Qwen-Edit consistency improvement

[KZF-kzf]

I found that there is no qualitative comparison of Qwen-Edit regarding this in the paper. I want to know the improvement in human-portrait consistency after the authors trained QwenEdit using the reward from the paper. May I ask if there are any specific qualitative or quantitative experiments that can be shown?

[Jayce-Ping]

Hi there! Sorry for the late reply. PaCo-Reward is designed to focus on three aspects of visual consistency: identity, logic, and style. For image editing, identity consistency (e.g., human faces) should theoretically improve after training with PaCo-Reward.

While we did not conduct a specific automated evaluation for human-portrait consistency in the editing task, our Text-to-ImageSet results show a significant improvement, as illustrated in Figure 7 (left). We believe these gains in consistency should also carry over to the editing task.