app.py

from __future__ import annotations
import math
import random
import gradio as gr
import numpy as np
import torch
from PIL import Image
from diffusers import StableDiffusionXLImg2ImgPipeline, EDMEulerScheduler, AutoencoderKL
from huggingface_hub import hf_hub_download

vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)

pipe_edit = StableDiffusionXLImg2ImgPipeline.from_single_file(
    hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors"),
    num_in_channels=8,
    is_cosxl_edit=True,
    vae=vae,
    torch_dtype=torch.float16,
)

pipe_edit.scheduler = EDMEulerScheduler(sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction")
pipe_edit.to("cuda")

refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-refiner-1.0", 
    vae=vae, 
    torch_dtype=torch.float16, 
    use_safetensors=True, 
    variant="fp16"
)
refiner.to("cuda")

def set_timesteps_patched(self, num_inference_steps: int, device=None):
    self.num_inference_steps = num_inference_steps
    ramp = np.linspace(0, 1, self.num_inference_steps)
    sigmas = torch.linspace(math.log(self.config.sigma_min), math.log(self.config.sigma_max), len(ramp)).exp().flip(0)
    sigmas = sigmas.to(dtype=torch.float32, device=device)
    self.timesteps = self.precondition_noise(sigmas)
    self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
    self._step_index = None
    self._begin_index = None
    self.sigmas = self.sigmas.to("cpu")

EDMEulerScheduler.set_timesteps = set_timesteps_patched

def king(input_image, instruction: str, negative_prompt: str = "", steps: int = 25, randomize_seed: bool = True, seed: int = 2404, guidance_scale: float = 6, progress=gr.Progress(track_tqdm=True)):
    input_image = Image.open(input_image).convert('RGB')
    if randomize_seed:
        seed = random.randint(0, 999999)
    generator = torch.manual_seed(seed)
    output_image = pipe_edit(
        instruction,
        negative_prompt=negative_prompt,
        image=input_image,
        guidance_scale=guidance_scale,
        image_guidance_scale=1.5,
        width=input_image.width,
        height=input_image.height,
        num_inference_steps=steps,
        generator=generator,
        output_type="latent",
    ).images
    refine = refiner(
        prompt=f"{instruction}, 4k, hd, high quality, masterpiece",
        negative_prompt=negative_prompt,
        guidance_scale=7.5,
        num_inference_steps=steps,
        image=output_image,
        generator=generator,
    ).images[0]  
    return seed, refine

css = '''
.gradio-container{max-width: 700px !important}
h1{text-align:center}
footer {
    visibility: hidden
}
'''

examples = [
    ["./supercar.png", "make it red"],
    ["./red_car.png", "add some snow"],
]

with gr.Blocks(css=css) as demo:
    gr.Markdown("# Image Editing\n### Note: First image generation takes time")
    with gr.Row():
        instruction = gr.Textbox(lines=1, label="Instruction", interactive=True)
        generate_button = gr.Button("Run", scale=0)
        
    with gr.Row():
        input_image = gr.Image(label="Image", type='filepath', interactive=True)

    with gr.Row():
        guidance_scale = gr.Number(value=6.0, step=0.1, label="Guidance Scale", interactive=True)
        steps = gr.Number(value=25, step=1, label="Steps", interactive=True)

    with gr.Accordion("Advanced options", open=False):
        with gr.Row():
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                value="(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, ugly, disgusting, blurry, amputation,(face asymmetry, eyes asymmetry, deformed eyes, open mouth)",
                visible=True
            )
        with gr.Row():
            randomize_seed = gr.Checkbox(label="Randomize Seed", value=True, interactive=True)
            seed = gr.Number(value=2404, step=1, label="Seed", interactive=True)

    gr.Examples(
        examples=examples,
        inputs=[input_image, instruction],
        outputs=[input_image],
        cache_examples=False,
    )

    generate_button.click(
        king,
        inputs=[input_image, instruction, negative_prompt, steps, randomize_seed, seed, guidance_scale],
        outputs=[seed, input_image],
    )

demo.queue(max_size=500).launch()