (MiniMax's 33B omni-modal model, ~42.5GB at full precision - check the HuggingFace repo and the community quant scene for the practical options), plus a ComfyUI build with native H3 support.

The gotcha

The pack's whole design philosophy - compiler-owned structure, no free-form edits - concentrates here. If the adapter rejects your prompt, it's almost always because h3_prompt and plan_context came from different sources. Keep them paired from one node and this is the smoothest part of the pipeline. Also note the community license reality: H3 weights are territory-restricted (US/EU/UK/Korea excluded), so double-check your region before committing to a workflow around it.