import gradio as gr import torch import requests from io import BytesIO from diffusers import StableDiffusionPipeline from diffusers import DDIMScheduler from utils import * from inversion_utils import * from modified_pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline from torch import autocast, inference_mode def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1): # inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf, # based on the code in https://github.com/inbarhub/DDPM_inversion # returns wt, zs, wts: # wt - inverted latent # wts - intermediate inverted latents # zs - noise maps sd_pipe.scheduler.set_timesteps(num_diffusion_steps) # vae encode image with autocast("cuda"), inference_mode(): w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float() # find Zs and wts - forward process wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps) return wt, zs, wts def sample(wt, zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1): # reverse process (via Zs and wT) w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:]) # vae decode image with autocast("cuda"), inference_mode(): x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample if x0_dec.dim()<4: x0_dec = x0_dec[None,:,:,:] img = image_grid(x0_dec) return img # load pipelines sd_model_id = "runwayml/stable-diffusion-v1-5" device = torch.device("cuda" if torch.cuda.is_available() else "cpu") sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id).to(device) sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler") sem_pipe = SemanticStableDiffusionPipeline.from_pretrained(sd_model_id).to(device) cache_examples = True def get_example(): case = [ [ 'examples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg', 'a man wearing a brown hoodie in a crowded street', 'a robot wearing a brown hoodie in a crowded street', 'painting', 'examples/ddpm_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png', 'examples/ddpm_sega_painting_of_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png' ]] return case def edit(input_image, src_prompt ="", tar_prompt="", steps=100, # src_cfg_scale, skip=36, tar_cfg_scale=15, edit_concept="", sega_edit_guidance=0, # warm_up=1, # neg_guidance=False, flip=False, left = 0, right = 0, top = 0, bottom = 0): # offsets=(0,0,0,0) x0 = load_512(input_image, left,right, top, bottom, device) # invert # wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale) wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps) if flip: wt, zs, wts = torch.flip(wt, [2, 3]),torch.flip(zs, [2, 3]),torch.flip(wts, [2, 3]) latnets = wts[skip].expand(1, -1, -1, -1) #pure DDPM output pure_ddpm_out = sample(wt, zs, wts, prompt_tar=tar_prompt, cfg_scale_tar=tar_cfg_scale, skip=skip) if not edit_concepts or not sega_edit_guidance: return pure_ddpm_out,pure_ddpm_out # SEGA edit_concepts = edit_concept.split(",") neg_guidance =[] for edit_concept in edit_concepts: if edit_concept.startswith("-"): neg_guidance.append(True) else: neg_guidance.append(False) edit_concepts = [concept.strip("+|-") for concept in edit_concepts] default_warm_up = [1]*len(edit_concepts) editing_args = dict( editing_prompt = edit_concepts, reverse_editing_direction = neg_guidance, edit_warmup_steps=default_warm_up, edit_guidance_scale=[sega_edit_guidance], edit_threshold=[.93], edit_momentum_scale=0.5, edit_mom_beta=0.6 ) sega_out = sem_pipe(prompt=tar_prompt,eta=1, latents=latnets, guidance_scale = tar_cfg_scale, num_images_per_prompt=1, num_inference_steps=steps, use_ddpm=True, wts=wts, zs=zs[skip:], **editing_args) yield pure_ddpm_out,sega_out.images[0] ######## # demo # ######## intro = """
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.