Fix: Proper RGBA -> RGB conversion for PIL images.

[huachenheli]

Directly converting RGBA to RGB via convert on PIL.Image produces a strange background as demonstrated in the picture:

• Original: https://upload.wikimedia.org/wikipedia/commons/4/47/PNG_transparency_demonstration_1.png
• Bad conversion:
• Desired conversion (this PR):
  

Test plan:

Unit test: pytest tests/multimodal/test_image.py -s

Local server:
vllm serve /home/huachenheli/local/llm/huggingface/llama4/Llama-4-Scout-17B-16E-Instruct --gpu-memory-utilization 0.5 --tensor-parallel-size 8 --max-model-len 65536

curl -0 -v -X POST localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d "@-" << EOF
{
    "messages": [{
        "role": "user",
        "model": "/home/huachenheli/local/llm/huggingface/llama4/Llama-4-Scout-17B-16E-Instruct",
        "content": [
            {"type": "text", "text": "What is the background color of this image, excluding four dice?"},
            {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/4/47/PNG_transparency_demonstration_1.png"}}
        ]
    }]
}
EOF

Without this change:

"content":"The background color of the image, excluding the four dice, appears to be a multicolored striped pattern. The colors include blue, white, green, purple, red, and yellow. However, if we had to choose a dominant color or a color that is most representative of the background, it would be difficult to pinpoint a single color due to the striped nature of the background.\n\nHowever, if the question is asking for a color that is visible behind the dice and not obstructed by them, then the answer could be inferred as black, since there is a black border around the image. \n\nTherefore, the background color of the image, excluding the four dice, is black."

With this change:

"content":"The background color of the image, excluding the four dice, is white."

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mgoin]

Maybe you could just make a image_to_image_mode function that has this conditional inside of it

[DarkLight1337]

Thanks for reporting and fixing this issue! Can you add a unit test to avoid future regressions?

[huachenheli]

Thanks for reporting and fixing this issue! Can you add a unit test to avoid future regressions?

Done.

[DarkLight1337]

Hmm actually, I see other places where .convert is used, can you update them as well?

[huachenheli]

Hmm actually, I see other places where .convert is used, can you update them as well?

Let me check.

[huachenheli]

Updated existing call sites.

[DarkLight1337]

Please also merge from main to fix CI failures.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ChenheliHua.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork