app.py

import gradio as gr
import spaces
from gradio_litmodel3d import LitModel3D
import os
import torch
import numpy as np
import imageio
import uuid
from easydict import EasyDict as edict
from PIL import Image
from trellis.pipelines import TrellisImageTo3DPipeline
from trellis.representations import Gaussian, MeshExtractResult
from trellis.utils import render_utils, postprocessing_utils
from transformers import pipeline as translation_pipeline
from diffusers import FluxPipeline
from typing import *


MAX_SEED = np.iinfo(np.int32).max
TMP_DIR = "/tmp/Trellis-demo"
os.makedirs(TMP_DIR, exist_ok=True)

# 메모리 관련 환경 변수
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128'
os.environ['TRANSFORMERS_CACHE'] = '/tmp/transformers_cache'
os.environ['TORCH_HOME'] = '/tmp/torch_home'
os.environ['HF_HOME'] = '/tmp/huggingface'
os.environ['XDG_CACHE_HOME'] = '/tmp/cache'
os.environ['SPCONV_ALGO'] = 'native'
os.environ['WARP_USE_CPU'] = '1'

def initialize_models():
    global pipeline, translator, flux_pipe
    
    try:
        # 캐시 디렉토리 생성
        for dir_path in ['/tmp/transformers_cache', '/tmp/torch_home', '/tmp/huggingface', '/tmp/cache']:
            os.makedirs(dir_path, exist_ok=True)
        
        # Trellis 파이프라인 초기화
        pipeline = TrellisImageTo3DPipeline.from_pretrained(
            "JeffreyXiang/TRELLIS-image-large"
        )
        
        # 번역기 초기화
        translator = translation_pipeline(
            "translation",
            model="Helsinki-NLP/opus-mt-ko-en",
            device="cpu"
        )
        
        # Flux 파이프라인 초기화 (즉시 로드)
        flux_pipe = FluxPipeline.from_pretrained(
            "black-forest-labs/FLUX.1-dev",
            torch_dtype=torch.float32
        )
        
        print("Models initialized successfully")
        return True
        
    except Exception as e:
        print(f"Model initialization error: {str(e)}")
        return False

def load_flux_pipe():
    """Flux 파이프라인을 필요할 때만 로드"""
    global flux_pipe
    if flux_pipe is None:
        flux_pipe = FluxPipeline.from_pretrained(
            "black-forest-labs/FLUX.1-dev",
            torch_dtype=torch.float32
        )
    return flux_pipe

def free_memory():
    """메모리를 정리하는 유틸리티 함수"""
    import gc
    gc.collect()
    
    if torch.cuda.is_available():
        with torch.cuda.device('cuda'):
            torch.cuda.empty_cache()
    
    # 임시 파일 정리
    for dir_path in ['/tmp/transformers_cache', '/tmp/torch_home', '/tmp/huggingface', '/tmp/cache']:
        if os.path.exists(dir_path):
            for file in os.listdir(dir_path):
                file_path = os.path.join(dir_path, file)
                try:
                    if os.path.isfile(file_path):
                        os.unlink(file_path)
                except Exception as e:
                    print(f'Error deleting {file_path}: {e}')


@spaces.GPU
def setup_gpu_model(model):
    """GPU 설정이 필요한 모델을 처리하는 함수"""
    if torch.cuda.is_available():
        model = model.to("cuda")
    return model


def translate_if_korean(text):
    if any(ord('가') <= ord(char) <= ord('힣') for char in text):
        translated = translator(text)[0]['translation_text']
        return translated
    return text

@spaces.GPU
def preprocess_image(image: Image.Image) -> Tuple[str, Image.Image]:
    try:
        trial_id = str(uuid.uuid4())
        
        # 이미지가 너무 작은 경우 크기 조정
        min_size = 64
        if image.size[0] < min_size or image.size[1] < min_size:
            ratio = min_size / min(image.size)
            new_size = tuple(int(dim * ratio) for dim in image.size)
            image = image.resize(new_size, Image.LANCZOS)
            
        processed_image = pipeline.preprocess_image(image)
        processed_image.save(f"{TMP_DIR}/{trial_id}.png")
        return trial_id, processed_image
        
    except Exception as e:
        print(f"Error in preprocess_image: {str(e)}")
        return None, None

