[Frontend] Add MiniMax H3 video upscale workflow (WF-07) by xuexueligao · Pull Request #7472 · vllm-project/vllm-omni

xuexueligao · 2026-09-13T09:01:51Z

Purpose

Adds WF-07 from #7380: generate H3 video and audio through vLLM-Omni, upscale the frames with SeedVR2 3B FP16, and save both videos with the generated audio and FPS.

The PR adds UI/API workflow templates and setup notes using existing nodes. No node implementation or test files are added.

Test Plan

Install the SeedVR2 ComfyUI plugin and its DiT/VAE weights, and configure the H3 service as described in the workflow setup.
Import the UI workflow, set the service URL/model and server-side LoRA path, and run it. An API workflow is also included.
Compare the two saved MP4s and check their frame count, FPS, duration and audio.

Test Result

Executed the complete workflow against live H3 and ComfyUI services on 2026-09-15. The H3 generation node used a fresh seed. Download the exact workflow used for this run.

	Original	SeedVR2 output
Resolution	1344 x 768	2688 x 1536
Frames / FPS	124 / 24	124 / 24
Video duration	5.166667 s	5.166667 s
Audio	AAC, 32 kHz stereo	AAC, 32 kHz stereo

Original video · Upscaled video with audio · Side-by-side comparison

The comparison shows a full-frame overview followed by matching central crops. The original crop is enlarged 2x with bicubic interpolation for display; the SeedVR2 crop uses native pixels.

Measured audio lag before/after upscale: 0 ms. Media measurements and run records. These measurements check timing and audio preservation.

Formatting and Markdown checks passed.

AI assistance: Codex assisted with the workflow, documentation and validation.

This is my first contribution to vLLM-Omni. Please let me know if any changes are needed, and I'll update the PR.

chatgpt-codex-connector · 2026-09-13T09:02:03Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-13T09:06:14.714275Z	`bb9ff78`	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

vllm-omni-review-bot · 2026-09-13T09:06:15Z

This PR appears to be related to model: MinimaxH3.

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

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

@xuexueligao, 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.

vllm-omni-review-bot · 2026-09-13T09:09:51Z

Omni ReviewBot triage note

Resolved as of aa929967cc86: the high-priority or low-quality signal noted on an earlier commit no longer applies.

xuexueligao · 2026-09-14T02:03:33Z

Self-review updated for cacafda. Codex performed the following checks during PR preparation.

Reviewed the four-file PR diff: README, concise setup notes and UI/API workflows. The previously added E2E case and validation script are removed; existing project tests are unchanged.
Loaded the SeedVR2 workflow in ComfyUI and ran H3 generation, frame upscaling and video saving against the real services. Both saved videos have 124 frames at 24 FPS, with AAC 32 kHz stereo audio. Resolution changes from 1344 x 768 to 2688 x 1536; measured audio lag is 0 ms.
Checked workflow connections, file endings, line endings, whitespace, spelling and Markdown. The description links the new SeedVR2 workflow, screenshots and comparison videos. Run records.

princepride · 2026-09-14T03:23:40Z

@xuexueligao can you also share the result after upscale the original video?

xuexueligao · 2026-09-14T03:40:44Z

@xuexueligao can you also share the result after upscale the original video?

ok

princepride · 2026-09-14T03:41:34Z

pls refer to this pr: #7456

xuexueligao · 2026-09-14T04:03:46Z

Added the upscaled video with audio (2688 x 1536) and the original video (1344 x 768) to the PR description for comparison, along with the ComfyUI workflow screenshot and workflow JSON.

Also synced with main after #7456 merged: the workflow now uses its duration input and existing audio/aspect-ratio handling. The duplicate production changes are removed. The test changes contain one live E2E case only, and its rerun passed (1 passed, 14 warnings). The description and self-review link the final run's configuration and raw results.

@amchen2310

Add remote H3 generation with frame upscaling and preservation of the returned audio and FPS. Include UI/API workflows, aspect-ratio forwarding, validation commands and CPU regression coverage. The bytes_to_video audio fix is taken from YichenLiu (@amchen2310), PR vllm-project#6782, commit f1ab717. Its original implementation is retained with the project-required SPDX header. Refs: vllm-project#7380 Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

Replace the four CPU regression cases with one E2E case that submits the workflow to running ComfyUI and H3 services and verifies both MP4 outputs. Record the UI workflow and seed with the request, reject cached generation, and document service prerequisites. Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

princepride

I think we can remove all test related file if we haven't add any new node. And can we use SeedVR2 as the upscale model？

princepride · 2026-09-15T06:27:02Z

+## Recorded runtime configuration
+
+The H3 service used vLLM-Omni commit `33fff2155e0bf9bc8688f7ffb92ef508efdf6954` plus the original worktree changes. The ComfyUI extension and E2E runner use this PR on base `58adeec05f151e18542323cb4644b009b83cfef3`, including the interface changes from #7456. The ComfyUI host used commit `cbbc9dab1f03d0d9a6caa8a8be7d77a7e37e1e44`. The server environment contained vLLM `0.29.0+cu129`, PyTorch `2.13.0+cu129`, and NVIDIA driver `570.190`. H3 used two RTX 4090 GPUs with TP2, distributed layerwise offload, eight resident layers and cuDNN attention.


pls remove them, we don't need.

