stable_diffusion_demo.py

import spaces
from PIL import Image
from tqdm.auto import tqdm
import torch
from transformers import CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, UNet2DConditionModel, PNDMScheduler, LMSDiscreteScheduler, StableDiffusionPipeline, UniPCMultistepScheduler
from torchvision import transforms

torch_device = "cuda" if torch.cuda.is_available() else ("mps" if torch.mps.is_available() else "cpu")

torch_dtype = torch.float16 if torch_device in ["cuda", "mps"] else torch.float32

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5", 
    torch_dtype=torch_dtype, 
    use_safetensors=True, 
    safety_checker = None).to(torch_device)

pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)

# pipe.enable_model_cpu_offload() <--- disable for ZeroGPU

@spaces.GPU
def StableDiffusion(uncond_embeddings, text_embeddings, height, width, num_inference_steps, guidance_scale, seed):
    batch_size=1

    generator = None

    if seed:
        generator=torch.manual_seed(seed)
    
    output = pipe(
        prompt = text_embeddings,
        negative_prompt = uncond_embeddings,
        height = height,
        width = width,
        num_inference_steps = num_inference_steps,
        guidance_scale = guidance_scale,
        generator = generator
    ).images[0]
    return output