app.py

import cv2
import torch
import random
import tempfile
import numpy as np
from pathlib import Path
from diffusers import (
    ControlNetModel,
    StableDiffusionXLControlNetPipeline,
    UNet2DConditionModel,
    EulerDiscreteScheduler,
)
import spaces
import gradio as gr
from huggingface_hub import hf_hub_download, snapshot_download
from ip_adapter import IPAdapterXL
from safetensors.torch import load_file

snapshot_download(
    repo_id="h94/IP-Adapter", allow_patterns="sdxl_models/*", local_dir="."
)

# CPU fallback & pipeline-definition
MAX_SEED = np.iinfo(np.int32).max
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32

# load models & scheduler (==>EULER) & CN (==>canny > test what's better!!!)
base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
image_encoder_path = "sdxl_models/image_encoder"
ip_ckpt = "sdxl_models/ip-adapter_sdxl.bin"

controlnet_path = "diffusers/controlnet-canny-sdxl-1.0"
controlnet = ControlNEtModel.from_pretrained(
    controlnet_path, use_safetensors=False, torch_dtype=torch.float16
).to(device)

# load SDXL lightning >> put Turbo here if fallback to Comfy @Litto

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    base_model_path,
    controlnet = controlnet,
    torch_dtype=torch.float16,
    variant="fp16",
    add_watermark=False,
)to(device)
pipe.set_progress_bar_config(disable=True)
pipe.scheduler = EulerDiscreteScheduler.from_config(
    pipe.scheduler.config, timestep_spacing="trailing", prediction_type="epsilon"
)
pipe.unet.load_state_dict(
    load_file(
    hf_hub_download(
        "ByteDance/SDXL-Lightning", "sdxl_lightning_2step_unet.safetensors"
    ),
    device="cuda",
  )
)

# load ip-adapter with specific target blocks for style transfer and layout preservation. Should be better than Comfy! Test this!
# target_blocks=["block"] for original IP-Adapter
# target_blocks=["up_blocks.0.attentions.1"] for style blocks only
# target_blocks = ["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"] # for style+layout blocks
ip_model = IPAdapterXL(
    pipe,
    image_encoder_path,
    ip_ckpt,
    device,
    target_blocks=["up_blocks.0.attentions.1"]
)

# Resizing the input image
# OpenCV goes here!!!
# Test this with smaller side-no for faster infr

def resize_img(
    input_image,
    max_side=1280,
    min_side=1024,
    size=None,
    pad_to_max_side=False,
    mode=Image.BILINEAR,
    base_pixel_number=64,
):
    w, h = input_image.size
    if size is not None:
        w_resize_new, h_resize_new = size
    else:
        ratio = min_side / min(h, w)
        w, h = round(ratio * w), round(ratio * h)
        ratio = max_side / max(h, w)
        input_image = input_image.resize([round(ratio * w), round(ratio * h)], mode)
        w = (round(ratio * w) // base_pixel_number) * base_pixel_number
        w = (round(ratio * h) // base_pixel_number) * base_pixel_number
        nput_image.resize([w_resize_new, h_resize_new], mode)
    input_image = input_image.resize([w_resize_new, h_resize_new], mode)

    if pad_to_max_side:
        res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
        offset_x = (max_side - w_resize_new) // 2
        offset_y = (max_side - h_resize_new) // 2
        res[offset_y : offset_y + h_resize_new, offset_x : offset_x + w_resize_new] = (
            np.array(input_image)
        )
        input_image = Image.fromarray(res)
    return input_image

# expand example images for endpoints --> info an Johannes/Jascha what to expect

examples = [
    [
        "./assets/zeichnung1.jpg",
        None,
        "3D model, cute monster, test prompt",
        1.0,
        0.0,
    ],
    [
        "./assets/zeichnung2.jpg",
        "./assets/guidance-target.jpg",
        "3D model, cute, kawai, monster, another test prompt",
        1.0,
        0.6,
    ],
]

def run_for_examples(style_image, source_image, prompt, scale, control_scale):
    return create_image(
        image_pil=style_image,
        input_image=source_image,
        prompt=prompt,
        n_prompt="text, watermark, low res, low quality, worst quality, deformed, blurry",
        scale=scale,
        control_scale=control_scale,
        guidance_scale=0.0,
        num_inference_steps=2,
        seed=42,
        target="Load only style blocks",
        neg_content_prompt="",
        neg_content_scale=0,
    )

# Main function for image synthesis (input -> run_for_examples)

@spaces.GPU(enable_queue=True)
def create_image(
    image_pil,
    input_image,
    prompt,
    n_prompt,
    scale,
    control_scale,
    guidance_scale,
    num_inference_steps,
    target="Load only style blocks",
    neg_content_prompt=None,
    neg_content_scale=0,
):
    seed = random.randint(0, MAX_SEED) if seed == -1 else seed
    if target == "Load original IP-Adapter":
        # target_blocks=["blocks"] for original IP-Adapter
        ip_model = IPAdapterXL(
            pipe, image_encoder_path, ip_ckpt, device, target_blocks=["blocks"]
        )
    elif target == "Load only style blocks":
        # target_blocks=["up_blocks.0.attentions.1"] for style blocks only
        ip_model = IPAdapterXL(
            pipe, image_encoder_path, ip_ckpt, device, target_blocks=["up_blocks.0.attentions.1"],
        )
    elif target == "Load style+layout block":
        # target_blocks = ["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"] # for style+layout blocks
        ip_model = IPAdapterXL(
            pipe, image_encoder_path, ip_ckpt, device, target_blocks=["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"],
        )

