app_humanwild.py

import spaces
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL, EulerAncestralDiscreteScheduler
from diffusers.utils import load_image
from PIL import Image
import torch
import numpy as np
import cv2
import gradio as gr
from torchvision import transforms 
import fire
import os

controlnet = ControlNetModel.from_pretrained(
    "geyongtao/HumanWild",
    torch_dtype=torch.float16
).to('cuda')

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

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet,
    vae=vae,
    torch_dtype=torch.float16,
    use_safetensors=True,
    low_cpu_mem_usage=True,
    offload_state_dict=True,
).to('cuda')
pipe.controlnet.to(memory_format=torch.channels_last)

# pipe.enable_xformers_memory_efficient_attention()
pipe.force_zeros_for_empty_prompt = False


def resize_image(image):
    image = image.convert('RGB')
    current_size = image.size
    if current_size[0] > current_size[1]:
        center_cropped_image = transforms.functional.center_crop(image, (current_size[1], current_size[1]))
    else:
        center_cropped_image = transforms.functional.center_crop(image, (current_size[0], current_size[0]))
    resized_image = transforms.functional.resize(center_cropped_image, (1024, 1024))
    return resized_image

def get_normal_map(image):
    image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
    with torch.no_grad(), torch.autocast("cuda"):
        depth_map = depth_estimator(image).predicted_depth
    image = transforms.functional.center_crop(image, min(image.shape[-2:]))
    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)
    image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
    image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
    return image


@spaces.GPU
def generate_(prompt, negative_prompt, normal_image, num_steps, controlnet_conditioning_scale, seed):
    generator = torch.Generator("cuda").manual_seed(seed)    
    images = pipe(
        prompt, 
        negative_prompt=negative_prompt, 
        image=normal_image, 
        num_inference_steps=num_steps, 
        controlnet_conditioning_scale=float(controlnet_conditioning_scale),
        num_images_per_prompt=2,
        generator=generator,
    ).images
    return images

@spaces.GPU
def process(normal_image, prompt, negative_prompt, num_steps, controlnet_conditioning_scale, seed):
    # resize input_image to 1024x1024
    normal_image = resize_image(normal_image)
    # depth_image = get_depth_map(input_image)
    images = generate_(prompt, negative_prompt, normal_image, num_steps, controlnet_conditioning_scale, seed)

    return [images[0], images[1]]


def run_demo():
    
    _TITLE = '''3D Human Reconstruction in the Wild with Synthetic Data Using Generative Models'''

    block = gr.Blocks().queue()
    
    with block:
        gr.Markdown("# 3D Human Reconstruction in the Wild with Synthetic Data Using Generative Models ")
        gr.HTML('''
          <p style="margin-bottom: 10px; font-size: 94%">
            This is a demo for Surface Normal ControlNet that using
            <a href="https://huggingface.co/geyongtao/HumanWild" target="_blank"> HumanWild model</a> pretrained weight. 
            <a style="display:inline-block; margin-left: .5em" href='https://github.com/YongtaoGe/WildHuman/'><img src='https://img.shields.io/github/stars/YongtaoGe/WildHuman?style=social' /></a>
          </p>
        ''')
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(sources=None, type="pil") # None for upload, ctrl+v and webcam

                example_folder = os.path.join(os.path.dirname(__file__), "./assets")
                example_fns = [os.path.join(example_folder, example) for example in os.listdir(example_folder)]
                gr.Examples(
                    examples=example_fns,
                    inputs=[input_image],
                    cache_examples=False,
                    label='Examples (click one of the images below to start)',
                    examples_per_page=30
                )
            
                prompt = gr.Textbox(label="Prompt", value="a person, in the wild")
                negative_prompt = gr.Textbox(visible=False, label="Negative prompt", value="Logo,Watermark,Text,Ugly,Morbid,Extra fingers,Poorly drawn hands,Mutation,Blurry,Extra limbs,Gross proportions,Missing arms,Mutated hands,Long neck,Duplicate,Mutilated,Mutilated hands,Poorly drawn face,Deformed,Bad anatomy,Cloned face,Malformed limbs,Missing legs,Too many fingers")
                num_steps = gr.Slider(label="Number of steps", minimum=25, maximum=50, value=30, step=1)
                controlnet_conditioning_scale = gr.Slider(label="ControlNet conditioning scale", minimum=0.1, maximum=1.0, value=0.95, step=0.05)
                seed = gr.Slider(label="Seed", minimum=0, maximum=2147483647, step=1, randomize=True,)
                run_button = gr.Button(value="Run")

            with gr.Column():
                result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery", columns=[2], height='auto')
        ips = [input_image, prompt, negative_prompt, num_steps, controlnet_conditioning_scale, seed]
        
        run_button.click(fn=process, inputs=ips, outputs=[result_gallery])
    
    block.launch(debug = True)

if __name__ == '__main__':
    fire.Fire(run_demo)