Confusion about the quality of the generated images

showlab / DatasetDM

[NeurIPS2023] DatasetDM:Synthesizing Data with Perception Annotations Using Diffusion Models

https://weijiawu.github.io/DatasetDM_page/

299 stars 12 forks source link

Confusion about the quality of the generated images #1

Closed Yuqifan1117 closed 1 year ago

Yuqifan1117 commented 1 year ago

Thank you for providing such excellent work! I have a question I would like to consult with you.

I noticed that if I input the GPT4 generated prompt directly into the Stable Diffusion v1.5, the resulting images may not be of good quality (e.g. missing objects, object relation errors). Have you done any post-processing on the generated images? (Not mentioned in the paper architecture diagram Fig.2)

I'm looking forward to your answers to these questions. Thank you.

weijiawu commented 1 year ago

Thank you for your attention. We don't have extensive post-processing. In fact, the quality of our synthetic data isn't that promising either; it only brings significant improvements in certain few-shot settings.