What is the reason why removing one person from an image can affect others?

[Happyflytoyou]

Thanks for your great work!
I test some image editing case, such as "remove the right person in the image", execute this command: python infer.py/infer_mobile.py --prompt "remove the right person in the image" --image_path ./input/test_dreamlite.jpg , here are the processing results of the model. You can see that although the person on the right has been removed, it has caused the other two faces to deform. Have you tested this situation? How was it resolved?
We look forward to your reply.

[minh-nguyenhoang]

This is a common problem for such small model, you would need something like masking to overcome this problem.

[Carlofkl]

Hi @Happyflytoyou , thanks for the feedback!

As a lightweight 0.39B on-device model, DreamLite does have known limitations on face/portrait editing. This is an inherent trade-off of the compact model capacity. We are actively working on improving these aspects in future versions.

If you have specific use cases in mind, feel free to share and we can suggest whether DreamLite is a good fit.

Thanks for trying it out!

[Happyflytoyou]

@minh-nguyenhoang @Carlofkl Thanks for your reply！

I understand what you mean, but I think this model should support both comprehension and generative tasks to some extent. In theory, it should be able to process only specific areas of comprehension? I will also try using the optimization strategy of segmentation mask，and look forward to your future optimized versions!