app.py

import gradio as gr
import spaces
from gradio_litmodel3d import LitModel3D

import os
import shutil
os.environ['SPCONV_ALGO'] = 'native'
from typing import *
import torch
import numpy as np
import imageio
import uuid
import trimesh
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


MAX_SEED = np.iinfo(np.int32).max
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
os.makedirs(TMP_DIR, exist_ok=True)


def start_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    print(f'Creating user directory: {user_dir}')
    os.makedirs(user_dir, exist_ok=True)
    
def end_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    print(f'Removing user directory: {user_dir}')
    shutil.rmtree(user_dir)

def preprocess_image(image: Image.Image) -> Tuple[str, Image.Image]:
    """
    Preprocess the input image.
    """
    processed_image = pipeline.preprocess_image(image)
    return 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']

def get_seed(randomize_seed: bool, seed: int) -> int:
    """
    Get the random seed.
    """
    return np.random.randint(0, MAX_SEED) if randomize_seed else seed

@spaces.GPU
def image_to_3d(
    image: Image.Image,
    seed: int,
    ss_guidance_strength: float,
    ss_sampling_steps: int,
    slat_guidance_strength: float,
    slat_sampling_steps: int,
    req: gr.Request,
) -> Tuple[dict, str]:
    """
    Convert an image to a 3D model.
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    
    # First stage: Generate sparse structure
    outputs = pipeline.run(
        image,
        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,
        },
    )
    
    # Clear CUDA cache after structure generation
    torch.cuda.empty_cache()

    # Second stage: Generate video preview in batches
    video_frames = []
    video_geo_frames = []
    batch_size = 30  # Process 30 frames at a time
    num_frames = 120
    
    for i in range(0, num_frames, batch_size):
        end_idx = min(i + batch_size, num_frames)
        batch_frames = render_utils.render_video(
            outputs['gaussian'][0], 
            num_frames=end_idx - i,
            start_frame=i
        )['color']
        video_frames.extend(batch_frames)
        
        batch_geo = render_utils.render_video(
            outputs['mesh'][0], 
            num_frames=end_idx - i,
            start_frame=i
        )['normal']
        video_geo_frames.extend(batch_geo)
        
        # Clear cache after each batch
        torch.cuda.empty_cache()
    
    # Combine frames and save video
    video = [np.concatenate([video_frames[i], video_geo_frames[i]], axis=1) 
            for i in range(len(video_frames))]
    trial_id = str(uuid.uuid4())
    video_path = os.path.join(user_dir, f"{trial_id}.mp4")
    imageio.mimsave(video_path, video, fps=15)
    
    # Clear video data
    del video_frames
    del video_geo_frames
    del video
    torch.cuda.empty_cache()
    
    # Pack state
    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
    return state, video_path

@spaces.GPU
def extract_high_quality_mesh(
    state: dict,
    req: gr.Request,
) -> Tuple[str, str]:
    """
    Save raw mesh data directly with correct orientation.
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    
    # Get the raw mesh data from state
    vertices = state['mesh']['vertices']  # Already in numpy format from pack_state
    faces = state['mesh']['faces']
    trial_id = state['trial_id']
    
    # Rotate vertices from z-up to y-up
    rotation_matrix = np.array([[1, 0, 0], [0, 0, -1], [0, 1, 0]])
    rotated_vertices = vertices @ rotation_matrix
    
    # Create mesh and save
    simple_mesh = trimesh.Trimesh(vertices=rotated_vertices, faces=faces)
    glb_path = os.path.join(user_dir, f"{trial_id}_full.glb")
    simple_mesh.export(glb_path)
    
    return glb_path, glb_path
    
@spaces.GPU
def extract_reduced_glb(
    state: dict,
    mesh_simplify: float,
    texture_size: int,
    req: gr.Request,
) -> Tuple[str, str]:
    """
    Extract a reduced-quality GLB file with texturing.
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    gs, mesh, trial_id = unpack_state(state)
    
    # Clear cache before GLB generation
    torch.cuda.empty_cache()
    
    glb = postprocessing_utils.to_glb(
        gs,
        mesh,
        simplify=mesh_simplify,
        texture_size=texture_size,
        verbose=True
    )
    glb_path = os.path.join(user_dir, f"{trial_id}_reduced.glb")
    glb.export(glb_path)
    
    # Final cleanup
    torch.cuda.empty_cache()
    return glb_path, glb_path

with gr.Blocks(delete_cache=(600, 600)) as demo:
    gr.Markdown("""
    ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
    * Upload an image and click "Generate" to create a 3D asset
    * After generation:
        * Click "Extract Full GLB" for maximum detail mesh (untextured)
        * Or use GLB Extraction Settings for a reduced textured version
    """)
    
    with gr.Row():
        with gr.Column():
            image_prompt = gr.Image(label="Image Prompt", format="png", 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, 500, 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, 500, label="Sampling Steps", value=12, step=1)

            generate_btn = gr.Button("Generate")
            extract_full_btn = gr.Button("Extract Full GLB", interactive=False)
            
            with gr.Accordion(label="GLB Extraction Settings", open=False):
                mesh_simplify = gr.Slider(0.0, 0.98, label="Simplify", value=0.95, step=0.01)
                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
            
            extract_reduced_btn = gr.Button("Extract Reduced GLB", interactive=False)

        with gr.Column():
            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
            model_output = LitModel3D(label="3D Model Preview", exposure=20.0, height=300)
            gr.Markdown("### Download Options")
            with gr.Row():
                download_full = gr.DownloadButton(label="Download Full-Quality GLB", interactive=False)
                download_reduced = gr.DownloadButton(label="Download Reduced GLB", interactive=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=[image_prompt],
            run_on_click=True,
            examples_per_page=64,
        )

    # Event handlers
    demo.load(start_session)
    demo.unload(end_session)
    
    image_prompt.upload(
        preprocess_image,
        inputs=[image_prompt],
        outputs=[image_prompt],
    )

    generate_btn.click(
        get_seed,
        inputs=[randomize_seed, seed],
        outputs=[seed],
    ).then(
        image_to_3d,
        inputs=[image_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
        outputs=[output_buf, video_output],
    ).then(
        lambda: [gr.Button(interactive=True), gr.Button(interactive=True), gr.Button(interactive=False), gr.Button(interactive=False)],
        outputs=[extract_full_btn, extract_reduced_btn, download_full, download_reduced],
    )

    extract_full_btn.click(
        extract_high_quality_mesh,
        inputs=[output_buf],
        outputs=[model_output, download_full],
    ).then(
        lambda: gr.Button(interactive=True),
        outputs=[download_full],
    )

    extract_reduced_btn.click(
        extract_reduced_glb,
        inputs=[output_buf, mesh_simplify, texture_size],
        outputs=[model_output, download_reduced],
    ).then(
        lambda: gr.Button(interactive=True),
        outputs=[download_reduced],
    )

if __name__ == "__main__":
    pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
    pipeline.cuda()
    try:
        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))
    except:
        pass
    demo.launch()