edit_app.py

from __future__ import annotations

import math
import random
import sys
from argparse import ArgumentParser

import einops
import gradio as gr
import k_diffusion as K
import numpy as np
import torch
import torch.nn as nn
from einops import rearrange
from omegaconf import OmegaConf
from PIL import Image, ImageOps
from torch import autocast
from huggingface_hub import hf_hub_download

sys.path.append("./stable_diffusion")

from stable_diffusion.ldm.util import instantiate_from_config


help_text = """
If you're not getting what you want, there may be a few reasons:
1. Is the image not changing enough? Your Image CFG weight may be too high. This value dictates how similar the output should be to the input. It's possible your edit requires larger changes from the original image, and your Image CFG weight isn't allowing that. Alternatively, your Text CFG weight may be too low. This value dictates how much to listen to the text instruction. The default Image CFG of 1.5 and Text CFG of 7.5 are a good starting point, but aren't necessarily optimal for each edit. Try:
    * Decreasing the Image CFG weight, or
    * Incerasing the Text CFG weight, or
2. Conversely, is the image changing too much, such that the details in the original image aren't preserved? Try:
    * Increasing the Image CFG weight, or
    * Decreasing the Text CFG weight
3. Try generating results with different random seeds by setting "Randomize Seed" and running generation multiple times. You can also try setting "Randomize CFG" to sample new Text CFG and Image CFG values each time.
4. Rephrasing the instruction sometimes improves results (e.g., "turn him into a dog" vs. "make him a dog" vs. "as a dog").
5. Increasing the number of steps sometimes improves results.
6. Do faces look weird? The Stable Diffusion autoencoder has a hard time with faces that are small in the image. Try:
    * Cropping the image so the face takes up a larger portion of the frame.
"""


example_instructions = [
    "Make it a picasso painting",
    "as if it were by modigliani",
    "convert to a bronze statue",
    "Turn it into an anime.",
    "have it look like a graphic novel",
    "make him gain weight",
    "what would he look like bald?",
    "Have him smile",
    "Put him in a cocktail party.",
    "move him at the beach.",
    "add dramatic lighting",
    "Convert to black and white",
    "What if it were snowing?",
    "Give him a leather jacket",
    "Turn him into a cyborg!",
    "make him wear a beanie",
]


class CFGDenoiser(nn.Module):
    def __init__(self, model):
        super().__init__()
        self.inner_model = model

    def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
        cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
        cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
        cfg_cond = {
            "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
            "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
        }
        out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
        return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)


def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
    print(f"Loading model from {ckpt}")
    pl_sd = torch.load(ckpt, map_location="cpu")
    if "global_step" in pl_sd:
        print(f"Global Step: {pl_sd['global_step']}")
    sd = pl_sd["state_dict"]
    if vae_ckpt is not None:
        print(f"Loading VAE from {vae_ckpt}")
        vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
        sd = {
            k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
            for k, v in sd.items()
        }
    model = instantiate_from_config(config.model)
    m, u = model.load_state_dict(sd, strict=False)
    if len(m) > 0 and verbose:
        print("missing keys:")
        print(m)
    if len(u) > 0 and verbose:
        print("unexpected keys:")
        print(u)
    return model


def main():
    parser = ArgumentParser()
    parser.add_argument("--resolution", default=512, type=int)
    parser.add_argument("--config", default="configs/generate.yaml", type=str)
    parser.add_argument("--ckpt", default="instruct-pix2pix-00-22000.ckpt", type=str)
    parser.add_argument("--vae-ckpt", default=None, type=str)
    args = parser.parse_args()

    args.ckpt = hf_hub_download(repo_id="diffusers/pix2pix-sd", filename="instruct-pix2pix-00-22000.ckpt")

    config = OmegaConf.load(args.config)
    model = load_model_from_config(config, args.ckpt, args.vae_ckpt)
    model.eval().cuda()
    model_wrap = K.external.CompVisDenoiser(model)
    model_wrap_cfg = CFGDenoiser(model_wrap)
    null_token = model.get_learned_conditioning([""])
    example_image = Image.open("imgs/example.jpg").convert("RGB")

    def load_example(
        steps: int,
        randomize_seed: bool,
        seed: int,
        randomize_cfg: bool,
        text_cfg_scale: float,
        image_cfg_scale: float,
    ):
        example_instruction = random.choice(example_instructions)
        return [example_image, example_instruction] + generate(
            example_image,
            example_instruction,
            steps,
            randomize_seed,
            seed,
            randomize_cfg,
            text_cfg_scale,
            image_cfg_scale,
        )

