Black picture results
Anybody had a problem with getting solid black pictures as results after training the model? I managed to train with 1 steps and 100 steps fine. Then when I went to 1000 steps I got black pictures after one training but the next was fine. I tried training further - for 2000 and 2500 steps but both cases result in black pictures. Any idea what may be causing the issue?
Hi,
@FidgetSpinoza
FYI, the training functionality of this Space is basically a GUI wrapper of the script train_dreambooth_lora.py in this directory of the diffusers library. So the forum for diffusers might also be a good place to ask this question. I think you will find more experts there.
@hysts
Thanks, I'll ask there. Although I have an idea what might be going on. I think that for some reason safety checker is flagging each generated picture. I'll try fiddling with it, disable the NSFW checker and see if that does any good.
@FidgetSpinoza did you get it to work?
@FidgetSpinoza
@Cossette
I am facing black pictures too when using 2000 or 3000 steps.
Have you guys figured out why?
How I can find my trained Lora? I saved as Private but cannot find, thnx
I personally found that it is because the learning_rate setting is too high, try 4e-5 to 1e-5.
But would like someone to explain why too high results in black
I personally found that it is because the learning_rate setting is too high, try 4e-5 to 1e-5.
But would like someone to explain why too high results in black
I tried it, but it stills generate only black images.
I personally found that it is because the learning_rate setting is too high, try 4e-5 to 1e-5.
But would like someone to explain why too high results in blackI tried it, but it stills generate only black images.
have you try "safety_checker=None" and "requires_safety_checker=false:" ?like this:
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base",
torch_dtype=torch.float16,
safety_checker=None,
requires_safety_checker=false
)
I personally found that it is because the learning_rate setting is too high, try 4e-5 to 1e-5.
But would like someone to explain why too high results in blackI tried it, but it stills generate only black images.
have you try "safety_checker=None" and "requires_safety_checker=false:" ?like this:
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base",
torch_dtype=torch.float16,
safety_checker=None,
requires_safety_checker=false
)
Yes, its not a safety_checker issue, by generation pipeline has it disabled by default
to fix black imgs- uncheck adam8w