H3 Video Outpaint · 扩画范围预览 (EXP)

MiniMaxH3VideoOutpaintGeometryPreviewT8 does one small thing and does it without touching a GPU: it shows you where your original footage sits on the new canvas and which region is going to be invented. It's the second node in the shipped Geometry_Preview_EXP workflow, and honestly it should be the first thing anyone runs on a new clip.

The pitch is simple. Outpainting a 16:9 clip into 9:16 means you're adding a lot of sky and floor, or one huge side wall of nothing. Getting the anchor wrong doesn't produce an error - it produces a video with the subject jammed against one edge and a weird amount of newly generated ceiling. Preview costs a couple of seconds.

How it works

You feed it plan from the Plan node. It picks one source frame - frame_index, zero-based, and it must be a frame that actually exists in the file or it throws - decodes just that frame in its own CPU ffmpeg process, and draws the two rectangles: the source rectangle and the area that will be filled. There's no VAE, no diffusion weights, and no full image batch materialised. preview_max_edge (default 768, range 64–1536) only caps the size of the thumbnail you look at; it does not change the plan, the canvas, or anything about the output.

Outputs are plan passed straight through - handy for mid-graph inspection - plus geometry_preview as an IMAGE and preview_report as JSON. The report is explicit about what you're looking at: model_called: false, source_modified: false, scope: source_frame_and_geometry_only_not_generated_or_accepted. Read it once and the node's honesty will save you from mistaking a checkerboard for a result.

What you're actually looking at

The blue checkerboard marks the region that will be generated. It is not a preview of the outpaint. Nothing in it came from a model, and it tells you nothing about whether the seams will be clean - the pack's own docs are blunt that some sections still come out with strips, repeated textures or visible seams even on a completed 32-second run. This node answers "is my framing right?", full stop.

There's one real trap in the node description, and it's the kind of thing this author likes to spell out: connecting the preview branch alongside your sampling branch does not pause generation. Run it in the standalone geometry-preview workflow, look, adjust, then go back to the generation graph. Gluing it into the heavy graph just gets you a thumbnail that renders after everything has already queued.

Inputs that matter

• frame_index - pick a frame that represents the shot. Frame 0 is often a fade or a title card.
• preview_max_edge - the thumbnail cap. Bump it if you're eyeballing a subtle anchor shift on a wide canvas.

That's the whole node. plan in, small image out.

Install

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

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

No pip dependencies - requirements.txt is deliberately empty so the install can't swap out ComfyUI's Torch. This node needs ffmpeg on PATH for the single-frame decode and nothing else: no H3 weights, no Qwen3-VL encoder, no VAEs. That makes it the one outpaint node you can run on a laptop to check your geometry while the desktop is busy.

When it lies to you

Two failure modes worth knowing. frame_index out of range throws rather than clamping, so a stale index after you swap source files will look like a broken node. And on very small canvases the report sets source_thumbnail_subpixel: true - your source rectangle is less than a pixel wide in the preview and the drawing is useless. Raise preview_max_edge or preview a larger plan.