def pack_state(gs: Gaussian, mesh: MeshExtractResult, trial_id: str) -> dict:
    return {
        'gaussian': {
            **gs.init_params,
            '_xyz': gs._xyz.cpu().numpy(),
            '_features_dc': gs._features_dc.cpu().numpy(),
            '_scaling': gs._scaling.cpu().numpy(),
            '_rotation': gs._rotation.cpu().numpy(),
            '_opacity': gs._opacity.cpu().numpy(),
        },
        'mesh': {
            'vertices': mesh.vertices.cpu().numpy(),
            'faces': mesh.faces.cpu().numpy(),
        },
        'trial_id': trial_id,
    }

def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
    gs = Gaussian(
        aabb=state['gaussian']['aabb'],
        sh_degree=state['gaussian']['sh_degree'],
        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
        scaling_bias=state['gaussian']['scaling_bias'],
        opacity_bias=state['gaussian']['opacity_bias'],
        scaling_activation=state['gaussian']['scaling_activation'],
    )
    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
    
    mesh = edict(
        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
    )
    
    return gs, mesh, state['trial_id']

@spaces.GPU
def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_strength: float,
                ss_sampling_steps: int, slat_guidance_strength: float, slat_sampling_steps: int):
    try:
        if randomize_seed:
            seed = np.random.randint(0, MAX_SEED)
            
        input_image = Image.open(f"{TMP_DIR}/{trial_id}.png")
        
        # 이미지 크기 제한
        max_size = 512
        if max(input_image.size) > max_size:
            ratio = max_size / max(input_image.size)
            input_image = input_image.resize(
                (int(input_image.size[0] * ratio), 
                 int(input_image.size[1] * ratio)),
                Image.LANCZOS
            )
        
        if torch.cuda.is_available():
            pipeline.to("cuda")
        
        with torch.no_grad():
            outputs = pipeline.run(
                input_image,
                seed=seed,
                formats=["gaussian", "mesh"],
                preprocess_image=False,
                sparse_structure_sampler_params={
                    "steps": min(ss_sampling_steps, 15),
                    "cfg_strength": ss_guidance_strength,
                },
                slat_sampler_params={
                    "steps": min(slat_sampling_steps, 15),
                    "cfg_strength": slat_guidance_strength,
                }
            )
        
        # 비디오 프레임 수 감소
        video = render_utils.render_video(outputs['gaussian'][0], num_frames=30)['color']
        video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=30)['normal']
        video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
        
        trial_id = str(uuid.uuid4())
        video_path = f"{TMP_DIR}/{trial_id}.mp4"
        os.makedirs(os.path.dirname(video_path), exist_ok=True)
        imageio.mimsave(video_path, video, fps=15)
        
        state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
        
        if torch.cuda.is_available():
            pipeline.to("cpu")
        
        return state, video_path
        
    except Exception as e:
        print(f"Error in image_to_3d: {str(e)}")
        if torch.cuda.is_available():
            pipeline.to("cpu")
        raise e

@spaces.GPU
def generate_image_from_text(prompt, height, width, guidance_scale, num_steps):
    try:
        global flux_pipe
        
        if torch.cuda.is_available():
            flux_pipe = flux_pipe.to("cuda")
            flux_pipe = flux_pipe.to(torch.float16)
        
        # 이미지 크기 제한
        height = min(height, 512)
        width = min(width, 512)
        
        # 프롬프트 처리
        base_prompt = "wbgmsst, 3D, white background"
        translated_prompt = translate_if_korean(prompt)
        final_prompt = f"{translated_prompt}, {base_prompt}"
        
        print(f"Generating image with prompt: {final_prompt}")
        
        with torch.inference_mode():
            output = flux_pipe(
                prompt=[final_prompt],
                height=height,
                width=width,
                guidance_scale=min(guidance_scale, 10.0),
                num_inference_steps=min(num_steps, 30)
            )
            
            image = output.images[0]
        
        if torch.cuda.is_available():
            flux_pipe = flux_pipe.to("cpu")
        
        return image
        
    except Exception as e:
        print(f"Error in generate_image_from_text: {str(e)}")
        if torch.cuda.is_available():
            flux_pipe = flux_pipe.to("cpu")
        raise e

@spaces.GPU
def extract_glb(state: dict, mesh_simplify: float, texture_size: int) -> Tuple[str, str]:
    gs, mesh, trial_id = unpack_state(state)
    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
    glb_path = f"{TMP_DIR}/{trial_id}.glb"
    glb.export(glb_path)
    return glb_path, glb_path

def activate_button() -> gr.Button:
    return gr.Button(interactive=True)

def deactivate_button() -> gr.Button:
    return gr.Button(interactive=False)

css = """
footer {
    visibility: hidden;
}
"""

