app.py

import torch
import gradio as gr
from PIL import Image
import random

from diffusers import (
    DiffusionPipeline,
    AutoencoderKL,
    StableDiffusionControlNetPipeline,
    ControlNetModel,
    StableDiffusionControlNetImg2ImgPipeline,
    DPMSolverMultistepScheduler,
    EulerDiscreteScheduler
)
import tempfile
import time
from share_btn import community_icon_html, loading_icon_html, share_js
import user_history
from illusion_style import css
import os

BASE_MODEL = "SG161222/Realistic_Vision_V5.1_noVAE"

# Initialize both pipelines on CPU with float32
# Using float32 because CPU doesn't support half-precision (float16) well
device = "cpu"
torch_dtype = torch.float32

vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch_dtype)
controlnet = ControlNetModel.from_pretrained("monster-labs/control_v1p_sd15_qrcode_monster", torch_dtype=torch_dtype)

# Safety checker disabled by default to save memory/CPU cycles
SAFETY_CHECKER_ENABLED = False 

main_pipe = StableDiffusionControlNetPipeline.from_pretrained(
    BASE_MODEL,
    controlnet=controlnet,
    vae=vae,
    safety_checker=None,
    feature_extractor=None,
    torch_dtype=torch_dtype,
).to(device)

image_pipe = StableDiffusionControlNetImg2ImgPipeline(**main_pipe.components)

# Sampler map
SAMPLER_MAP = {
    "DPM++ Karras SDE": lambda config: DPMSolverMultistepScheduler.from_config(config, use_karras=True, algorithm_type="sde-dpmsolver++"),
    "Euler": lambda config: EulerDiscreteScheduler.from_config(config),
}

def center_crop_resize(img, output_size=(512, 512)):
    width, height = img.size
    new_dimension = min(width, height)
    left = (width - new_dimension)/2
    top = (height - new_dimension)/2
    right = (width + new_dimension)/2
    bottom = (height + new_dimension)/2
    img = img.crop((left, top, right, bottom))
    img = img.resize(output_size)
    return img

def common_upscale(samples, width, height, upscale_method, crop=False):
    return torch.nn.functional.interpolate(samples, size=(height, width), mode=upscale_method)

def upscale(samples, upscale_method, scale_by):
    width = round(samples.shape[3] * scale_by)
    height = round(samples.shape[2] * scale_by)
    s = common_upscale(samples, width, height, upscale_method)
    return s

def check_inputs(prompt: str, control_image: Image.Image):
    if control_image is None:
        raise gr.Error("Please select or upload an Input Illusion")
    if prompt is None or prompt == "":
        raise gr.Error("Prompt is required")

# Inference function
def inference(
    control_image: Image.Image,
    prompt: str,
    negative_prompt: str,
    guidance_scale: float = 8.0,
    controlnet_conditioning_scale: float = 1,
    control_guidance_start: float = 1,    
    control_guidance_end: float = 1,
    upscaler_strength: float = 0.5,
    seed: int = -1,
    sampler = "DPM++ Karras SDE",
    progress = gr.Progress(track_tqdm=True),
    profile: gr.OAuthProfile | None = None,
):
    start_time = time.time()
    
    control_image_small = center_crop_resize(control_image)
    control_image_large = center_crop_resize(control_image, (1024, 1024))

    main_pipe.scheduler = SAMPLER_MAP[sampler](main_pipe.scheduler.config)
    my_seed = random.randint(0, 2**32 - 1) if seed == -1 else seed
    generator = torch.Generator(device=device).manual_seed(my_seed)
    
    # Reducing steps for CPU performance
    out = main_pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        image=control_image_small,
        guidance_scale=float(guidance_scale),
        controlnet_conditioning_scale=float(controlnet_conditioning_scale),
        generator=generator,
        control_guidance_start=float(control_guidance_start),
        control_guidance_end=float(control_guidance_end),
        num_inference_steps=12, # Dropped steps for speed
        output_type="latent"
    )
    
    upscaled_latents = upscale(out.images if hasattr(out, 'images') else out[0], "nearest-exact", 2)
    
    out_image = image_pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        control_image=control_image_large,        
        image=upscaled_latents,
        guidance_scale=float(guidance_scale),
        generator=generator,
        num_inference_steps=15, # Dropped steps for speed
        strength=upscaler_strength,
        control_guidance_start=float(control_guidance_start),
        control_guidance_end=float(control_guidance_end),
        controlnet_conditioning_scale=float(controlnet_conditioning_scale)
    )
    
    end_time = time.time()
    print(f"Inference took {end_time-start_time}s")

    user_history.save_image(
        label=prompt,
        image=out_image["images"][0],
        profile=profile,
        metadata={"prompt": prompt, "seed": my_seed},
    )

    return out_image["images"][0], gr.update(visible=True), gr.update(visible=True), my_seed
        
with gr.Blocks(css=css) as app:
    gr.Markdown("<div style='text-align: center;'><h1>Illusion Diffusion CPU 🌀</h1></div>")

    with gr.Row():
        with gr.Column():
            control_image = gr.Image(label="Input Illusion", type="pil")
            controlnet_conditioning_scale = gr.Slider(minimum=0.0, maximum=5.0, step=0.01, value=0.8, label="Illusion strength")
            prompt = gr.Textbox(label="Prompt", placeholder="Medieval village scene")
            negative_prompt = gr.Textbox(label="Negative Prompt", value="low quality")
            
            with gr.Accordion(label="Advanced Options", open=False):
                guidance_scale = gr.Slider(minimum=0.0, maximum=50.0, step=0.25, value=7.5, label="Guidance Scale")
                sampler = gr.Dropdown(choices=list(SAMPLER_MAP.keys()), value="Euler")
                control_start = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0, label="Start of ControlNet")
                control_end = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1, label="End of ControlNet")
                strength = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1, label="Strength of the upscaler")
                seed = gr.Slider(minimum=-1, maximum=9999999999, step=1, value=-1, label="Seed")
                used_seed = gr.Number(label="Last seed used", interactive=False)
            run_btn = gr.Button("Run")
            
        with gr.Column():
            result_image = gr.Image(label="Output", interactive=False)
            with gr.Group(visible=False) as share_group:
                share_button = gr.Button("Share to community")

    run_btn.click(
        check_inputs,
        inputs=[prompt, control_image],
        queue=False
    ).success(
        inference,
        inputs=[control_image, prompt, negative_prompt, guidance_scale, controlnet_conditioning_scale, control_start, control_end, strength, seed, sampler],
        outputs=[result_image, result_image, share_group, used_seed])

app.queue(max_size=10).launch()