StatsMerging: Statistics-Guided Model Merging via Task-Specific Teacher Distillation

Model merging has emerged as a promising solution to accommodate multiple large models within constrained memory budgets. We present StatsMerging, a novel lightweight learning-based model merging method guided by weight distribution statistics without requiring ground truth labels or test samples. StatsMerging offers three key advantages: (1) It uniquely leverages singular values from singular value decomposition (SVD) to capture task-specific weight distributions, serving as a proxy for task importance to guide task coefficient prediction; (2) It employs a lightweight learner StatsMergeLearner to model the weight distributions of task-specific pre-trained models, improving generalization and enhancing adaptation to unseen samples; (3) It introduces Task-Specific Teacher Distillation for merging vision models with heterogeneous architectures, a merging learning paradigm that avoids costly ground-truth labels by task-specific teacher distillation. Notably, we present two types of knowledge distillation, (a) distilling knowledge from task-specific models to StatsMergeLearner; and (b) distilling knowledge from models with heterogeneous architectures prior to merging. Extensive experiments across eight tasks demonstrate the effectiveness of StatsMerging. Our results show that StatsMerging outperforms state-of-the-art techniques in terms of overall accuracy, generalization to unseen tasks, and robustness to image quality variations.

Joint Flow And Feature Refinement Using Attention For Video Restoration

Recent advancements in video restoration have focused on recovering high-quality video frames from low-quality inputs. Compared with static images, the performance of video restoration significantly depends on efficient exploitation of temporal correlations among successive video frames. The numerous techniques make use of temporal information via flow-based strategies or recurrent architectures. However, these methods often encounter difficulties in preserving temporal consistency as they utilize degraded input video frames. To resolve this issue, we propose a novel video restoration framework named Joint Flow and Feature Refinement using Attention (JFFRA). The proposed JFFRA is based on key philosophy of iteratively enhancing data through the synergistic collaboration of flow (alignment) and restoration. By leveraging previously enhanced features to refine flow and vice versa, JFFRA enables efficient feature enhancement using temporal information. This interplay between flow and restoration is executed at multiple scales, reducing the dependence on precise flow estimation. Moreover, we incorporate an occlusion-aware temporal loss function to enhance the network's capability in eliminating flickering artifacts. Comprehensive experiments validate the versatility of JFFRA across various restoration tasks such as denoising, deblurring, and super-resolution. Our method demonstrates a remarkable performance improvement of up to 1.62 dB compared to state-of-the-art approaches.