app.py

import torch
import numpy as np
import gradio as gr
from diffusers import (
    StableDiffusionXLPipeline, 
    AutoPipelineForInpainting, 
    TCDScheduler,
    ControlNetModel,
    StableDiffusionXLControlNetPipeline,
    MotionAdapter,
    AnimateDiffPipeline
)
from diffusers.utils import make_image_grid, export_to_gif
from PIL import Image
import io
import requests
from transformers import DPTImageProcessor, DPTForDepthEstimation

# Available models
AVAILABLE_MODELS = {
    "Stable Diffusion XL": "stabilityai/stable-diffusion-xl-base-1.0",
    "Animagine XL 3.0": "cagliostrolab/animagine-xl-3.0",
}

# Available LoRA styles
AVAILABLE_LORAS = {
    "TCD": "h1t/TCD-SDXL-LoRA",
    "Papercut": "TheLastBen/Papercut_SDXL",
}

def get_depth_map(image):
    # Initialize depth estimator
    depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas")
    feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
    
    # Process image
    image = feature_extractor(images=image, return_tensors="pt").pixel_values
    with torch.no_grad():
        depth_map = depth_estimator(image).predicted_depth

    # Resize and normalize depth map
    depth_map = torch.nn.functional.interpolate(
        depth_map.unsqueeze(1),
        size=(1024, 1024),
        mode="bicubic",
        align_corners=False,
    )
    depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
    depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
    depth_map = (depth_map - depth_min) / (depth_max - depth_min)
    image = torch.cat([depth_map] * 3, dim=1)

    # Convert to PIL Image
    image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
    image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
    return image

def load_image_from_url(url):
    response = requests.get(url)
    return Image.open(io.BytesIO(response.content)).convert("RGB")

def generate_image(prompt, seed, num_steps, guidance_scale, eta):
    # Initialize the pipeline
    base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
    tcd_lora_id = "h1t/TCD-SDXL-LoRA"
    
    # Use CPU for inference
    pipe = StableDiffusionXLPipeline.from_pretrained(
        base_model_id,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load and fuse LoRA weights
    pipe.load_lora_weights(tcd_lora_id)
    pipe.fuse_lora()
    
    # Generate the image
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        eta=eta,
        generator=generator,
    ).images[0]
    
    return image

def generate_community_image(prompt, model_name, seed, num_steps, guidance_scale, eta):
    # Initialize the pipeline
    base_model_id = AVAILABLE_MODELS[model_name]
    tcd_lora_id = "h1t/TCD-SDXL-LoRA"
    
    # Use CPU for inference
    pipe = StableDiffusionXLPipeline.from_pretrained(
        base_model_id,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load and fuse LoRA weights
    pipe.load_lora_weights(tcd_lora_id)
    pipe.fuse_lora()
    
    # Generate the image
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        eta=eta,
        generator=generator,
    ).images[0]
    
    return image

def generate_style_mix(prompt, seed, num_steps, guidance_scale, eta, style_weight):
    # Initialize the pipeline
    base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
    tcd_lora_id = "h1t/TCD-SDXL-LoRA"
    styled_lora_id = "TheLastBen/Papercut_SDXL"
    
    # Use CPU for inference
    pipe = StableDiffusionXLPipeline.from_pretrained(
        base_model_id,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load multiple LoRA weights
    pipe.load_lora_weights(tcd_lora_id, adapter_name="tcd")
    pipe.load_lora_weights(styled_lora_id, adapter_name="style")
    
    # Set adapter weights
    pipe.set_adapters(["tcd", "style"], adapter_weights=[1.0, style_weight])
    
    # Generate the image
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        eta=eta,
        generator=generator,
    ).images[0]
    
    return image

def generate_controlnet(prompt, init_image, seed, num_steps, guidance_scale, eta, controlnet_scale):
    # Initialize the pipeline
    base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
    controlnet_id = "diffusers/controlnet-depth-sdxl-1.0"
    tcd_lora_id = "h1t/TCD-SDXL-LoRA"
    
    # Initialize ControlNet
    controlnet = ControlNetModel.from_pretrained(
        controlnet_id,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    
    # Initialize pipeline
    pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
        base_model_id,
        controlnet=controlnet,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load and fuse LoRA weights
    pipe.load_lora_weights(tcd_lora_id)
    pipe.fuse_lora()
    
    # Generate depth map
    depth_image = get_depth_map(init_image)
    
    # Generate the image
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        image=depth_image,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        eta=eta,
        controlnet_conditioning_scale=controlnet_scale,
        generator=generator,
    ).images[0]
    