# Gradio 인터페이스 정의
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
    gr.Markdown("""
    # Craft3D : 3D Asset Creation & Text-to-Image Generation
    """)
    
    with gr.Tabs():
        with gr.TabItem("Image to 3D"):
            with gr.Row():
                with gr.Column():
                    image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil", height=300)
                    
                    with gr.Accordion(label="Generation Settings", open=False):
                        seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                        randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                        gr.Markdown("Stage 1: Sparse Structure Generation")
                        with gr.Row():
                            ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
                            ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
                        gr.Markdown("Stage 2: Structured Latent Generation")
                        with gr.Row():
                            slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
                            slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)

                    generate_btn = gr.Button("Generate")
                    
                    with gr.Accordion(label="GLB Extraction Settings", open=False):
                        mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
                        texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
                    
                    extract_glb_btn = gr.Button("Extract GLB", interactive=False)

                with gr.Column():
                    video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
                    model_output = LitModel3D(label="Extracted GLB", exposure=20.0, height=300)
                    download_glb = gr.DownloadButton(label="Download GLB", interactive=False)

        with gr.TabItem("Text to Image"):
            with gr.Row():
                with gr.Column():
                    text_prompt = gr.Textbox(
                        label="Text Prompt",
                        placeholder="Enter your image description...",
                        lines=3
                    )
                    
                    with gr.Row():
                        txt2img_height = gr.Slider(256, 1024, value=512, step=64, label="Height")
                        txt2img_width = gr.Slider(256, 1024, value=512, step=64, label="Width")
                    
                    with gr.Row():
                        guidance_scale = gr.Slider(1.0, 20.0, value=7.5, label="Guidance Scale")
                        num_steps = gr.Slider(1, 50, value=20, label="Number of Steps")
                    
                    generate_txt2img_btn = gr.Button("Generate Image")
                
                with gr.Column():
                    txt2img_output = gr.Image(label="Generated Image")
    
    trial_id = gr.Textbox(visible=False)
    output_buf = gr.State()


    # Example images
    with gr.Row():
        examples = gr.Examples(
            examples=[
                f'assets/example_image/{image}'
                for image in os.listdir("assets/example_image")
            ],
            inputs=[image_prompt],
            fn=preprocess_image,
            outputs=[trial_id, image_prompt],
            run_on_click=True,
            examples_per_page=32,  # 예제 수 감소
            cache_examples=False  # 예제 캐싱 비활성화는 Examples 컴포넌트에서 설정
        )
    


    # Handlers
    image_prompt.upload(
        preprocess_image,
        inputs=[image_prompt],
        outputs=[trial_id, image_prompt],
    )
    
    image_prompt.clear(
        lambda: '',
        outputs=[trial_id],
    )

    generate_btn.click(
        image_to_3d,
        inputs=[trial_id, seed, randomize_seed, ss_guidance_strength, ss_sampling_steps, 
                slat_guidance_strength, slat_sampling_steps],
        outputs=[output_buf, video_output],
        concurrency_limit=1
    ).then(
        activate_button,
        outputs=[extract_glb_btn]
    )

    extract_glb_btn.click(
        extract_glb,
        inputs=[output_buf, mesh_simplify, texture_size],
        outputs=[model_output, download_glb],
        concurrency_limit=1
    ).then(
        activate_button,
        outputs=[download_glb]
    )
    generate_txt2img_btn.click(
        generate_image_from_text,
        inputs=[text_prompt, txt2img_height, txt2img_width, guidance_scale, num_steps],
        outputs=[txt2img_output],
        concurrency_limit=1,
        show_progress=True  # 진행 상황 표시
    )


if __name__ == "__main__":
    # 메모리 정리
    free_memory()
    
    # 모델 초기화
    if not initialize_models():
        print("Failed to initialize models")
        exit(1)
    
    try:
        # rembg 사전 로드 시도 (매우 작은 이미지로)
        test_image = Image.fromarray(np.ones((32, 32, 3), dtype=np.uint8) * 255)
        pipeline.preprocess_image(test_image)
    except Exception as e:
        print(f"Warning: Failed to preload rembg: {str(e)}")
    
    # Gradio 앱 실행
    demo.queue(max_size=3).launch(
        share=True,
        max_threads=1,
        show_error=True
    )