Skip to content

[Bugfix][MiniMax-H3] Preserve reference image geometry - #6992

Closed
heyuanliu-intel wants to merge 1 commit into
vllm-project:mainfrom
heyuanliu-intel:pr/h3-reference-image-geometry
Closed

heyuanliu-intel wants to merge 1 commit into
vllm-project:mainfrom
heyuanliu-intel:pr/h3-reference-image-geometry

Conversation

@heyuanliu-intel

@heyuanliu-intel heyuanliu-intel commented Sep 3, 2026 •

Copy link
Copy Markdown
Contributor

Purpose

Problem: serving_video rescales every reference image to the request canvas. MiniMax-H3 expects reference images at their native geometry, so ref2va silently replaces the reference aspect ratio with the canvas one.

Fix: register an H3-specific reference_image_size_resolver through the existing model_extras/registry.py extension point that keeps the source geometry. Pipelines without a resolver keep the serving-side resize.

Test Plan

Reproduce: send a ref2va request with a 48x32 reference image and a 96x64 canvas; the engine receives a 96x64 image.

pytest -q tests/model_extras/test_model_extras.py -k minimax_h3_preserves_reference_image_size
pytest -q tests/entrypoints/openai_api/test_video_server.py -k minimax_h3_preserves_reference_geometry

vLLM Version: as pinned by requirements/. vLLM-Omni Commit: rebased onto current main.

Test Result

CPU tests pass: the engine receives the 48x32 image while sampling params still carry (96, 64). Dropping the registry entry makes the serving test observe 96x64 again.

@chatgpt-codex-connector

Copy link
Copy Markdown

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

@vllm-omni-review-bot

Copy link
Copy Markdown

This PR appears to be related to model: MinimaxH3.

Model owners: @david6666666

@heyuanliu-intel, 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

vllm-omni-review-bot commented Sep 3, 2026 •

Copy link
Copy Markdown

Omni ReviewBot triage note

Automated triage of commit 870fefba2f1a produced:

  • Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

## Purpose

**Problem:** `serving_video` rescales every reference image to the request canvas. MiniMax-H3 expects reference images at their native geometry, so ref2va silently replaces the reference aspect ratio with the canvas one.

**Fix:** register an H3-specific `reference_image_size_resolver` through the existing `model_extras/registry.py` extension point that keeps the source geometry. Pipelines without a resolver keep the serving-side resize.

## Test Plan

**Reproduce:** send a ref2va request with a 48x32 reference image and a 96x64 canvas; the engine receives a 96x64 image.

```bash
pytest -q tests/model_extras/test_model_extras.py -k minimax_h3_preserves_reference_image_size
pytest -q tests/entrypoints/openai_api/test_video_server.py -k minimax_h3_preserves_reference_geometry
```

**vLLM Version:** as pinned by `requirements/`. **vLLM-Omni Commit:** rebased onto current `main`.

## Test Result

CPU tests pass: the engine receives the 48x32 image while sampling params still carry (96, 64). Dropping the registry entry makes the serving test observe 96x64 again.

Signed-off-by: heyuanliu-intel <heyuan.liu@intel.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working diffusion codes related to diffusion models

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants