app.py

import os
import subprocess

# Убираем pyenv
os.environ.pop("PYENV_VERSION", None)

# Установка зависимостей
subprocess.run(["pip", "install", "torch", "wheel"], check=True)
subprocess.run([
    "pip", "install", "--no-build-isolation", 
    "diso@git+https://github.com/SarahWeiii/diso.git"
], check=True)

# Импорты (перенесены после установки зависимостей)
import gradio as gr
import uuid
import torch
import zipfile
import requests
import traceback
import trimesh
import numpy as np
from trimesh.exchange.gltf import export_glb

from inference_triposg import run_triposg
from triposg.pipelines.pipeline_triposg import TripoSGPipeline
from briarmbg import BriaRMBG


print("Trimesh version:", trimesh.__version__)

# Настройки устройства
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

# Загрузка весов
weights_dir = "pretrained_weights"
triposg_path = os.path.join(weights_dir, "TripoSG")
rmbg_path = os.path.join(weights_dir, "RMBG-1.4")

if not (os.path.exists(triposg_path) and os.path.exists(rmbg_path)):
    print("📦 Downloading pretrained weights...")
    url = "https://huggingface.co/datasets/endlesstools/pretrained-assets/resolve/main/pretrained_models.zip"
    zip_path = "pretrained_models.zip"

    with requests.get(url, stream=True) as r:
        r.raise_for_status()
        with open(zip_path, "wb") as f:
            for chunk in r.iter_content(chunk_size=8192):
                f.write(chunk)

    print("📦 Extracting weights...")
    with zipfile.ZipFile(zip_path, "r") as zip_ref:
        zip_ref.extractall(weights_dir)

    os.remove(zip_path)
    print("✅ Weights ready.")

# Загрузка моделей
pipe = TripoSGPipeline.from_pretrained(triposg_path).to(device, dtype)
rmbg_net = BriaRMBG.from_pretrained(rmbg_path).to(device)
rmbg_net.eval()

# Генерация .glb
# def generate(image_path, face_number=50000, guidance_scale=5.0, num_steps=25):
#     print("[API CALL] image_path received:", image_path)
#     print("[API CALL] File exists:", os.path.exists(image_path))

#     temp_id = str(uuid.uuid4())
#     output_path = f"/tmp/{temp_id}.glb"
#     print("[DEBUG] Generating mesh from:", image_path)

#     try:
#         mesh = run_triposg(
#             pipe=pipe,
#             image_input=image_path,
#             rmbg_net=rmbg_net,
#             seed=42,
#             num_inference_steps=int(num_steps),
#             guidance_scale=float(guidance_scale),
#             faces=int(face_number),
#         )

#         if mesh is None or mesh.vertices.shape[0] == 0 or mesh.faces.shape[0] == 0:
#             raise ValueError("Mesh generation returned an empty mesh")

#         mesh = trimesh.Trimesh(vertices=mesh.vertices, faces=mesh.faces)
#         mesh.rezero()
#         mesh.fix_normals()
#         mesh.apply_translation(-mesh.center_mass)  

#         # Масштабируем, чтобы модель вписывалась в размер 1x1x1 
#         # Если нужно будет подгонять под размер в Endless Tools, то можно использовать:
#         # scale_factor = 1.0 / np.max(np.linalg.norm(mesh.vertices, axis=1))
#         # mesh.apply_scale(scale_factor)


#         glb_data = mesh.export(file_type='glb')
#         with open(output_path, "wb") as f:
#             f.write(glb_data)

#         print(f"[DEBUG] Mesh saved to {output_path}")
#         return output_path if os.path.exists(output_path) else None

#     except Exception as e:
#         print("[ERROR]", e)
#         traceback.print_exc()
#         return f"Error: {e}"

def generate(image_path, face_number=50000, guidance_scale=5.0, num_steps=25):
    print("[API CALL] image_path received:", image_path)
    print("[API CALL] File exists:", os.path.exists(image_path))

    temp_id = str(uuid.uuid4())
    output_path = f"/tmp/{temp_id}.glb"
    print("[DEBUG] Generating mesh from:", image_path)

    try:
        mesh = run_triposg(
            pipe=pipe,
            image_input=image_path,
            rmbg_net=rmbg_net,
            seed=42,
            num_inference_steps=int(num_steps),
            guidance_scale=float(guidance_scale),
            faces=int(face_number),
        )

        if mesh is None or mesh.vertices.shape[0] == 0 or mesh.faces.shape[0] == 0:
            raise ValueError("Mesh generation returned an empty mesh")

        # 🔧 Пересоздаём Trimesh и гарантируем чистоту геометрии
        mesh = trimesh.Trimesh(vertices=mesh.vertices, faces=mesh.faces, process=True)

        # ✅ Центрируем модель
        mesh.apply_translation(-mesh.center_mass)

        # ✅ Масштабируем к единичному размеру (все модели ~одинаковые)
        scale_factor = 1.0 / np.max(np.linalg.norm(mesh.vertices, axis=1))
        mesh.apply_scale(scale_factor)

        # ✅ Гарантированно пересчитываем нормали
        mesh.fix_normals()

        # print("[DEBUG] Normals present:", mesh.has_vertex_normals)
        if hasattr(mesh, "vertex_normals"):
            print("[DEBUG] Normals shape:", mesh.vertex_normals.shape)
        else:
            print("[DEBUG] Normals missing.")


        # 💾 Сохраняем GLB
        glb_data = mesh.export(file_type='glb')
        with open(output_path, "wb") as f:
            f.write(glb_data)

        print(f"[DEBUG] Mesh saved to {output_path}")
        return output_path if os.path.exists(output_path) else None

    except Exception as e:
        print("[ERROR]", e)
        traceback.print_exc()
        return f"Error: {e}"


# Интерфейс Gradio
demo = gr.Interface(
    fn=generate,
    inputs=gr.Image(type="filepath", label="Upload image"),
    outputs=gr.File(label="Download .glb"),
    title="TripoSG Image to 3D",
    description="Upload an image to generate a 3D model (.glb)",
)

# Запуск
demo.launch()