H3 Video Outpaint · 区域引导模型 (EXP)

Compiling region prompts into a guidance object doesn't do anything on its own. MiniMaxH3VideoOutpaintRegionalModelT8 is the node that pushes them into the diffusion model, and it's the step in the regional chain where things either work or fail loudly.

You take a native H3 MODEL - the one out of your checkpoint/UNET loader - plus the prepared handle from the guided Prepare, and you get back a patched MODEL, the same prepared passed through, and a report. That patched model goes into the candidate or sampler node exactly where the unpatched one would have gone.

The one knob

query_chunk_rows, default 256, range 32–2048. It's the number of query rows processed per chunk when applying the regional routing. Lower is gentler on VRAM; higher may be marginally faster. On a 16GB card, if you're already near the ceiling from the sampling canvas, this is the first dial to turn down - and if you're nowhere near it, leave it alone. It is not a quality control.

What it refuses, and why that's the interesting part

The node's own description is unusually candid: it binds prepared regional prompts to a native H3 model, tolerates the audited KJ low-VRAM attention/FFN patches, and rejects LoRAs, other attention owners and unknown wrappers instead of stacking them silently.

That's the right call. In ComfyUI every MODEL-wrangling node in the graph is competing for the same patch chain, and the failure mode when two of them interleave is not an exception - it's a generation that runs fine and looks wrong. The compatibility matrix for this route marks Turbo, SPEED, SLA, VDN, Fast H3 and Prompt Relay as unsupported, fail-closed, rather than waving them through with a warning. If you're wondering why outpaint doesn't work with the eight-step OpenVDN workflow everyone likes: different model architecture and execution contract, per the policy table. That's not laziness, it's that nobody verified the combination and shipping an unverified clever path is how you get a thousand "why does my outpaint look like porridge" threads.

Reading the report

regional_model_report is worth a look once, because it tells you the nature of what got installed. It reports directly_routed_queries: "target_video_only", audio_directly_routed: false, dense_s_by_s_mask_created: false, and - the honest one - semantics_cannot_leak_indirectly_guaranteed: false.

Translate that: the routing biases which query rows see which regional text, it leaves the audio branch alone, and it is not an airtight spatial mask. Attention doesn't respect your rectangle just because you drew one, so a bit of "sky" phrasing can still colour the bottom strip. The report flags human_review_required: true for exactly that reason. Anyone who's used regional prompting at image scale will recognise this - it's the same deal, at video scale, with a heavier cost for getting it wrong.

How it fits the graph

Guided route: Plan → Guidance → guided Prepare → Regional Model → Candidate/Sample → Confirm → Continue → Compose. The regional conditioning provider only survives that path; hand it to the ordinary sampler and it errors, which is the intended guardrail rather than a bug.

Install

Manager: search MiniMax H3 Audio T8, then restart ComfyUI completely. Manual:

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

requirements.txt installs nothing deliberately - the pack keeps Torch, torchaudio, numpy, Pillow and safetensors coming from your ComfyUI install so a node install can't break your CUDA stack. You'll need the H3 FL2VA base model, the Qwen3-VL encoder and both VAEs placed per the pack README, plus ComfyUI-KJNodes for the audited MiniMaxLowVRAMAttention and MiniMaxChunkFeedForward nodes the shipped workflows use, and ffmpeg on PATH.

If it errors

Two likely messages. "Connect the output of Prepare Regional Guidance" means you wired a plain prepared handle in - the node checks the conditioning object's type and refuses anything that isn't the regional provider. A rejection naming a LoRA or an attention wrapper means the compatibility matrix is doing its job: pull the offending node out, or run the audit node first and find out before you queue an hour of sampling.