| from diffusers import StableDiffusionXLPipeline, AutoencoderKL, UNet2DConditionModel, LCMScheduler, DPMSolverMultistepScheduler | |
| import torch | |
| loaded_pipe = None | |
| loaded_pipe_id = None | |
| def load_model(pipe_id, unet_model_id): | |
| global loaded_pipe, loaded_pipe_id | |
| if loaded_pipe_id != pipe_id: | |
| unet = UNet2DConditionModel.from_pretrained( | |
| unet_model_id, | |
| torch_dtype=torch.float16, | |
| variant="fp16", | |
| ) | |
| vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) | |
| loaded_pipe = StableDiffusionXLPipeline.from_pretrained( | |
| pipe_id, unet=unet, vae=vae, torch_dtype=torch.float16, variant="fp16", | |
| ).to("cuda") | |
| loaded_pipe_id = pipe_id | |
| return loaded_pipe | |
| def set_scheduler(pipe, scheduler_type): | |
| if scheduler_type == "LCM": | |
| pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config) | |
| elif scheduler_type == "DPM++ 2M Karras": | |
| pipe.scheduler = DPMSolverMultistepScheduler(use_karras_sigmas="yes") | |
| return pipe | |
| def generate_image(prompt, num_inference_steps, seed, guidance_scale, negative_prompt, pipe_id="KBlueLeaf/kohaku-xl-beta7.1", unet_model_id="latent-consistency/lcm-sdxl", scheduler_type="LCM"): | |
| global loaded_pipe | |
| pipe = load_model(pipe_id, unet_model_id) | |
| pipe = set_scheduler(pipe, scheduler_type) | |
| generator = torch.manual_seed(seed) | |
| image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, generator=generator, guidance_scale=guidance_scale).images[0] | |
| return image | |