Challenges in Achieving Fine Reconstructions

[alimohammadiamirhossein]

Hi,
I was analyzing the reconstruction results using flux-schnell, but I couldn’t achieve plausible outcomes. Am I missing something, or is it inherently challenging to achieve fine reconstructions with flux-schnell?

[wangjiangshan0725]

Hi @alimohammadiamirhossein , we suggest you refer to this issue #13 to check if the command is correct.

[alimohammadiamirhossein]

Hi again,

Thank you for your response.
Yes I checked that before.

This is my command:

python edit.py --source_prompt "a photo of Spiderman"
--target_prompt "a photo of Spiderman"
--guidance 1
--source_img_dir '/examples/source/spiderman.jpg'
--num_steps 4
--inject 0
--name 'flux-schnell'
--output_dir '/examples/edit-result/spiderman/'

Am I missing something in the setup? Does your method only support flux-dev, or is it also capable of performing reconstruction with flux-schnell?

Thank you for your help!

[wangjiangshan0725]

We have not explored the performance of inversion on the flux-schnell. Maybe you can firstly test the performance of vanilla rectified flow inversion on flux-schnell and compare the performance. Thanks!

[alimohammadiamirhossein]

Thanks! It works fine with vanilla rectified flow inversion. I have one more question: when trying to change the color of an image, such as turning a white shirt into a red one, the edit doesn’t seem to work. I also couldn’t find examples of this type of transformation in your paper. Is this a limitation of flux editing, or am I using the wrong parameters? Based on the parameters in the other .sh files you provided, it seems this doesn’t work with any of those parameter sets. Once again, thank you for your answers.