Inpaint is slow because of preparation time and model switching

[CyrusVorwald]

Inpaint switches and loads models at initialization and again halfway through the steps. It takes about 4x as long as regular image generation between that and the preparation time. I tried setting inpaint_disable_initial_latent to true but it didn't impact timing very much.

[konieshadow]

To reduce cold start time, it only download base model files of the preset when startup. And when using adavance functions such as upscale, inpaint or outpaint, downloading on demand will happend. The interface provided by replicate cannot achieve very flexible use purposes.

[CyrusVorwald]

This occurs every time inpaint is run, even if done repeatedly within a short time period.

[konieshadow]

This occurs every time inpaint is run, even if done repeatedly within a short time period.

I don't have much time to run much tests now. Can you compare it to local deployment (python main.py) and origin Fooocus?

[jareddr]

Sorry to ask a semi-unrelated question in this thread but how does one get inpainting to work?

What format should the image mask be? I've tried both transparent + white pixels for my mask and transparent+black pixels for my mask. Every time I try I just get a new image generation back taking up the whole canvas.

Here is how I have it set up on replicate inpaint image

inpaint mask (I've tried both black and white masks)

additional in paint prompt

Then this is the output I get

[jareddr]

Sorry I answered my own question by looking at the code. Mask is working great with black = no change and white = in paint.