    def generate(
        input_image: Image.Image,
        instruction: str,
        steps: int,
        randomize_seed: bool,
        seed: int,
        randomize_cfg: bool,
        text_cfg_scale: float,
        image_cfg_scale: float,
    ):
        seed = random.randint(0, 100000) if randomize_seed else seed
        text_cfg_scale = round(random.uniform(6.0, 9.0), ndigits=2) if randomize_cfg else text_cfg_scale
        image_cfg_scale = round(random.uniform(1.2, 1.8), ndigits=2) if randomize_cfg else image_cfg_scale

        width, height = input_image.size
        factor = args.resolution / max(width, height)
        factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
        width = int((width * factor) // 64) * 64
        height = int((height * factor) // 64) * 64
        input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)

        if instruction == "":
            return [input_image, seed]

        with torch.no_grad(), autocast("cuda"), model.ema_scope():
            cond = {}
            cond["c_crossattn"] = [model.get_learned_conditioning([instruction])]
            input_image = 2 * torch.tensor(np.array(input_image)).float() / 255 - 1
            input_image = rearrange(input_image, "h w c -> 1 c h w").to(model.device)
            cond["c_concat"] = [model.encode_first_stage(input_image).mode()]

            uncond = {}
            uncond["c_crossattn"] = [null_token]
            uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])]

            sigmas = model_wrap.get_sigmas(steps)

            extra_args = {
                "cond": cond,
                "uncond": uncond,
                "text_cfg_scale": text_cfg_scale,
                "image_cfg_scale": image_cfg_scale,
            }
            torch.manual_seed(seed)
            z = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
            z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args)
            x = model.decode_first_stage(z)
            x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
            x = 255.0 * rearrange(x, "1 c h w -> h w c")
            edited_image = Image.fromarray(x.type(torch.uint8).cpu().numpy())

            return [seed, text_cfg_scale, image_cfg_scale, edited_image]

    def reset():
        return [0, "Randomize Seed", 1371, "Fix CFG", 7.5, 1.5, None]

    with gr.Blocks(css="footer {visibility: hidden}") as demo:
        with gr.Row():
            with gr.Column(scale=1, min_width=100):
                generate_button = gr.Button("Generate")
            with gr.Column(scale=1, min_width=100):
                load_button = gr.Button("Load Example")
            with gr.Column(scale=1, min_width=100):
                reset_button = gr.Button("Reset")
            with gr.Column(scale=3):
                instruction = gr.Textbox(lines=1, label="Edit Instruction", interactive=True)

        with gr.Row():
            input_image = gr.Image(label="Input Image", type="pil", interactive=True)
            edited_image = gr.Image(label=f"Edited Image", type="pil", interactive=False)
            input_image.style(height=512, width=512)
            edited_image.style(height=512, width=512)

        with gr.Row():
            steps = gr.Number(value=100, precision=0, label="Steps", interactive=True)
            randomize_seed = gr.Radio(
                ["Fix Seed", "Randomize Seed"],
                value="Randomize Seed",
                type="index",
                show_label=False,
                interactive=True,
            )
            seed = gr.Number(value=1371, precision=0, label="Seed", interactive=True)
            randomize_cfg = gr.Radio(
                ["Fix CFG", "Randomize CFG"],
                value="Fix CFG",
                type="index",
                show_label=False,
                interactive=True,
            )
            text_cfg_scale = gr.Number(value=7.5, label=f"Text CFG", interactive=True)
            image_cfg_scale = gr.Number(value=1.5, label=f"Image CFG", interactive=True)

        gr.Markdown(help_text)

        load_button.click(
            fn=load_example,
            inputs=[
                steps,
                randomize_seed,
                seed,
                randomize_cfg,
                text_cfg_scale,
                image_cfg_scale,
            ],
            outputs=[input_image, instruction, seed, text_cfg_scale, image_cfg_scale, edited_image],
        )
        generate_button.click(
            fn=generate,
            inputs=[
                input_image,
                instruction,
                steps,
                randomize_seed,
                seed,
                randomize_cfg,
                text_cfg_scale,
                image_cfg_scale,
            ],
            outputs=[seed, text_cfg_scale, image_cfg_scale, edited_image],
        )
        reset_button.click(
            fn=reset,
            inputs=[],
            outputs=[steps, randomize_seed, seed, randomize_cfg, text_cfg_scale, image_cfg_scale, edited_image],
        )

    demo.queue(concurrency_count=1)
    demo.launch(share=True)


if __name__ == "__main__":
    main()