[Frontend] Add MiniMax-H3 latent editing workflows (WF-05)

[FayeSpica]

Purpose

Implement WF-05 from #7380 using the latent-edit interfaces from #7465 and #7575.

1. Add one ComfyUI workflow covering object removal, inpainting, continuation, and extension.
2. Build independent spatial masks with native ComfyUI nodes and calculate temporal masks with MiniMax-H3 Temporal Mask.
3. Preview each mask as a red overlay on source video using native compositing nodes, with Preview above Result for each case.
4. Upload video and audio masks as JSON files to match the server API, including small masks and audio scalars.
5. Add workflow usage guides for the four editing cases.

Workflow Screenshot

Test Plan

Run MiniMax-H3 FL2VA:

vllm serve /root/.cache/huggingface/hub/models--MiniMaxAI--MiniMax-H3/snapshots/6818f6c32d12b210915e44ad56a4228c2608f160/FL2VA \
  --omni \
  --host 0.0.0.0 --port 8000 \
  --trust-remote-code \
  --served-model-name MiniMaxAI/MiniMax-H3 \
  --task-type fl2va \
  --num-gpus 8 --usp 8 --ring 1 \
  --text-encoder-tp-size 8 \
  --enable-distributed-layerwise-offload \
  --vae-parallel-mode tile \
  --vae-use-tiling \
  --vae-patch-parallel-size 8 \
  --enable-diffusion-pipeline-profiler \
  --diffusion-attention-backend FLASH_ATTN \
  --init-timeout 1800

Use the same source clip for all four cases. Object removal and inpainting have independent generated rectangular masks, currently 180 × 168 at x=1074, y=493 on a 1344 × 768 canvas. Adapt their position and size to the input. Continuation and extension use automatically calculated temporal masks. The mask image and static preview below are historical input illustrations, not required workflow inputs.

Case	Expected edit	Duration	Audio mask
Object removal	Reconstruct surrounding wall in the bag region	5 s	0
Inpainting	Curled orange cat gently swaying its tail	5 s	0
Continuation	Overcast sky, torrential rain and distant lightning; preserve_fraction=0.2	5 s	0.8
Extension	Summer starry night; camera follows a rising firework upward until it bursts	10 s	0.9

Input	Media
Input video	school_generated3.mp4
Original spatial mask (reference only)
Static mask preview (input illustration)	WF05-mask-preview.mp4

Run each case's Mask Preview and Result outputs separately against a compatible H3 service. Inspect spatial coverage, temporal preservation, and the generated video/audio. Running the entire workflow submits all four generation branches.

vLLM Version: v0.29.0

vLLM-Omni Commit: On the EDIT-02 stack merged with main at 23f412644.

Test Result

Object removal

Tips: Include handles, shadows, and some surrounding background in the mask. In this clip, expanding each side by 32 pixels improved removal: (x, y, width, height) changed from (1106, 525, 116, 104) to (1074, 493, 180, 168). This is a clip-specific margin. For symmetric expansion by d, use (x-d, y-d, width+2d, height+2d) within the canvas bounds.

Output	Media
Mask Preview	WF05-object-removal-mask-preview.mp4
Result	WF05-object-removal.mp4

Inpainting

Output	Media
Mask Preview	WF05-inpainting-mask-preview.mp4
Result	WF05-inpainting.mp4

Continuation

Output	Media
Mask Preview	WF05-continuation-mask-preview.mp4
Result	WF05-continuation.mp4

Extension

Output	Media
Mask Preview	WF05-extension-mask-preview.mp4
Result	WF05-extension.mp4

The attachments above are manual preview and generation examples, not placeholders. Prompt adherence can vary across inputs.

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.
(anything written below this line will be removed by GitHub Actions)

[FayeSpica]

@princepride Could you review this PR when you have time? Thanks!

[vllm-omni-review-bot]

This PR appears to be related to model: MinimaxH3.

Model owners: @david6666666 @alex-jw-brooks @fhfuih

Routing: @david6666666 via model owner; @alex-jw-brooks via CODEOWNERS; @fhfuih via CODEOWNERS

@FayeSpica, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[princepride]

Great Job!!! Thank you so much, I succeed reproduce it!

[princepride]

LGTM