I’m using this model on my m2. I tried multiple prompts but only net images are generated (attached below). Do you know why? Thanks
- detailed prompt below:
args = get_args_parser().parse_args([ "--prompt", "a photo of an astronaut riding a horse on mars", "-i", "coreml-stable-diffusion-2-base/original/packages", "-o", "output", "--compute-unit", "ALL", "--model-version", "stabilityai/stable-diffusion-2-base", "--seed", "93"]) pipeline.main(args)
Hi @skcoirz , thanks for the report! Could you please report your macOS version as well?
Furthermore, do you mind testing the following:
- Setting up an environment with python-3.8 as recommended by 
- Exporting the models locally per 
If the locally exported models succeed, the ones published on the hub might need to be updated.
Yeah sure, Atila.
- 13.3.1 (22E261)
- Mac Mini M2 Pro
Let me try the following. Will get back and let you know how it works. Thanks!
- Switch to Python 3.8
- Convert the Model Locally.
python -m python_coreml_stable_diffusion.torch2coreml --convert-unet --convert-text-encoder --convert-vae-decoder --convert-safety-checker -o models --model-version stabilityai/stable-diffusion-2-1 --attention-implementation ORIGINAL
python -m python_coreml_stable_diffusion.pipeline --prompt "a photo of an astronaut riding a horse on mars" -i models/stabilityai_stable-diffusion-2-1 --model-version stabilityai/stable-diffusion-2-1 -o output --compute-unit ALL --seed 93
Thank you @skcoirz ! Did you mean to use
--model-version stabilityai/stable-diffusion-2-1 and not
--model-version stabilityai/stable-diffusion-2-1-base? The base model generates a 512x512 output whereas the non-base model generates 768x768 and the results and testing we published so far were using the former configuration. Nonetheless, I will report this issue for the non-base version but it would be great if you could confirm that the base version still works on your end.
Interesting! base got the same mosaic output with
—compute-unit ALL, but it’s able to generate a good image with
CPU_AND_GPU. Looks like a bug somewhere in ALL branch?
I also re-tried 2.1 non-base with
CPU_AND_GPU, the generated image is pure black. Any theory? Thanks so much!