Spaces:
Running
on
Zero
Thoughts on this space
Add a real-time preview for depth and Canny masks, updating as you change their values. This would allow users to set them more precisely, as there are no universal values that will look good on any given input, requiring to generate a couple of images in order to find a perfect setting (and therefore spending a ZeroGPU quota). If you are determined to generate a very good image, it will take many attempts. That's frustrating.
It also would be more convenient if it was possible to configure thresholds for Canny and depth. Sometimes Canny detects too many edges, resulting in very awkward images (especially when depth estimation fails to tell any difference, confusing the model). Sometimes Canny doesn't detect edges properly, and parts of the image will be replaced completely.
Aaaand, there was a very good good space on HF that I liked a lot, called "Text guided FLUX inpainting". It had a textbox for the model that handled generating a mask based on a given prompt, and a textbox for the FLUX itself. Sadly, this space was removed shortly after, but it's concept was one of the best that I've ever seen. Instead of manually drawing an inpainting mask through a crappy Gradio drawing tool (fun fact: on Android, after touching this piece of shit at least once, the whole space would become unusable, because all buttons including "generate" will not be clickable) a specialized model will draw a mask for you. YES, I'm hinting you to add this magic wand to this space, allowing to make the model focus on certain objects, foreground, background, whatever, instead of re-generating the whole image. And, of course, there would be a third slider to set how much this mask will affect the image. Just imagine, you can edit a background on the selfie with your bro, without editing you and your bro!
Aaaaaaaaaaand other things to add:
Generating images without providing a style image, based only on a text prompt
Consistent slider values (0-50 for first and 0-1 for second? 1.5 steps?)