>>38600455...tried doing my own take on "the schizo prompt" - I kept the same prompt as before, removed the input/inpaint image, changed the seed, but still did 50 steps/12 scale on euler_a
this was the negative prompt, kek: fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws, fused digits, extra digits, missing digits, claws, paws,
LOOK MA, NO HANDS!
>>38600606I fuck with it sometimes. On a 12C/24T Ivy Bridge-E system, a single 512x768 image at 20 steps on DPM++ 2M takes ~10 minutes. I wouldn't advise using it. It's not non-deterministic, but the output is different from using pytorch with CUDA or ROCm for some reason. The "Hello Asuka" troubleshooting guide has a specific "you're using a CPU" example in it, and I got that same image.
