app_base.py

import torch
import spaces
import random
import os

import numpy as np
import gradio as gr

from PIL import Image
from diffusers import FluxControlNetPipeline, FluxControlNetModel
from diffusers.models import FluxMultiControlNetModel

from controlnet_aux import (
    CannyDetector,
    MidasDetector,
)

from huggingface_hub import login

USE_ZERO_GPU = os.environ.get("USE_ZERO_GPU", "0") == "1"

HF_TOKEN = os.environ.get("HF_TOKEN") 
if HF_TOKEN:
    login(token=HF_TOKEN)

MAX_SEED = np.iinfo(np.int32).max
MAX_SIZE = 1024

styles = [
    "3D Animation",
    "Maomu Ghibli",
]

device = "cuda" if torch.cuda.is_available() else "cpu"

base_model = 'black-forest-labs/FLUX.1-dev'
controlnet_model_union = 'Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro'

controlnet_union = FluxControlNetModel.from_pretrained(controlnet_model_union, torch_dtype=torch.bfloat16)
controlnet = FluxMultiControlNetModel([controlnet_union]) # we always recommend loading via FluxMultiControlNetModel

pipe = FluxControlNetPipeline.from_pretrained(base_model, controlnet=controlnet, torch_dtype=torch.bfloat16)
pipe.to(device)

pipe.unload_lora_weights()

pipe.load_lora_weights(
    "vzhizhi6611/OminiControlArt",
    weight_name=f"v0/3d_animation.safetensors",
    adapter_name="3d_animation",
)
pipe.load_lora_weights(
    "vzhizhi6611/OminiControlArt",
    weight_name=f"v0/maomu_ghibli.safetensors",
    adapter_name="maomu_ghibli",
)

canny_detector = CannyDetector()
midas_detector = MidasDetector.from_pretrained(
    "valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
)
midas_detector = midas_detector.to(device)

def infer(
    input_image,
    prompt,
    style,
    num_inference_steps=24,
    guidance_scale=3.5,
    seed=42,
    randomize_seed=False,
    canny_weight=0.2,
    depth_weight=0.4,
    canny_detect=0.375,
    depth_detect=0.5,
):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    
     # Set Adapter
    activate_adapter_name = {
        "3D Animation": "3d_animation",
        "Maomu Ghibli": "maomu_ghibli",
    }[style]
    pipe.set_adapters(activate_adapter_name)

    control_mode_depth = 2
    control_mode_canny = 0

    w, h = input_image.size
    factor = max(w, h) / MAX_SIZE
    width = int(w / factor)
    height = int(h / factor)

    input_image = input_image.resize((width, height), Image.LANCZOS)

    canny_image = canny_detector(input_image, detect_resolution=int(MAX_SIZE * canny_detect), image_resolution=MAX_SIZE)
    depth_image = midas_detector(input_image, detect_resolution=int(MAX_SIZE * depth_detect), image_resolution=MAX_SIZE)

    control_image = []
    control_mode = []
    controlnet_conditioning_scale = []
    if depth_weight > 0:
        control_mode.append(control_mode_depth)
        controlnet_conditioning_scale.append(depth_weight)
        control_image.append(depth_image)
    if canny_weight > 0:
        control_mode.append(control_mode_canny)
        controlnet_conditioning_scale.append(canny_weight)
        control_image.append(canny_image)

    result_image = pipe(
        prompt, 
        control_image=control_image,
        control_mode=control_mode,
        width=width,
        height=height,
        controlnet_conditioning_scale=controlnet_conditioning_scale,
        num_inference_steps=num_inference_steps, 
        guidance_scale=guidance_scale,
        generator=torch.Generator().manual_seed(seed),
    ).images[0]
    
    return result_image, canny_image, depth_image, seed

if USE_ZERO_GPU:
    infer = spaces.GPU(infer, duration=30)

def create_demo() -> gr.Blocks:

    with gr.Blocks() as demo:
        cropper = gr.State()
        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(label="Prompt", lines=1, value="3d animation style selfie")
                num_inference_steps = gr.Slider(minimum=1, maximum=100, value=24, step=1, label="Num Inference Steps")
                guidance_scale = gr.Slider(minimum=0, maximum=20, value=3.5, step=0.5, label="Guidance Scale")
                with gr.Accordion("Advanced Options", open=False):
                    canny_weight = gr.Slider(minimum=0, maximum=1, value=0.3, step=0.01, label="Canny Weight")
                    depth_weight = gr.Slider(minimum=0, maximum=1, value=0.6, step=0.01, label="Depth Weight")
                    canny_detect = gr.Slider(minimum=0.1, maximum=1, value=0.375, step=0.025, label="Canny Detect")
                    depth_detect = gr.Slider(minimum=0.1, maximum=1, value=0.375, step=0.025, label="Depth Detect")
            with gr.Column():
                seed = gr.Number(label="Seed", value=42)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
                style = gr.Dropdown(label="Style", choices=styles, value=styles[0])
                g_btn = gr.Button("Generate Image")
        
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(label="Input Image", type="pil", interactive=True)
                canny_image = gr.Image(label="Canny Image", type="pil", interactive=False)
            with gr.Column():
                result_image = gr.Image(label="Result Image", type="pil", interactive=False)
                depth_image = gr.Image(label="Depth Image", type="pil", interactive=False)
                seed_output = gr.Number(label="Seed Output", interactive=False)

        g_btn.click(
            fn=infer,
            inputs=[
                input_image,
                prompt,
                style,
                num_inference_steps,
                guidance_scale,
                seed,
                randomize_seed,
                canny_weight,
                depth_weight,
                canny_detect,
                depth_detect
            ],
            outputs=[result_image, canny_image, depth_image, seed_output],
        )

    return demo