search

HaozheZhao / UltraEdit

173 stars 9 forks source link

ImagenHub Integration #4

Closed vinesmsuic closed 3 months ago

vinesmsuic commented 4 months ago

Very cool work!

We added the implementation of UltraEdit on ImagenHub (ICLR 2024). Feel free to checkout the results on our ImagenHub benchmark and verify the code implementation :)

Best, Max

chenllliang commented 4 months ago

It seems awesome!

HaozheZhao commented 4 months ago

Hi Max,

Thanks for your message and the opportunity to work on ImagenHub!

We noticed a minor issue with the model inference for the free-form image editing setting in the ImageHub. After training with the region-based image editing data, the model has become quite sensitive to the mask input. Hence, when using the model for free-form image editing, we recommend setting a higher SFG scale, such as 10.5 or above, for better results when providing the model with a blank mask.

Alternatively, you can use this model that has been trained specifically with only free-form image editing dataset which does not support the mask input. During the inference of this model, you can just set the mask_img to be None.

vinesmsuic commented 3 months ago

Now its changed to the free-form model! https://github.com/TIGER-AI-Lab/ImagenHub/pull/33

We also re-run the experiment on Visualization: https://chromaica.github.io/Museum/ImagenHub_Text-Guided_IE/