import gradio as gr import torch import numpy as np import requests import random 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 import re def randomize_seed_fn(seed, randomize_seed): if randomize_seed: seed = random.randint(0, np.iinfo(np.int32).max) torch.manual_seed(seed) return seed 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 zs, wts def sample(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) def get_example(): case = [ [ 'examples/source_a_cat_sitting_next_to_a_mirror.jpeg', 'a cat sitting next to a mirror', 'watercolor painting of a cat sitting next to a mirror', 100, 36, 15, '+Schnauzer dog, -cat', 5.5, 1, 'examples/ddpm_watercolor_painting_a_cat_sitting_next_to_a_mirror.png', 'examples/ddpm_sega_watercolor_painting_a_cat_sitting_next_to_a_mirror_plus_dog_minus_cat.png' ], [ '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', 100, 36, 15, '+painting', 10, 1, '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' ], [ 'examples/source_wall_with_framed_photos.jpeg', '', '', 100, 36, 15, '+pink drawings of muffins', 10, 1, 'examples/ddpm_wall_with_framed_photos.png', 'examples/ddpm_sega_plus_pink_drawings_of_muffins.png' ], [ 'examples/source_an_empty_room_with_concrete_walls.jpg', 'an empty room with concrete walls', 'glass walls', 100, 36, 17, '+giant elephant', 10, 1, 'examples/ddpm_glass_walls.png', 'examples/ddpm_sega_glass_walls_gian_elephant.png' ]] return case def invert_and_reconstruct( input_image, do_inversion, seed, randomize_seed, wts, zs, src_prompt ="", tar_prompt="", steps=100, src_cfg_scale = 3.5, skip=36, tar_cfg_scale=15, ): x0 = load_512(input_image, device=device) if do_inversion or randomize_seed: # invert and retrieve noise maps and latent zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale) wts = gr.State(value=wts_tensor) zs = gr.State(value=zs_tensor) do_inversion = False # output = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale) # return output, wts, zs, do_inversion return wts, zs, do_inversion def edit(input_image, wts, zs, tar_prompt, steps, skip, tar_cfg_scale, edit_concept_1,edit_concept_2,edit_concept_3, guidnace_scale_1,guidnace_scale_2,guidnace_scale_3, warmup_1, warmup_2, warmup_3, neg_guidance_1, neg_guidance_2, neg_guidance_3, threshold_1, threshold_2, threshold_3 ): # SEGA # parse concepts and neg guidance editing_args = dict( editing_prompt = [edit_concept_1,edit_concept_2,edit_concept_3], reverse_editing_direction = [ neg_guidance_1, neg_guidance_2, neg_guidance_3,], edit_warmup_steps=[warmup_1, warmup_2, warmup_3,], edit_guidance_scale=[guidnace_scale_1,guidnace_scale_2,guidnace_scale_3], edit_threshold=[threshold_1, threshold_2, threshold_3], edit_momentum_scale=0.5, edit_mom_beta=0.6, eta=1, ) latnets = wts.value[skip].expand(1, -1, -1, -1) sega_out = sem_pipe(prompt=tar_prompt, latents=latnets, guidance_scale = tar_cfg_scale, num_images_per_prompt=1, num_inference_steps=steps, use_ddpm=True, wts=wts.value, zs=zs.value[skip:], **editing_args) return sega_out.images[0] ######## # demo # ######## intro = """

Edit Friendly DDPM X Semantic Guidance

An Edit Friendly DDPM Noise Space: Inversion and Manipulations X SEGA: Instructing Diffusion using Semantic Dimensions

