app.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 

controlnet = ControlNetModel.from_pretrained(
    "briaai/BRIA-2.2-ControlNet-Depth",
    torch_dtype=torch.float16
).to('cuda')

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "briaai/BRIA-2.2",
    controlnet=controlnet,
    torch_dtype=torch.float16,
    device_map='auto',
    low_cpu_mem_usage=True,
    offload_state_dict=True,
).to('cuda').to(torch.float16)
pipe.scheduler = EulerAncestralDiscreteScheduler(
    beta_start=0.00085,
    beta_end=0.012,
    beta_schedule="scaled_linear",
    num_train_timesteps=1000,
    steps_offset=1
)
# pipe.enable_freeu(b1=1.1, b2=1.1, s1=0.5, s2=0.7)
# pipe.enable_xformers_memory_efficient_attention()
pipe.force_zeros_for_empty_prompt = False

from transformers import DPTFeatureExtractor, DPTForDepthEstimation
depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")

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_depth_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, canny_image, num_steps, controlnet_conditioning_scale, seed):
    generator = torch.Generator("cuda").manual_seed(seed)    
    images = pipe(
    prompt, negative_prompt=negative_prompt, image=canny_image, num_inference_steps=num_steps, controlnet_conditioning_scale=float(controlnet_conditioning_scale),
    generator=generator,
    ).images
    return images

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

    return [depth_image, images[0]]



block = gr.Blocks().queue()

with block:
    gr.Markdown("## BRIA 2.2 ControlNet Depth")
    gr.HTML('''
      <p style="margin-bottom: 10px; font-size: 94%">
        This is a demo for ControlNet Depth that using
        <a href="https://huggingface.co/briaai/BRIA-2.2" target="_blank">BRIA 2.2 text-to-image model</a> as backbone. 
        Trained on licensed data, BRIA 2.2 provide full legal liability coverage for copyright and privacy infringement.
      </p>
    ''')
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(sources=None, type="pil") # None for upload, ctrl+v and webcam
            prompt = gr.Textbox(label="Prompt")
            negative_prompt = gr.Textbox(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=100, value=50, step=1)
            controlnet_conditioning_scale = gr.Slider(label="ControlNet conditioning scale", minimum=0.1, maximum=2.0, value=1.0, 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)