|
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): |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
sd_pipe.scheduler.set_timesteps(num_diffusion_steps) |
|
|
|
|
|
with autocast("cuda"), inference_mode(): |
|
w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float() |
|
|
|
|
|
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): |
|
|
|
|
|
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:]) |
|
|
|
|
|
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 |
|
|
|
|
|
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) |
|
latents, wts, zs = None, None, None |
|
|
|
def invert_and_reconstruct(input_image, |
|
src_prompt, |
|
tar_prompt, |
|
steps, |
|
|
|
skip, |
|
tar_cfg_scale): |
|
offsets=(0,0,0,0) |
|
x0 = load_512(input_image, *offsets, device) |
|
|
|
|
|
|
|
|
|
wt, zs, wts = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps) |
|
latnets = wts[skip].expand(1, -1, -1, -1) |
|
|
|
eta = 1 |
|
|
|
pure_ddpm_out = sample(wt, zs, wts, prompt_tar=tar_prompt, |
|
cfg_scale_tar=tar_cfg_scale, skip=skip, |
|
eta = eta) |
|
return pure_ddpm_out |
|
|
|
def edit( tar_prompt, |
|
steps, |
|
edit_concept, |
|
sega_edit_guidance, |
|
warm_up, |
|
neg_guidance): |
|
|
|
editing_args = dict( |
|
editing_prompt = [edit_concept], |
|
reverse_editing_direction = [neg_guidance], |
|
edit_warmup_steps=[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=eta, latents=latnets, |
|
num_images_per_prompt=1, |
|
num_inference_steps=steps, |
|
use_ddpm=True, wts=wts, zs=zs[skip:], **editing_args) |
|
return sega_out.images[0] |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
intro = """<h1 style="font-weight: 900; margin-bottom: 7px;"> |
|
Edit Friendly DDPM X Semantic Guidance: Editing Real Images |
|
</h1> |
|
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. |
|
<br/> |
|
<a href="https://huggingface.co/spaces/LinoyTsaban/ddpm_sega?duplicate=true"> |
|
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> |
|
<p/>""" |
|
with gr.Blocks() as demo: |
|
gr.HTML(intro) |
|
|
|
|
|
with gr.Row(): |
|
input_image = gr.Image(label="Input Image",interactive=False) |
|
ddpm_edited_image = gr.Image(label=f"Reconstructed Image", interactive=False) |
|
sega_edited_image = gr.Image(label=f"Edited Image", interactive=False) |
|
input_image.style(height=512, width=512) |
|
ddpm_edited_image.style(height=512, width=512) |
|
sega_edited_image.style(height=512, width=512) |
|
|
|
with gr.Row(): |
|
with gr.Column(scale=1, min_width=100): |
|
generate_button = gr.Button("Invert") |
|
with gr.Column(scale=1, min_width=100): |
|
edit_button = gr.Button("Edit") |
|
|
|
|
|
|
|
|
|
|
|
with gr.Row(): |
|
src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True) |
|
|
|
tar_prompt = gr.Textbox(lines=1, label="Target Prompt", interactive=True) |
|
|
|
with gr.Row(): |
|
|
|
steps = gr.Number(value=100, precision=0, label="Steps", interactive=True) |
|
|
|
|
|
skip = gr.Number(value=36, precision=0, label="Skip", interactive=True) |
|
tar_cfg_scale = gr.Number(value=15, label=f"Reconstruction CFG", interactive=True) |
|
|
|
edit_concept = gr.Textbox(lines=1, label="Edit Concept", interactive=True) |
|
sega_edit_guidance = gr.Number(value=5, label=f"SEGA CFG", interactive=True) |
|
warm_up = gr.Number(value=5, label=f"Warm-up Steps", interactive=True) |
|
neg_guidance = gr.Checkbox(label="SEGA negative_guidance") |
|
|
|
|
|
|
|
|
|
generate_button.click( |
|
fn=invert_and_reconstruct, |
|
inputs=[input_image, |
|
src_prompt, |
|
tar_prompt, |
|
steps, |
|
src_cfg_scale, |
|
skip, |
|
tar_cfg_scale |
|
], |
|
outputs=[ddpm_edited_image], |
|
) |
|
|
|
edit_button.click( |
|
fn=edit, |
|
inputs=[tar_prompt, |
|
steps, |
|
edit_concept, |
|
sega_edit_guidance, |
|
warm_up, |
|
neg_guidance |
|
], |
|
outputs=[sega_edited_image], |
|
) |
|
|
|
|
|
demo.queue(concurrency_count=1) |
|
demo.launch(share=False) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|