For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. Duplicate Space

""" with gr.Blocks(css='style.css') as demo: def add_concept(sega_concepts_counter): if sega_concepts_counter == 1: return row2.update(visible=True), row3.update(visible=False), plus.update(visible=True), 2 else: return row2.update(visible=True), row3.update(visible=True), plus.update(visible=False), 3 def reset_do_inversion(): do_inversion = True return do_inversion gr.HTML(intro) wts = gr.State() zs = gr.State() do_inversion = gr.State(value=True) sega_concepts_counter = gr.State(1) with gr.Row(): input_image = gr.Image(label="Input Image", interactive=True) # ddpm_edited_image = gr.Image(label=f"DDPM Reconstructed Image", interactive=False, visible=False) sega_edited_image = gr.Image(label=f"DDPM + SEGA Edited Image", interactive=False) input_image.style(height=365, width=365) # ddpm_edited_image.style(height=512, width=512) sega_edited_image.style(height=365, width=365) with gr.Tabs() as tabs: with gr.TabItem('1. Describe the desired output', id=0): with gr.Row().style(mobile_collapse=False, equal_height=True): tar_prompt = gr.Textbox( label="Edit Concept", show_label=False, max_lines=1, placeholder="Enter your 1st edit prompt", ) with gr.TabItem('2. Add SEGA edit concepts', id=1): # with gr.Group(): with gr.Row().style(mobile_collapse=False, equal_height=True): # tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="") neg_guidance_1 = gr.Checkbox( label='Negative Guidance') warmup_1 = gr.Slider(label='Warmup', minimum=0, maximum=50, value=10, step=1, interactive=True) guidnace_scale_1 = gr.Slider(label='Scale', minimum=1, maximum=10, value=5, step=0.25, interactive=True) threshold_1 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99, value=0.95, steps=0.01, interactive=True) edit_concept_1 = gr.Textbox( label="Edit Concept", show_label=False, max_lines=1, placeholder="Enter your 1st edit prompt", ) with gr.Row(visible=False) as row2: # tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="") neg_guidance_2 = gr.Checkbox( label='Negative Guidance',visible=True) warmup_2 = gr.Slider(label='Warmup', minimum=0, maximum=50, value=10, step=1, visible=True,interactive=True) guidnace_scale_2 = gr.Slider(label='Scale', minimum=1, maximum=10, value=5, step=0.25,visible=True, interactive=True) threshold_2 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99, value=0.95, steps=0.01,visible=True, interactive=True) edit_concept_2 = gr.Textbox( label="Edit Concept", show_label=False,visible=True, max_lines=1, placeholder="Enter your 2st edit prompt", ) with gr.Row(visible=False) as row3: # tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True, placeholder="") neg_guidance_3 = gr.Checkbox( label='Negative Guidance',visible=True) warmup_3 = gr.Slider(label='Warmup', minimum=0, maximum=50, value=10, step=1, visible=True,interactive=True) guidnace_scale_3 = gr.Slider(label='Scale', minimum=1, maximum=10, value=5, step=0.25,visible=True, interactive=True) threshold_3 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99, value=0.95, steps=0.01,visible=True, interactive=True) edit_concept_3 = gr.Textbox( label="Edit Concept", show_label=False,visible=True, max_lines=1, placeholder="Enter your 3rd edit prompt", ) with gr.Row().style(mobile_collapse=False, equal_height=True): plus = gr.Button("+") with gr.Row(): with gr.Column(scale=1, min_width=100): run_button = gr.Button("Run") # with gr.Column(scale=1, min_width=100): # edit_button = gr.Button("Edit") with gr.Accordion("Advanced Options", open=False): with gr.Row(): with gr.Column(): src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="") steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True) src_cfg_scale = gr.Number(value=3.5, label=f"Source Guidance Scale", interactive=True) seed = gr.Number(value=0, precision=0, label="Seed", interactive=True) randomize_seed = gr.Checkbox(label='Randomize seed', value=False) with gr.Column(): skip = gr.Slider(minimum=0, maximum=40, value=36, label="Skip Steps", interactive=True) tar_cfg_scale = gr.Slider(minimum=7, maximum=18,value=15, label=f"Guidance Scale", interactive=True) # gr.Markdown(help_text) plus.click(fn = add_concept, inputs=sega_concepts_counter, outputs= [row2, row3, plus, sega_concepts_counter], queue = False) run_button.click( fn = randomize_seed_fn, inputs = [seed, randomize_seed], outputs = [seed], queue = False).then( fn=invert_and_reconstruct, inputs=[input_image, do_inversion, seed, randomize_seed, wts, zs, src_prompt, tar_prompt, steps, src_cfg_scale, skip, tar_cfg_scale, ], # outputs=[ddpm_edited_image, wts, zs, do_inversion], outputs=[wts, zs, do_inversion], ).success( fn=edit, inputs=[input_image, wts, zs, tar_prompt, steps, skip, tar_cfg_scale, edit_concept_1,edit_concept_2,edit_concept_3, guidnace_scale_1,guidnace_scale_2,guidnace_scale_3, warmup_1, warmup_2, warmup_3, neg_guidance_1, neg_guidance_2, neg_guidance_3, threshold_1, threshold_2, threshold_3 ], outputs=[sega_edited_image], ) input_image.change( fn = reset_do_inversion, outputs = [do_inversion] ) # gr.Examples( # label='Examples', # examples=get_example(), # inputs=[input_image, src_prompt, tar_prompt, steps, # # src_cfg_scale, # skip, # tar_cfg_scale, # # edit_concept, # sega_edit_guidance, # warm_up, # # neg_guidance, # ddpm_edited_image, sega_edited_image # ], # outputs=[ddpm_edited_image, sega_edited_image], # # fn=edit, # # cache_examples=True # ) demo.queue() demo.launch(share=False)