SDXL Training

[Benji4564]

Hello (again ;) ),
your pretrained checkpoints as well a multiple slider work very well, but I have proplems training my own. I used your sample prompts and settings for textsliders, but there where no change in the result. All generated images look the same.

[chabrarahul]

Hello,

In the paper figure 14, it is tested that strengths of edits in the range (-3 to 3) for the slider values (CLIP score and LPIPS). I wanted to know why did you choose max till 3. Does it depend on the guidance ? In the prompt-age-slider-GPT.yaml default guidance is 4. Does the sliders strength will be maximum at 4 ? The thing is when I tested slider for various values it was not showing the desired changed at slider value 4 so I have to increase and then finally at 7 or 8 it was showing me changes. Is that possible even if the guidance is set to 4 it still can show beyond 4 ?

[Benji4564]

I think I more or less fixed the problem, but found another. The SDXL training works, but you cannot add other loras with the pipe.load_lora_weights function, because thenn you only get the same image

[chabrarahul]

Hey @Benji4564 , How did you solve the training issue ? I am also the same issue where there is not effect after training.

[liujingwen-bmil]

I am also the same issue where there is not effect after training, Is there any good solution?

[rohitgandikota]

Hey guys, do you still have this issue? Happy to help if the issue still persists

[Ranking666]

I am also the same issue where there is not effect after training

[rohitgandikota]

We have released a latest update with diffusers support. This should better help understand how to get the desired effect with high slider scales.

Some suggestions:

1. instead of using single prompt - use multiple prompts (see trainscripts/textsliders/data/prompts.yaml)
2. It is better to have attributes to disentangle against. For example, use the attributes male, female, white, black, hispanic, asian when training person related sliders. So that they are disentangled
3. Increase the slider scale during inference - this usually works!