fp16 does not work with euler_a scheduler

[KingStorm]

Describe the bug

fp16 does not work with euler_a scheduler: it displays nearly black image.

using checkpoint: stable-diffusion-v1.5

Reproduction

• Input prompt: a boy.
• euler_a scheduler: default parameters.
• load stable-diffusion-v1.5 with fp16

Logs

No response

System Info

Ubuntu 18, 2080Ti

[patrickvonplaten]

Hey @KingStorm,

I think it's the safety checker disabling your image because it detected nsfw, could you maybe copy-paste a reproducible code snippet here?

[jwkirchenbauer]

Possibly related, tried to run the MWE from the model card and am getting fully black images too.

import torch
from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler

model_id = "stabilityai/stable-diffusion-2"

# Use the Euler scheduler here instead
scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, revision="fp16", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
# pipe.enable_attention_slicing() # seems to not affect this behavior either way

prompt = "a photo of an astronaut riding a horse on mars"

# orig example size, works fine
# image = pipe(prompt, height=768, width=768).images[0]

# widescreen version was what I wanted, and this works fine too
# image = pipe(prompt, height=512, width=1080).images[0]

# but at the "1080p" res, a black image every time (I think)
image = pipe(prompt, height=1080, width=1920).images[0]
    
image.save("astronaut_rides_horse.png")

[patrickvonplaten]

@jwkirchenbauer,

Thanks a lot for the reproducible code snippet. Since it's a different issue I opened it for #1552 .

This PR: https://github.com/huggingface/diffusers/actions/runs/3621830689/jobs/6105766777 should solve the problem regarding the memory. You were 100% right enable_attention_slicing() didn't work for SD 2.

[jwkirchenbauer]

Sorry, will do next time!
Wasn't sure whether it was the fp16 or the euler sched or whatnot, but I found this issue in my initial search.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.