Some questions about image editing

[huangjie-nlp]

Dear SDXL team,
We consistently observe two issues when replacing backgrounds:

Generated scenes deviate significantly from text prompts (e.g., requesting 'Chanel store entrance' outputs generic street views)

Critical details like architectural elements/textures appear blurred even at 4K resolution.
Could you advise optimal workflows to resolve this? We've attached test cases for reference.
Sincerely,

Below are some of our examples：

Original image：

prompt：
Replace the background with a scene in front of a Chanel store, 4K resolution, highly detailed

result：

We have tried numerous examples, but these issues persist in all results.

We would appreciate it if the author could provide insights into the persistent issues we're facing with background replacement in SDXL. Thank you for your support.
@timudk @akx @voletiv