OrthoTryOn: Geometric Orthogonalization for Conflict-Free Unified Fashion Generation

A research team has introduced OrthoTryOn, a unified framework designed to eliminate interference when a single model performs both virtual try-on and garment reconstruction, according to a preprint posted to arXiv on 26 Jun 2026 [1][2]. The framework addresses a persistent problem in multi-task learning known as negative transfer, where sharing parameters across semantically distinct tasks causes severe gradient conflict and degrades performance [2]. OrthoTryOn mitigates this within a shared Low-Rank Adaptation, or LoRA, module [1][2]. Its first mechanism, Orthogonal Subspace Projection (OSP), applies task-specific orthogonal rotations to bottleneck features, mapping them into decorrelated coordinate frames so the tasks no longer compete for the same representational space [2]. To handle interference that persists at inference time, the authors propose Fisher-guided Negative Guidance (FNG), a parameter-free strategy that uses diagonal Fisher information to measure inter-task sensitivity overlap and explicitly repels generation trajectories from the most confusable task via Classifier-Free Guidance [2]. The paper reports that OrthoTryOn not only avoids the performance collapse typical of naive unified training but surpasses independently trained task-specific models, achieving state-of-the-art results across multiple benchmarks while generalizing across diverse diffusion backbones [1][2]. Code for the project has been made available on GitHub through the NJU-PCALab repository [2]. The preprint was submitted to arXiv, the open-access repository that hosts e-prints across physics, computer science, and related fields and which, as of late 2024, receives roughly 24,000 new articles per month [6]. The work appears within the Computer Vision and Pattern Recognition section of the repository [1].