How to evaluate object identity and mask precision

[fermigo66-collab]

Hi Authors, thanks for the great work, I got a question about Figure 16, how do you evaluate the identity and mask precision? I could not find their evaluation method in the paper. Thanks.

[nini0919]

Hi, thanks for your question.

Actually, the main purpose of Figure 16 is to illustrate whether our CoT-based data construction strategy helps improve performance on editing tasks. For this figure, we use samples from the ImgEdit benchmark as a qualitative visualization to verify the effectiveness of our CoT data construction.

Regarding the object identity and mask precision metrics, they are not newly introduced or re-evaluated by us in this figure. Instead, these results come directly from the GPT-4.1 evaluation results provided in the ImgEdit benchmark.

Hope this helps clarify it.

[fermigo66-collab]

Hi, thanks for your question.

Actually, the main purpose of Figure 16 is to illustrate whether our CoT-based data construction strategy helps improve performance on editing tasks. For this figure, we use samples from the ImgEdit benchmark as a qualitative visualization to verify the effectiveness of our CoT data construction.

Regarding the object identity and mask precision metrics, they are not newly introduced or re-evaluated by us in this figure. Instead, these results come directly from the GPT-4.1 evaluation results provided in the ImgEdit benchmark.

Hope this helps clarify it.

Thanks a lot for the clarification, appreciated.