$ show YixunLiang/UniTEX
UniTEX is a research-focused repository implementing a two-stage framework for high-fidelity 3D texture generation from arbitrary 3D shapes. Stage 1 uses LoRA-adapted Diffusion Transformers for multi-view texture synthesis, while Stage 2 employs a transformer-based Large Texturing Model (LTM) to predict Texture Functions in unified 3D space, bypassing UV mapping limitations. The repository integrates super-resolution (TSD_SR) for detail enhancement and provides tools for mesh processing, rendering, and texture reprojection. It includes training scripts, model architectures, and utility libraries, indicating its primary purpose is to support experimentation and reproduction of the CVPR 2026 paper's methodology rather than serving as a standalone end-user application.
The LLM identified 4 files worth highlighting for reuse:
Implements vertex color and UV texture optimization from multi-view images using gradient descent. Solves the recurring problem of fusing texture information from multiple camera views onto a 3D mesh. The module is self-contained, depends only on core PyTorch and numpy, and can be extracted for any texture refinement task requiring photometric consistency across views.
Provides a complete implementation of deformable convolutions (DCNv2) with optimized CUDA kernels and PyTorch wrappers. Solves the problem of adaptive receptive fields in convolutional networks, useful for tasks like video super-resolution and image restoration. The module includes forward/backward functions, handles modulated deformable convolutions, and can be reused in other projects by importing the DCNv2Pack modules.
Implements an Exponential Moving Average (EMA) callback for PyTorch Lightning to maintain moving averages of model weights during training. Solves the common problem of training instability in GANs and diffusion models by using EMA weights for evaluation. The component is framework-agnostic (only requires PyTorch Lightning), isolated, and can be dropped into any training loop with minimal modification.
Provides comprehensive bidirectional conversions between quaternions, rotation matrices, Euler angles, axis-angle, and 6D representations. Solves the ubiquitous problem of representing and converting 3D rotations in graphics and vision applications. The module is self-contained, has no external dependencies beyond numpy, and offers a clean, tested API for rotation operations that can be used in any 3D processing pipeline.