🎨 C3-TRELLIS-Gradio

A comprehensive comparison and implementation of different TRELLIS model variants for 3D asset generation from images. This project analyzes and consolidates multiple TRELLIS implementations to create an optimized Gradio interface.

🎯 Project Overview

This project examines four different TRELLIS implementations:

• 🔧 FurkanGozukara-TRELLIS: Extended implementation with API and additional features
• ⚡ TRELLIS-Imagen3D: Simplified, working implementation by cavargas10
• 🎯 TRELLIS-innoai: Multi-image support with both single and multi-image tabs
• 🌐 microsoft-TRELLIS: The original official implementation

The main contribution is creating a unified app.py for TRELLIS-innoai that combines the best features from all implementations while removing unnecessary dependencies.

🚀 Key Features

✨ Unified Implementation

• 🎭 Dual-mode interface: Single image and multi-image 3D generation in one app
• 🔧 No LitModel3D dependency: Uses standard Gradio components for better compatibility
• ✂️ Split image functionality: Automatically splits concatenated multi-view images
• ✨ Clean UI: Based on the working TRELLIS-Imagen3D interface design

🖼️ Multi-Image Support

• 📷 Upload multiple views of the same object for improved 3D reconstruction
• 🎲 Supports both stochastic and multidiffusion algorithms
• 🧪 Experimental feature that works best with consistent object views

📦 Export Options

• 🎮 GLB files: Industry-standard 3D format with texture
• ✨ Gaussian splats: PLY format for 3D Gaussian representation
• 🏛️ Adjustable mesh simplification and texture resolution

📋 Implementation Details

🔄 Key Changes Made

1. 🚫 Removed gradio_litmodel3d dependency: Replaced with standard gr.Model3D
2. ✅ Preserved split_image functionality: Essential for processing multi-view images
3. 🔗 Unified configuration: Server runs on 0.0.0.0:7860 like Imagen3D
4. 🧤 Simplified codebase: Removed unnecessary docstrings while maintaining functionality

⚙️ Technical Architecture

# Core pipeline configuration
pipeline = TrellisImageTo3DPipeline.from_pretrained("jetx/trellis-image-large")
pipeline.cuda()

# Two-stage generation process
# Stage 1: Sparse Structure Generation (guidance: 7.5, steps: 12)
# Stage 2: Detail Enhancement (guidance: 3.0, steps: 12)

🔧 Installation

1. Clone the repository:

git clone https://github.com/yourusername/c3-trellis-gradio.git
cd c3-trellis-gradio

2. Install dependencies:

pip install -r requirements.txt

3. Run the application:

python app.py

The application will be available at http://localhost:7860

🏆 Honorable Mentions

This project builds upon and references several TRELLIS implementations:

🏛️ Original Implementation

• microsoft/TRELLIS: The official TRELLIS repository with full training code and documentation
• Paper: Structured 3D Latents for Scalable and Versatile 3D Generation

🤗 Hugging Face Spaces

• cavargas10/TRELLIS-Imagen3D: Clean, working implementation that served as the base for our unified approach
• innoai/TRELLIS: Implementation with multi-image support that inspired our dual-mode interface

🧠 Model Weights

• jetx/trellis-image-large: Mirror of the TRELLIS model weights on Hugging Face

💡 Usage Example

import os
os.environ['SPCONV_ALGO'] = 'native'

from PIL import Image
from trellis.pipelines import TrellisImageTo3DPipeline
from trellis.utils import render_utils, postprocessing_utils

# Load pipeline
pipeline = TrellisImageTo3DPipeline.from_pretrained("jetx/trellis-image-large")
pipeline.cuda()

# Single image generation
image = Image.open("path/to/image.png")
outputs = pipeline.run(image, seed=1)

# Extract outputs
glb = postprocessing_utils.to_glb(
    outputs['gaussian'][0],
    outputs['mesh'][0],
    simplify=0.95,
    texture_size=1024
)
glb.export("output.glb")

🛠️ Project Structure

c3-trellis-gradio/
├── TRELLIS-innoai/          # Unified implementation
│   ├── app.py               # Main Gradio interface
│   ├── requirements.txt     # Dependencies (no litmodel3d)
│   └── assets/              # Example images
├── TRELLIS-Imagen3D/        # Reference implementation
├── FurkanGozukara-TRELLIS/  # Extended features
└── microsoft-TRELLIS/       # Original implementation

📄 License

This project is licensed under the MIT License, following the original TRELLIS licensing.

🙏 Acknowledgments

Special thanks to:

• 💙 The Microsoft Research team for creating TRELLIS
• ⭐ cavargas10 for the clean Imagen3D implementation
• 🌟 innoai for the multi-image support implementation
• 🤗 The Hugging Face community for hosting the demos and models

📜 Citation

If you use this project, please cite the original TRELLIS paper:

@article{xiang2024structured,
    title   = {Structured 3D Latents for Scalable and Versatile 3D Generation},
    author  = {Xiang, Jianfeng and Lv, Zelong and Xu, Sicheng and Deng, Yu and Wang, Ruicheng and Zhang, Bowen and Chen, Dong and Tong, Xin and Yang, Jiaolong},
    journal = {arXiv preprint arXiv:2412.01506},
    year    = {2024}
}