Artifacts (possibly related to MPS)

[jwooldridge234]

Hey! Your model looks really cool- just wondering if you can point me in the right direction as to how to solve this issue. I'm using this code:

import torch
from diffusers import AutoPipelineForText2Image
from diffusers.pipelines.wuerstchen import DEFAULT_STAGE_C_TIMESTEPS
from pathlib import Path
import time

DIR_NAME="./images/"
dirpath = Path(DIR_NAME)
# create parent dir if doesn't exist
dirpath.mkdir(parents=True, exist_ok=True)

pipe = AutoPipelineForText2Image.from_pretrained("warp-ai/wuerstchen", torch_dtype=torch.float16).to("mps")

caption = "A grim woman wearing rusty atompunk power-armor, holding a massive gauss rifle, standing on a cliff overlooking a vast desert, 70mm film still"
negative = "3d, cartoon, doll, lowres"
images = pipe(
    prompt=caption, 
    negative_prompt=negative,
    width=1280,
    height=1024,
    prior_timesteps=DEFAULT_STAGE_C_TIMESTEPS,
    prior_guidance_scale=4.0,
    num_images_per_prompt=1,
).images

for idx, image in enumerate(images):
    image_name = f'{time.time()}.png'
    image_path = dirpath / image_name
    image.save(image_path)

This is the console output I get:

Loading pipeline components...: 100%|█████████████| 5/5 [00:07<00:00,  1.45s/it]
Loading pipeline components...: 100%|█████████████| 4/4 [00:11<00:00,  2.75s/it]
100%|███████████████████████████████████████████| 29/29 [00:40<00:00,  1.41s/it]
  0%|                                                    | 0/12 [00:00<?, ?it/s]/Users/jackwooldridge/StableDiffusion/diffusers/venv/lib/python3.9/site-packages/torch/nn/functional.py:4027: UserWarning: The operator 'aten::_upsample_bicubic2d_aa.out' is not currently supported on the MPS backend and will fall back to run on the CPU. This may have performance implications. (Triggered internally at /Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/mps/MPSFallback.mm:13.)
  return torch._C._nn._upsample_bicubic2d_aa(input, output_size, align_corners, scale_factors)
100%|███████████████████████████████████████████| 12/12 [00:30<00:00,  2.54s/it]

And here's the image that gets output at the end:

I've tried with no negative prompts and different output sizes. The output always seems to have this distortion.