Image-to-3D-Video-Asset-Generator

Image-to-3D-Video-Asset-Generator is an all-in-one generative 3D pipeline that transitions smoothly from textual concepts or reference images into fully realized 3D mesh assets (.glb), dynamic camera movements in 5-second MP4 videos, and clean bundle exports (.zip).

Powered by TRELLIS.2-4B for high-fidelity 3D structural reconstruction and Z-Image-Turbo for rapid initial concept generation, the application includes custom PBR baking, mesh decimation, Rerun 3D visualization, and automated offscreen video rendering via PyRender (with a Matplotlib fallback).

video_2026-07-30T151445.mp4

Key Features

• Cascaded Generative Pipeline: Unifies Text-to-Image (Z-Image-Turbo), Image-to-3D (TRELLIS.2-4B), 3D-to-Video camera moves, and single-click ZIP asset bundling.
• Advanced 3D Mesh Processing: Converts raw latents into decimation-targeted .glb models with custom PBR texture layouts and flex-GEMM autotuned acceleration.
• Interactive Rerun 3D Viewer: Renders spatial 3D views, bounding boxes, and orientation vectors directly within the web workspace.
• Customizable Camera Video Rendering: Offers 7 camera movement styles (Orbit, Zoom In, Zoom Out, Turntable, Spiral, Top Sweep, and Cinematic) with custom background and polygon material Overrides.
• Session-Isolated Asset Export: Packages the final background-removed image, GLB mesh, MP4 preview, and metadata into an organized ZIP archive.

---

Repository Structure

├── assets/
│   ├── app/
│   └── example_image/
├── example-images/
├── trellis2/
├── app.py
├── autotune_cache.json
├── LICENSE.txt
├── packages.txt
├── pre-requirements.txt
├── README.md
└── requirements.txt

Installation and Requirements

To configure the Image-to-3D-Video-Asset-Generator suite locally, set up a Python 3.12 environment with the exact CUDA 13.0 compiled dependencies listed below. A modern, high-performance CUDA-enabled GPU is strictly required to execute the models and compiled C++ extensions.

• Python Version: Python 3.12 is strictly required and recommended.
• PyTorch Version: torch==2.11.0 or above is required for compatibility with pre-built wheel binaries.
• CUDA Version: CUDA 13.0 is required (--extra-index-url https://download.pytorch.org/whl/cu130), matching the live Hugging Face Space environment.

Standard PIP Installation

1. Install System Packages Install the necessary EGL and X11 rendering libraries specified in packages.txt:

sudo apt-get update && sudo apt-get install -y \
    libegl1 \
    libgl1 \
    libglib2.0-0 \
    libxkbcommon0 \
    libdbus-1-3

2. Upgrade Package Manager Ensure your local package manager is updated:

pip install "pip>=23.0.0"

3. Install Core Dependencies Install PyTorch 2.11.0 + CUDA 13.0 wheels, custom compiled RTX Pro 6000 binaries (cumesh, flex_gemm, nvdiffrast, nvdiffrec_render, o_voxel), and secondary libraries from requirements.txt:

pip install -r requirements.txt

Core Requirements (requirements.txt)

--extra-index-url https://download.pytorch.org/whl/cu130
-f https://whl.natten.org

torch==2.11.0
torchvision==0.26.0
triton==3.6.0

flash_attn @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/flash_attention_2/flash_attn-2.8.3%2Bcu13torch2.11cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
cumesh @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/rtx_pro_6000_prebuild_wheels/cumesh-0.0.1%2Btorch2.11.0.cu130-cp312-cp312-linux_x86_64.whl
flex_gemm @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/rtx_pro_6000_prebuild_wheels/flex_gemm-1.0.0%2Btorch2.11.0.cu130-cp312-cp312-linux_x86_64.whl
nvdiffrast @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/rtx_pro_6000_prebuild_wheels/nvdiffrast-0.4.0%2Btorch2.11.0.cu130-cp312-cp312-linux_x86_64.whl
nvdiffrec_render @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/rtx_pro_6000_prebuild_wheels/nvdiffrec_render-0.0.0%2Btorch2.11.0.cu130-cp312-cp312-linux_x86_64.whl
o_voxel @ https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator/resolve/main/rtx_pro_6000_prebuild_wheels/o_voxel-0.0.1%2Btorch2.11.0.cu130-cp312-cp312-linux_x86_64.whl

natten==0.21.6+torch2110cu130

pillow==12.0.0
imageio==2.37.2
imageio-ffmpeg==0.6.0
opencv-python-headless==4.12.0.88
trimesh==4.10.1
plyfile==1.1.3
pyrender

transformers==4.57.3
diffusers==0.37.1
accelerate==1.13.0
huggingface-hub
spaces
gradio==6.14.0
gradio_rerun

timm==1.0.22
kornia==0.8.2
einops==0.8.2
jaxtyping
monopriors

scipy
jax
zstandard==0.25.0
tqdm==4.67.1
easydict==1.13
omegaconf
termcolor
pyserde
braceexpand
rerun-sdk

git+https://github.com/microsoft/MoGe.git

---

Usage

Once dependencies are verified, launch the application by running the main module script:

python app.py

The script will configure offscreen EGL bindings, autotune GEMM kernels, and load the pipeline weights before starting a local Gradio interface (typically at http://127.0.0.1:7860/).

1. Step 1 — Concept Input: Use the Text-to-Image-3D tab to generate a new 2D base image from a prompt (e.g., "A Plane 3D model"), or switch to Image-to-3D and upload a custom image file directly.
2. Step 2 — 3D Reconstruction: Configure resolution ($512$, $1024$, or $1536$), target face decimation count, and texture size parameters, then click 2. Generate 3D. Inspect the generated GLB model using the embedded Rerun 3D viewer.
3. Step 3 — Video Rendering: Select camera movement style, rotation speed, background color, and output resolution, then click 4. Generate Video (5 sec).
4. Step 4 — Asset Packaging: Click 6. Generate ZIP File to compress the preprocessed input image, .glb mesh, .mp4 camera path, and metadata into a single downloadable package.

Links and Source

• GitHub Repository: https://github.com/PRITHIVSAKTHIUR/Image-to-3D-Video-Asset-Generator.git
• Hugging Face Live Demo: https://huggingface.co/spaces/prithivMLmods/Image-to-3D-Video-Asset-Generator
• License: Apache License 2.0