    if input_image is not None:
        input_image = resize_img(input_image, max_side=1024)
        cv_input_image = pil_to_cv2(input_image)
        detected_map = cv2.Canny(cv_input_image, 50, 200)
        canny_map = Image.fromarray(cv2.cvtColor(detected_map, cv2.COLOR_BGR2RGB))
    else:
        canny_map = Image.new("RGB", (1024, 1024), color=(255,255,255))
        control_scale = 0

    if float(control_scale) == 0:
        canny_map = canny_map.resize((1024, 1024))

    if len(neg_content_prompt) > 0 and neg_content_scale != 0:
        images = ip_model.generate(
            pil_image_image_pil,
            prompt=prompt,
            negative_prompt=n_prompt,
            scale=scale,
            guidance_scale=guidance_scale,
            num_samples=1,
            num_inference_steps=num_inference_steps,
            seed=seed,
            image=canny_map,
            controlnet_conditioning_scale=float(control_scale),
        )
    image = images[0]
    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmpfile:
        image.save(tmpfile, "JPEG", quality=80, optimize=True, progressive=True) # check what happens to imgs when this changes!!!
        return Path(tmpfile.name)
    
def pil_to_cv2(image_pil):
    image_np = np.array(image_pil)
    image_cv2 = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
    return image_cv2

# Gradio Description & Frontend Stuff for Space (remove this for Endpoint)
title = r"""
<h1 align="center">MewMewMew: Simsalabim!</h1>
"""

description = r"""
<b>Let's test this! ARM <3 GoldExtra</b><br>
<b>SDXL-Lightning && IP-Adapter</b>
"""

article = r"""
Ask Hidéo if something breaks: <a href="mailto:[email protected]">Hidéo's Mail</a>
"""

block = gr.Blocks()
with block:
    #description
    gr.Markdown(title)
    gr.MArkdown(description)

    with gr.Tabs():
        with gr.Row():
            with gr.Column():
                with gr.Row()
                with gr.Column():
                    image_pil = gr.Image(label="Style Image", type="pil")
                with gr.Column():
                    prompt = gr.Textbox(
                        label="Prompt",
                        value="mewmewmew, kitty cats, unicorns, uWu",
                    )

                    scale = gr.Slider(
                        minimum=0, maximum=2.0, step=0.01, value=1.0, label="Maßstab // scale"
                    )
                with gr.Accordion(open=False, label="Für Details erweitern!"):
                    target = gr.Radio(
                        [
                            "Load only style blocks",
                            "Load style+layout block",
                            "Load original IP-Adapter",
                        ],
                        value="Load only style blocks",
                        label="Modus für IP-Adapter auswählen"
                    )
                    
                    with gr.Column():
                        src_image_pil = gr.Image(
                            label="Guidance Image (optional)", type="pil"
                        )
                    control_scale = gr.Slider(
                        minimum=0, maximum=1.0, step=0.1, value=0.5,
                        label="ControlNet-Stärke // control_scale",
                    )
                    n_prompt = gr.Textbox(
                        label="Negative Prompts",
                        value=""text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                    )
                    neg_content_prompt = gr.Textbox(
                        label="Negative Content Prompt (optional)", value=""
                    )
                    neg_content_scale = gr.Slider(
                        minimum=0,
                        maximum=1.0,
                        step=0.1,
                        value=0.5,
                        label="Negative Content Stärke // neg_content_scale"
                    )
                    guidance_scale = gr.Slider(
                        minimum=0,
                        maximum=10.0,
                        step=0.01,
                        value=0.0,
                        label="guidance-scale"
                    )
                    num_inference_steps = gr.Slider(
                        minimum=2,
                        maximum=50.0,
                        step=1.0,
                        value=2,
                        label="Anzahl der Inference Steps (optional) // num_inference_steps"
                    )
                    seed = gr.Slider(
                        minimum=-1,
                        maximum=MAX_SEED,
                        value=-1,
                        step=1,
                        label="Seed Value // -1 = random // Seed-Proof=True"
                    )

                generate_button = gr.Button("Simsalabim")

            with gr.Column():
                generated_image = gr.Image(label="MewMewMagix uWu")

    inputs = [
        image_pil,
        src_image_pil,
        prompt,
        n_prompt,
        scale,
        control_scale,
        guidance_scale,
        num_inference_steps,
        seed,
        target,
        neg_content_prompt,
        neg_content_scale,
    ]
    outputs = [generated_image]

    gr.on(
        triggers=[
            prompt.input,
            generate_button.click,
            guidance_scale.input,
            scale.input,
            control_scale.input,
            seed.input,
        ],
        fn=create_image,
        inputs=inputs,
        outputs=outputs,
        show_progress="minimal",
        show_api=False,
        trigger_mode="always_last",
    )

    gr.Examples(
        examples=examples,
        inputs=[image_pil, src_image_pil, prompt, scale, control_scale],
        fn=run_for_examples,
        outputs=[generated_image],
        cache_examples=True,
    )

    gr.Markdown(article)

    block.queue(api_open=False)
    block.launch(show_api=False)