import gradio as gr
import numpy as np
import torch
import spaces
from PIL import Image

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048

import os
hf_token = os.environ.get("HF_TOKEN")

PIPE = None
DEPTH_PROC = None
DEPTH_MODEL = None


def load_models():
    global PIPE, DEPTH_PROC, DEPTH_MODEL
    if PIPE is not None:
        return
    from diffusers import FluxControlPipeline
    from transformers import AutoImageProcessor, DepthAnythingForDepthEstimation
    print("Loading FLUX.1-Depth-dev ...", flush=True)
    PIPE = FluxControlPipeline.from_pretrained(
        "black-forest-labs/FLUX.1-Depth-dev",
        torch_dtype=torch.bfloat16,
        token=hf_token,
        low_cpu_mem_usage=True,
    ).to("cuda")
    PIPE.enable_sequential_cpu_offload()
    PIPE.enable_vae_tiling()
    PIPE.enable_vae_slicing()
    print("FLUX loaded.", flush=True)
    depth_id = "LiheYoung/depth-anything-large-hf"
    print("Loading DepthAnything ...", flush=True)
    DEPTH_PROC = AutoImageProcessor.from_pretrained(depth_id, token=hf_token)
    DEPTH_MODEL = DepthAnythingForDepthEstimation.from_pretrained(depth_id, token=hf_token).to("cuda")
    print("DepthAnything loaded.", flush=True)


def to_depth(image_pil):
    inputs = DEPTH_PROC(images=image_pil, return_tensors="pt").to("cuda")
    with torch.no_grad():
        out = DEPTH_MODEL(**inputs).predicted_depth
    depth = out.squeeze().cpu().float().numpy()
    depth = (depth - depth.min()) / (depth.max() - depth.min() + 1e-8)
    depth = (depth * 255.0).astype("uint8")
    return Image.fromarray(depth).convert("RGB")


@spaces.GPU
def infer(control_image, prompt, seed=42, randomize_seed=False, width=1024, height=1024,
          guidance_scale=3.5, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):
    try:
        load_models()
        if randomize_seed:
            seed = int(np.random.randint(0, MAX_SEED))
        control_image = to_depth(control_image.convert("RGB"))
        image = PIPE(
            prompt=prompt,
            control_image=control_image,
            height=height,
            width=width,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            generator=torch.Generator().manual_seed(seed),
        ).images[0]
        return image, seed
    except Exception as e:
        import traceback
        traceback.print_exc()
        raise


examples = [
    "a tiny astronaut hatching from an egg on the moon",
    "a cat holding a sign that says hello world",
    "an anime illustration of a wiener schnitzel",
]

css = """#col-container { margin: 0 auto; max-width: 520px; }"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(
            "# FLUX.1 Depth [dev]\n"
            "12B param rectified flow transformer structural conditioning, guidance-distilled. "
            "Non-commercial license."
        )
        control_image = gr.Image(label="Upload the image for control", type="pil")
        with gr.Row():
            prompt = gr.Text(label="Prompt", show_label=False, max_lines=1,
                             placeholder="Enter your prompt", container=False)
            run_button = gr.Button("Run", scale=0)
        result = gr.Image(label="Result", show_label=False)
        with gr.Accordion("Advanced Settings", open=False):
            seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
            with gr.Row():
                width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
                height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024)
            with gr.Row():
                guidance_scale = gr.Slider(label="Guidance Scale", minimum=1, maximum=30, step=0.5, value=10)
                num_inference_steps = gr.Slider(label="Number of inference steps", minimum=1, maximum=50, step=1, value=28)

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[control_image, prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
        outputs=[result, seed],
    )

demo.launch()