app.py

import gradio as gr
import spaces
from gradio_litmodel3d import LitModel3D
import os
import time
from os import path
import shutil
from datetime import datetime
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
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 diffusers import FluxPipeline
from transformers import pipeline

# Hugging Face 토큰 설정
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN is None:
    raise ValueError("HF_TOKEN environment variable is not set")

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

# Setup and initialization code
cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
PERSISTENT_DIR = os.environ.get("PERSISTENT_DIR", ".")
gallery_path = path.join(PERSISTENT_DIR, "gallery")

os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path
os.environ['SPCONV_ALGO'] = 'native'

torch.backends.cuda.matmul.allow_tf32 = True

# 번역기 초기화
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")

class timer:
    def __init__(self, method_name="timed process"):
        self.method = method_name
    def __enter__(self):
        self.start = time.time()
        print(f"{self.method} starts")
    def __exit__(self, exc_type, exc_val, exc_tb):
        end = time.time()
        print(f"{self.method} took {str(round(end - self.start, 2))}s")

def preprocess_image(image: Image.Image) -> Tuple[str, Image.Image]:
    trial_id = str(uuid.uuid4())
    processed_image = pipeline.preprocess_image(image)
    processed_image.save(f"{TMP_DIR}/{trial_id}.png")
    return trial_id, processed_image

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) -> Tuple[dict, str]:
    if randomize_seed:
        seed = np.random.randint(0, MAX_SEED)
    outputs = pipeline.run(
        Image.open(f"{TMP_DIR}/{trial_id}.png"),
        seed=seed,
        formats=["gaussian", "mesh"],
        preprocess_image=False,
        sparse_structure_sampler_params={
            "steps": ss_sampling_steps,
            "cfg_strength": ss_guidance_strength,
        },
        slat_sampler_params={
            "steps": slat_sampling_steps,
            "cfg_strength": slat_guidance_strength,
        },
    )
    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
    trial_id = 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)
    return state, video_path


@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)

@spaces.GPU
def text_to_image(prompt: str, height: int, width: int, steps: int, scales: float, seed: int) -> Image.Image:
    # 한글 감지 및 번역
    def contains_korean(text):
        return any(ord('가') <= ord(c) <= ord('힣') for c in text)
    
    # 프롬프트 전처리
    if contains_korean(prompt):
        translated = translator(prompt)[0]['translation_text']
        prompt = translated
    
    # 프롬프트 형식 강제
    formatted_prompt = f"wbgmsst, 3D, {prompt}, white background"
    
    with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
        try:
            generated_image = pipe(
                prompt=[formatted_prompt],
                generator=torch.Generator().manual_seed(int(seed)),
                num_inference_steps=int(steps),
                guidance_scale=float(scales),
                height=int(height),
                width=int(width),
                max_sequence_length=256
            ).images[0]
            
            trial_id = str(uuid.uuid4())
            generated_image.save(f"{TMP_DIR}/{trial_id}.png")
            return generated_image
            
        except Exception as e:
            print(f"Error in image generation: {str(e)}")
            return None

# Gradio Interface
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("""## Craft3D""")
    
    with gr.Row():
        with gr.Column():
            text_prompt = gr.Textbox(
                label="Text Prompt",
                placeholder="Describe what you want to create...",
                lines=3
            )
            
            with gr.Accordion("Image Generation Settings", open=False):
                with gr.Row():
                    height = gr.Slider(
                        label="Height",
                        minimum=256,
                        maximum=1152,
                        step=64,
                        value=1024
                    )
                    width = gr.Slider(
                        label="Width",
                        minimum=256,
                        maximum=1152,
                        step=64,
                        value=1024
                    )
                
                with gr.Row():
                    steps = gr.Slider(
                        label="Inference Steps",
                        minimum=6,
                        maximum=25,
                        step=1,
                        value=8
                    )
                    scales = gr.Slider(
                        label="Guidance Scale",
                        minimum=0.0,
                        maximum=5.0,
                        step=0.1,
                        value=3.5
                    )
                
                seed = gr.Number(
                    label="Seed",
                    value=lambda: torch.randint(0, MAX_SEED, (1,)).item(),
                    precision=0
                )
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
            
            generate_image_btn = gr.Button("Generate Image")
            
            image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil", height=300)
            
            with gr.Accordion("3D Generation Settings", open=False):
                ss_guidance_strength = gr.Slider(0.0, 10.0, label="Structure Guidance Strength", value=7.5, step=0.1)
                ss_sampling_steps = gr.Slider(1, 50, label="Structure Sampling Steps", value=12, step=1)
                slat_guidance_strength = gr.Slider(0.0, 10.0, label="Latent Guidance Strength", value=3.0, step=0.1)
                slat_sampling_steps = gr.Slider(1, 50, label="Latent Sampling Steps", value=12, step=1)

            generate_3d_btn = gr.Button("Generate 3D")
            
            with gr.Accordion("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)
            
    trial_id = gr.Textbox(visible=False)
    output_buf = gr.State()

    # Handlers
    generate_image_btn.click(
        text_to_image,
        inputs=[text_prompt, height, width, steps, scales, seed],
        outputs=[image_prompt]
    ).then(
        preprocess_image,
        inputs=[image_prompt],
        outputs=[trial_id, image_prompt]
    )

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

    generate_3d_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],
    ).then(
        activate_button,
        outputs=[extract_glb_btn],
    )

    video_output.clear(
        deactivate_button,
        outputs=[extract_glb_btn],
    )

    extract_glb_btn.click(
        extract_glb,
        inputs=[output_buf, mesh_simplify, texture_size],
        outputs=[model_output, download_glb],
    ).then(
        activate_button,
        outputs=[download_glb],
    )

    model_output.clear(
        deactivate_button,
        outputs=[download_glb],
    )

if __name__ == "__main__":
    # 3D 생성 파이프라인
    pipeline = TrellisImageTo3DPipeline.from_pretrained(
        "JeffreyXiang/TRELLIS-image-large",
        use_auth_token=HF_TOKEN
    )
    pipeline.cuda()
    
    # 이미지 생성 파이프라인
    pipe = FluxPipeline.from_pretrained(
        "black-forest-labs/FLUX.1-dev",
        torch_dtype=torch.bfloat16,
        use_auth_token=HF_TOKEN
    )
    
    # Hyper-SD LoRA 로드
    pipe.load_lora_weights(
        hf_hub_download(
            "ByteDance/Hyper-SD",
            "Hyper-FLUX.1-dev-8steps-lora.safetensors",
            use_auth_token=HF_TOKEN
        )
    )
    pipe.fuse_lora(lora_scale=0.125)
    pipe.to(device="cuda", dtype=torch.bfloat16)
    
    try:
        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))
    except:
        pass
        
    demo.launch(allowed_paths=[PERSISTENT_DIR])