    # Create a grid of the depth map and result
    grid = make_image_grid([depth_image, image], rows=1, cols=2)
    return grid

def inpaint_image(prompt, init_image, mask_image, seed, num_steps, guidance_scale, eta, strength):
    # Initialize the pipeline
    base_model_id = "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"
    tcd_lora_id = "h1t/TCD-SDXL-LoRA"
    
    # Use CPU for inference
    pipe = AutoPipelineForInpainting.from_pretrained(
        base_model_id,
        torch_dtype=torch.float32  # Use float32 for CPU
    )
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load and fuse LoRA weights
    pipe.load_lora_weights(tcd_lora_id)
    pipe.fuse_lora()
    
    # Generate the image
    generator = torch.Generator().manual_seed(seed)
    image = pipe(
        prompt=prompt,
        image=init_image,
        mask_image=mask_image,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        eta=eta,
        strength=strength,
        generator=generator,
    ).images[0]
    
    # Create a grid of the original image, mask, and result
    grid = make_image_grid([init_image, mask_image, image], rows=1, cols=3)
    return grid

def generate_animation(prompt, seed, num_steps, guidance_scale, eta, num_frames, motion_scale):
    # Initialize the pipeline
    base_model_id = "frankjoshua/toonyou_beta6"
    motion_adapter_id = "guoyww/animatediff-motion-adapter-v1-5"
    tcd_lora_id = "h1t/TCD-SD15-LoRA"
    motion_lora_id = "guoyww/animatediff-motion-lora-zoom-in"
    
    # Load motion adapter
    adapter = MotionAdapter.from_pretrained(motion_adapter_id)
    
    # Initialize pipeline with CPU optimization
    pipe = AnimateDiffPipeline.from_pretrained(
        base_model_id,
        motion_adapter=adapter,
        torch_dtype=torch.float32,  # Use float32 for CPU
        low_cpu_mem_usage=True,     # Enable low CPU memory usage
        use_safetensors=False       # Use standard PyTorch weights
    )
    
    # Set TCD scheduler
    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
    
    # Load LoRA weights
    pipe.load_lora_weights(tcd_lora_id, adapter_name="tcd")
    pipe.load_lora_weights(
        motion_lora_id,
        adapter_name="motion-lora"
    )
    
    # Set adapter weights
    pipe.set_adapters(["tcd", "motion-lora"], adapter_weights=[1.0, motion_scale])
    
    # Generate animation
    generator = torch.Generator().manual_seed(seed)
    frames = pipe(
        prompt=prompt,
        num_inference_steps=num_steps,
        guidance_scale=guidance_scale,
        cross_attention_kwargs={"scale": 1},
        num_frames=num_frames,
        eta=eta,
        generator=generator
    ).frames[0]
    
    # Export to GIF
    gif_path = "animation.gif"
    export_to_gif(frames, gif_path)
    return gif_path

# Create the Gradio interface
with gr.Blocks(title="TCD-SDXL Image Generator") as demo:
    gr.Markdown("# TCD-SDXL Image Generator")
    gr.Markdown("Generate images using Trajectory Consistency Distillation with Stable Diffusion XL. Note: This runs on CPU, so generation may take some time.")
    
    with gr.Tabs():
        with gr.TabItem("Text to Image"):
            with gr.Row():
                with gr.Column():
                    text_prompt = gr.Textbox(
                        label="Prompt",
                        value="Painting of the orange cat Otto von Garfield, Count of Bismarck-Schönhausen, Duke of Lauenburg, Minister-President of Prussia. Depicted wearing a Prussian Pickelhaube and eating his favorite meal - lasagna.",
                        lines=3
                    )
                    text_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    text_steps = gr.Slider(minimum=1, maximum=10, value=4, label="Number of Steps", step=1)
                    text_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    text_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    text_button = gr.Button("Generate")
                with gr.Column():
                    text_output = gr.Image(label="Generated Image")
            
            text_button.click(
                fn=generate_image,
                inputs=[text_prompt, text_seed, text_steps, text_guidance, text_eta],
                outputs=text_output
            )
        
        with gr.TabItem("Inpainting"):
            with gr.Row():
                with gr.Column():
                    inpaint_prompt = gr.Textbox(
                        label="Prompt",
                        value="a tiger sitting on a park bench",
                        lines=3
                    )
                    init_image = gr.Image(label="Initial Image", type="pil")
                    mask_image = gr.Image(label="Mask Image", type="pil")
                    inpaint_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    inpaint_steps = gr.Slider(minimum=1, maximum=10, value=8, label="Number of Steps", step=1)
                    inpaint_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    inpaint_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    inpaint_strength = gr.Slider(minimum=0, maximum=1, value=0.99, label="Strength")
                    inpaint_button = gr.Button("Inpaint")
                with gr.Column():
                    inpaint_output = gr.Image(label="Result (Original | Mask | Generated)")
            
