davidelobba/TEMU-VTOFF

Based on the README, here's the technical summary: Implements a dual-Diffusion Transformer (dual-DiT) architecture that combines pretrained feature extraction with text-enhanced generation, using a multimodal hybrid attention mechanism to integrate garment descriptions with person features for synthesizing occluded regions. A lightweight DINOv2-based garment aligner module conditions generation on target in-shop images rather than traditional denoising objectives. Supports multi-category garment handling (upper/lower/full-body) across Dress Code and VITON-HD datasets, with pre-extracted features from CLIP, OpenCLIP, and T5 encoders, and requires Stable Diffusion 3 Medium via HuggingFace.