H3 Video Outpaint · 准备区域引导缓存 (EXP)

You already know what preparation does - encode the source, the audio and the prompt once, cache it, stop paying for it on every window. MiniMaxH3VideoOutpaintPrepareGuidedT8 is that same stage for the region-guided route, where the conditioning isn't a single global prompt but a compiled spatial routing plan.

The reason it's a separate node instead of a checkbox on Prepare is worth understanding, because it tells you where this feature actually sits on the maturity curve. The pack's author deliberately didn't touch the ordinary Prepare node or any existing workflow when adding regional guidance. New node, new cache path, new run name. If the regional route produces garbage on your footage, your working setup is untouched.

The inputs

plan and guidance come in first - guidance is the output of the Guidance node, and this node cross-validates it: the guidance carries its own integrity hash and a plan_sha256, and if it belongs to a different plan the whole thing fails rather than mis-binding prompts to the wrong geometry.

Then the same crowd as the ordinary Prepare: clip, video_vae, prompt, shot_prompts_json, run_name, audio_track, audio_block_tokens, resume_audio, and the optional audio_vae. Identical rules apply - one shot prompt per shot when you supply the array, a hard error if the source has audio and you skipped the audio VAE, and the same run-name constraints.

The difference is what gets encoded. Instead of plain text conditioning, prepare_outpaint_regional_conditioning compiles your regions onto H3's native 32-pixel spatial token grid and produces a conditioning provider object. The report tells you the interesting numbers directly:

• regional_conditioning_sha256 and binding_sha256 - the identity of what got compiled.
• ordinary_prepare_modified: false - a receipt that this route stayed in its lane.
• generated_video_complete: false - the same honest default as the normal Prepare.

Outputs are prepared and preparation_report. The prepared handle here belongs to the Regional Model node, not the plain Sample node - the two carry different conditioning objects, and wiring the wrong one in fails rather than quietly degrading.

Keep the run names apart

This is the gotcha the workflow docs repeat, and it's easy to trigger. Ordinary and guided routes must not share a run_name. The cache root is run_name plus a slice of the plan hash, so two different conditioning encodes under one name is how you get a cache that half-matches and then fails an identity check on the second window. Use outpaint_01 for one and outpaint_guided_01 (the default here) for the other, which is exactly what the shipped workflows do. The reverse case is caught too: try to prepare a plain-prompt run into a name that already holds regional conditioning and the ordinary Prepare stops you with "regional conditioning already exists in this run; choose a new run_name for plain prompts".

You also need a fresh run name whenever you edit regions_json or person_boxes_json. The guidance hash is baked into the conditioning, so tweaking a regional prompt is not a cheap edit - it's a re-prepare.

Install

Manager search MiniMax H3 Audio T8, then restart ComfyUI completely; or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

No pip installs are needed for this node - the pack's requirements.txt is intentionally empty so it can't replace ComfyUI's Torch/CUDA stack. You still need the H3 model files in models/diffusion_models, models/text_encoders and models/vae, plus ffmpeg on PATH for the media handling. For the regional route specifically, the shipped 06_Generate_Guided_Candidate_EXP workflow is the thing to copy rather than the graph you assemble by hand: the validation between Guidance, this node and the Regional Model node is strict enough that a mis-typed wire shows up as a clear error, but only if the rest of the graph is the one the author tested against.