            inpaint_button.click(
                fn=inpaint_image,
                inputs=[
                    inpaint_prompt, init_image, mask_image, inpaint_seed,
                    inpaint_steps, inpaint_guidance, inpaint_eta, inpaint_strength
                ],
                outputs=inpaint_output
            )
            
        with gr.TabItem("Community Models"):
            with gr.Row():
                with gr.Column():
                    community_prompt = gr.Textbox(
                        label="Prompt",
                        value="A man, clad in a meticulously tailored military uniform, stands with unwavering resolve. The uniform boasts intricate details, and his eyes gleam with determination. Strands of vibrant, windswept hair peek out from beneath the brim of his cap.",
                        lines=3
                    )
                    model_dropdown = gr.Dropdown(
                        choices=list(AVAILABLE_MODELS.keys()),
                        value="Animagine XL 3.0",
                        label="Select Model"
                    )
                    community_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    community_steps = gr.Slider(minimum=1, maximum=10, value=8, label="Number of Steps", step=1)
                    community_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    community_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    community_button = gr.Button("Generate")
                with gr.Column():
                    community_output = gr.Image(label="Generated Image")
            
            community_button.click(
                fn=generate_community_image,
                inputs=[
                    community_prompt, model_dropdown, community_seed,
                    community_steps, community_guidance, community_eta
                ],
                outputs=community_output
            )
            
        with gr.TabItem("Style Mixing"):
            with gr.Row():
                with gr.Column():
                    style_prompt = gr.Textbox(
                        label="Prompt",
                        value="papercut of a winter mountain, snow",
                        lines=3
                    )
                    style_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    style_steps = gr.Slider(minimum=1, maximum=10, value=4, label="Number of Steps", step=1)
                    style_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    style_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    style_weight = gr.Slider(minimum=0, maximum=2, value=1.0, label="Style Weight", step=0.1)
                    style_button = gr.Button("Generate")
                with gr.Column():
                    style_output = gr.Image(label="Generated Image")
            
            style_button.click(
                fn=generate_style_mix,
                inputs=[
                    style_prompt, style_seed, style_steps,
                    style_guidance, style_eta, style_weight
                ],
                outputs=style_output
            )
            
        with gr.TabItem("ControlNet"):
            with gr.Row():
                with gr.Column():
                    control_prompt = gr.Textbox(
                        label="Prompt",
                        value="stormtrooper lecture, photorealistic",
                        lines=3
                    )
                    control_image = gr.Image(label="Input Image", type="pil")
                    control_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    control_steps = gr.Slider(minimum=1, maximum=10, value=4, label="Number of Steps", step=1)
                    control_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    control_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    control_scale = gr.Slider(minimum=0, maximum=1, value=0.5, label="ControlNet Scale", step=0.1)
                    control_button = gr.Button("Generate")
                with gr.Column():
                    control_output = gr.Image(label="Result (Depth Map | Generated)")
            
            control_button.click(
                fn=generate_controlnet,
                inputs=[
                    control_prompt, control_image, control_seed,
                    control_steps, control_guidance, control_eta, control_scale
                ],
                outputs=control_output
            )

        with gr.TabItem("Animation"):
            with gr.Row():
                with gr.Column():
                    anim_prompt = gr.Textbox(
                        label="Prompt",
                        value="best quality, masterpiece, 1girl, looking at viewer, blurry background, upper body, contemporary, dress",
                        lines=3
                    )
                    anim_seed = gr.Slider(minimum=0, maximum=2147483647, value=0, label="Seed", step=1)
                    anim_steps = gr.Slider(minimum=1, maximum=10, value=5, label="Number of Steps", step=1)
                    anim_guidance = gr.Slider(minimum=0, maximum=1, value=0, label="Guidance Scale")
                    anim_eta = gr.Slider(minimum=0, maximum=1, value=0.3, label="Eta")
                    anim_frames = gr.Slider(minimum=8, maximum=32, value=24, label="Number of Frames", step=1)
                    anim_motion_scale = gr.Slider(minimum=0, maximum=2, value=1.2, label="Motion Scale", step=0.1)
                    anim_button = gr.Button("Generate Animation")
                with gr.Column():
                    anim_output = gr.Image(label="Generated Animation", format="gif")
            
            anim_button.click(
                fn=generate_animation,
                inputs=[
                    anim_prompt, anim_seed, anim_steps,
                    anim_guidance, anim_eta, anim_frames,
                    anim_motion_scale
                ],
                outputs=anim_output
            )

if __name__ == "__main__":
    demo.launch()