app.py

import gradio as gr
import torch
import numpy as np
from diffusers import StableDiffusionXLImg2ImgPipeline
from transformers import DPTFeatureExtractor, DPTForDepthEstimation
from PIL import Image, ImageEnhance, ImageOps

device = "cpu"  # or "cuda" if you have a GPU
torch_dtype = torch.float32

print("Loading SDXL Img2Img model...")
pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch_dtype
).to(device)

print("Loading bas-relief LoRA weights with PEFT...")
pipe.load_lora_weights(
    "KappaNeuro/bas-relief",
    weight_name="BAS-RELIEF.safetensors",
    peft_backend="peft"
)

print("Loading DPT Depth Model...")
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
depth_model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large").to(device)


def enhance_depth_map(depth_arr: np.ndarray) -> Image.Image:
    d_min, d_max = depth_arr.min(), depth_arr.max()
    depth_stretched = (depth_arr - d_min) / (d_max - d_min + 1e-8)
    depth_stretched = (depth_stretched * 255).astype(np.uint8)

    depth_pil = Image.fromarray(depth_stretched)
    depth_pil = ImageOps.autocontrast(depth_pil)

    enhancer = ImageEnhance.Sharpness(depth_pil)
    depth_pil = enhancer.enhance(2.0)

    return depth_pil


def generate_bas_relief_and_depth(input_image: Image.Image):
    # Redimensionar a imagem para o tamanho esperado
    input_image = input_image.resize((512, 512))

    # Prompt fixo para ativar o LoRA
    prompt = "BAS-RELIEF"

    print("Gerando imagem no estilo baixo-relevo...")
    result = pipe(
        prompt=prompt,
        image=input_image,
        strength=0.7,  # Controla a intensidade da transformação
        num_inference_steps=15,
        guidance_scale=7.5
    )
    generated_image = result.images[0]

    print("Calculando mapa de profundidade...")
    inputs = feature_extractor(generated_image, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = depth_model(**inputs)
        predicted_depth = outputs.predicted_depth

    prediction = torch.nn.functional.interpolate(
        predicted_depth.unsqueeze(1),
        size=generated_image.size[::-1],
        mode="bicubic",
        align_corners=False
    ).squeeze()

    depth_map_pil = enhance_depth_map(prediction.cpu().numpy())

    return generated_image, depth_map_pil


title = "Conversor para Baixo-relevo (SDXL + LoRA) com Mapa de Profundidade"
description = (
    "Carrega stable-diffusion-xl-base-1.0 no CPU, aplica LoRA de 'KappaNeuro/bas-relief' "
    "para transformar imagens em baixo-relevo e calcula o mapa de profundidade correspondente."
)

iface = gr.Interface(
    fn=generate_bas_relief_and_depth,
    inputs=gr.Image(label="Imagem de Entrada", type="pil"),
    outputs=[
        gr.Image(label="Imagem em Baixo-relevo"),
        gr.Image(label="Mapa de Profundidade")
    ],
    title=title,
    description=description
)

if __name__ == "__main__":
    iface.launch()