Removed the recorded runtime configuration and the E2E instructions in cacafda. The document now contains only requirements and workflow usage.

Replace the frame upscaler with existing SeedVR2 nodes, trim workflow setup notes, and remove the workflow-only validation files. Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

xuexueligao · 2026-09-15T09:13:28Z

Updated in cacafda: removed the test-related files, shortened the setup notes, and replaced RealESRGAN with SeedVR2 3B FP16 using the existing ComfyUI plugin.

The PR description now includes the updated workflow, ComfyUI screenshots, original video, SeedVR2 output with audio, and side-by-side comparison. The new run preserves all 124 frames, 24 FPS and the generated audio.

…Video Upscale.api.json Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com>

princepride

LGTM

…t#7472) Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Signed-off-by: Matthieu Laneuville <matthieu.laneuville@surf.nl>

…t#7472) Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com>

xuexueligao requested review from NickCao, alex-jw-brooks and yenuo26 as code owners September 13, 2026 09:01

hsliuustc0106 added frontend code related to entrypoint diffusion codes related to diffusion models enhancement New feature or request labels Sep 14, 2026

xuexueligao added 3 commits September 14, 2026 12:06

[Frontend] Align WF-07 with the merged duration interface

684a4e0

Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

xuexueligao force-pushed the codex/wf07-h3-upscale-pr-20260913 branch from 9724dba to 684a4e0 Compare September 14, 2026 04:12

princepride requested changes Sep 15, 2026

View reviewed changes

[Frontend] Use SeedVR2 for WF-07 video upscale

cacafda

Replace the frame upscaler with existing SeedVR2 nodes, trim workflow setup notes, and remove the workflow-only validation files. Signed-off-by: xuezhihao <89690814+xuexueligao@users.noreply.github.com>

Delete apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni MiniMax H3 …

aa92996

…Video Upscale.api.json Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com>

princepride approved these changes Sep 15, 2026

View reviewed changes

princepride added the ready label to trigger buildkite CI label Sep 15, 2026

princepride enabled auto-merge (squash) September 15, 2026 09:24

princepride merged commit 4a8ac29 into vllm-project:main Sep 15, 2026
6 of 7 checks passed

princepride mentioned this pull request Sep 16, 2026

RFC: MiniMax H3 API, ComfyUI nodes, and workflow parity #7380

Open

10 of 17 tasks

		## Recorded runtime configuration

		The H3 service used vLLM-Omni commit `33fff2155e0bf9bc8688f7ffb92ef508efdf6954` plus the original worktree changes. The ComfyUI extension and E2E runner use this PR on base `58adeec05f151e18542323cb4644b009b83cfef3`, including the interface changes from #7456. The ComfyUI host used commit `cbbc9dab1f03d0d9a6caa8a8be7d77a7e37e1e44`. The server environment contained vLLM `0.29.0+cu129`, PyTorch `2.13.0+cu129`, and NVIDIA driver `570.190`. H3 used two RTX 4090 GPUs with TP2, distributed layerwise offload, eight resident layers and cuDNN attention.

Conversation

xuexueligao commented Sep 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Test Plan

Test Result

Uh oh!

chatgpt-codex-connector Bot commented Sep 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

vllm-omni-review-bot commented Sep 13, 2026

Uh oh!

vllm-omni-review-bot commented Sep 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Omni ReviewBot triage note

Uh oh!

xuexueligao commented Sep 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

princepride commented Sep 14, 2026

Uh oh!

xuexueligao commented Sep 14, 2026

Uh oh!

princepride commented Sep 14, 2026

Uh oh!

xuexueligao commented Sep 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

princepride left a comment

Choose a reason for hiding this comment

Uh oh!

princepride Sep 15, 2026

Choose a reason for hiding this comment

Uh oh!

xuexueligao Sep 15, 2026

Choose a reason for hiding this comment

Uh oh!

xuexueligao commented Sep 15, 2026

Uh oh!

princepride left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

xuexueligao commented Sep 13, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Sep 13, 2026 •

edited

Loading

vllm-omni-review-bot commented Sep 13, 2026 •

edited

Loading

xuexueligao commented Sep 14, 2026 •

edited

Loading

xuexueligao commented Sep 14, 2026 •

edited

Loading