search

ali-vilab / Ranni

https://ranni-t2i.github.io/Ranni/
Apache License 2.0
207 stars 15 forks source link

Why are the results I tried so different from the results in the paper? #7

Open Qsuperme-Q opened 5 months ago

Qsuperme-Q commented 5 months ago

20240409-203805

thss15fyt commented 5 months ago

Thanks for your interests. The current released version is based on the pure SDv2.1 and it not quite stable especially for local attribute binding. We improve it by ignoring the global token for better local control, but it is still worse than the paper version. The paper version is a larger private 3B diffusion model with better image quality and local sensibility (with different text conditioning). We are still working to develop and release better version of the panel-to-image model.

---- Replied Message ---- | From | @.> | | Date | 04/09/2024 20:39 | | To | ali-vilab/Ranni @.> | | Cc | Subscribed @.***> | | Subject | [ali-vilab/Ranni] Why are the results I tried so different from the results in the paper? (Issue #7) |

20240409-203805.jpeg (view on web)

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

thss15fyt commented 5 months ago

For the current version, try with smaller control scale and larger control stop for better image quality.

image
Qsuperme-Q commented 5 months ago

Thank you for your reply. Will this updated version be available for everyone to test?

thss15fyt commented 5 months ago

Thank you for your reply. Will this updated version be available for everyone to test?

Yes. We will keep working to release it publicly.