app.py

import time

import gradio as gr
import torch
from PIL import Image
from diffusers.image_processor import VaeImageProcessor
from diffusers.schedulers import (
    DDPMScheduler,
    DDIMScheduler,
    EulerAncestralDiscreteScheduler,
    DPMSolverMultistepScheduler,
    PNDMScheduler,
    EulerDiscreteScheduler,
)
from diffusers.utils.torch_utils import randn_tensor

MODEL_IDS = {
    'Stable Diffusion v1.4': 'CompVis/stable-diffusion-v1-4',
    'Stable Diffusion v3 medium': 'stabilityai/stable-diffusion-3-medium-diffusers',
}

SCHEDULERS = {
    "DDPMScheduler": DDPMScheduler,
    "DDIMScheduler": DDIMScheduler,
    "PNDMScheduler": PNDMScheduler,
    "EulerDiscreteScheduler": EulerDiscreteScheduler,
    "EulerAncestralDiscreteScheduler": EulerAncestralDiscreteScheduler,
    "DPMSolverMultistepScheduler": DPMSolverMultistepScheduler,
}


def inference(
        image_pil: Image.Image,
        model_name: str,
        scheduler_name: str,
        per_step_time: int = 1,
        n_total_steps: int = 1000,
):
    yield image_pil, ""

    scheduler = SCHEDULERS[scheduler_name].from_pretrained(MODEL_IDS[model_name], subfolder='scheduler')
    scheduler.set_timesteps(num_inference_steps=n_total_steps)
    timesteps = torch.flip(scheduler.timesteps, dims=[0])

    image_processor = VaeImageProcessor()
    image_tensor = image_processor.preprocess(image_pil)

    # Fix seed
    generator = torch.Generator().manual_seed(1117)
    noise = randn_tensor(image_tensor.shape, generator)

    for i, t in enumerate(timesteps):
        t = torch.tensor([t])

        noised_image_tensor = scheduler.add_noise(image_tensor, noise, timesteps=t)
        noised_image_pil = image_processor.postprocess(noised_image_tensor)[0]
        time.sleep(per_step_time)

        # language=HTML
        info_html = f"""
            <div class="info-step">
              <span class="step-number">Step {i + 1}</span> / {n_total_steps}
            </div>
        """

        yield noised_image_pil, info_html


if __name__ == '__main__':
    demo = gr.Interface(
        title='Noisescope',
        description='',
        fn=inference,
        inputs=[
            gr.Image(type='pil', label='Input Image'),
            gr.Dropdown(list(MODEL_IDS.keys()), value='Stable Diffusion v1.4', label='Model ID'),
            gr.Dropdown(list(SCHEDULERS.keys()), value='DDPMScheduler', label='Scheduler'),
            gr.Radio(choices=[0, 0.01, 0.1, 1], value=0, label='Per-Step time'),
            gr.Radio(choices=[10, 25, 50, 100, 1000], value=50, label='Total Steps'),
        ],
        outputs=[
            gr.Image(type='pil', label='Noised Image'),
            gr.HTML(label='Timestep Info'),
        ],
        # language=css
        css="""
            body { font-family: Arial, sans-serif; background-color: #f0f0f5; }
            h1 { color: #3c3c3c; }
            .gradio-container { max-width: 800px; margin: auto; padding: 20px; background: white; border-radius: 10px; box-shadow: 0px 0px 15px rgba(0, 0, 0, 0.1); }
            .info-step { padding: 10px; background: #3c3c3c; color: white; border-radius: 5px; margin-bottom: 10px; }
            .step-number { font-weight: bold; color: #FFD700; }
        """,
        cache_examples=True,
        examples=[
            [Image.open("assets/corgi.png"), 'Stable Diffusion v1.4', 'DDIMScheduler', 0, 50],
        ],
    )
    demo.launch()