Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2607.22772

Generative Video Compression with Adaptive Score Distillation

Naifu Xue·Zhaoyang Jia·Haosen Li·Zihan Zheng·Jiahao Li·Bin Li·Xiaoyi Zhang·Qi Meng·Yuan Zhang·Yan Lu
View original source ↗Read source document ↗Export metadata ↓

Abstract

Diffusion models provide strong generative capabilities for video compression at ultra-low bitrates. Existing diffusion-based video codecs adapt base models originally developed for text-conditioned generation, whereas diffusion models designed and trained specifically for compression remain unexplored. To fill this gap, we introduce our Generative Video Codec (GenVC), built on a video diffusion model trained from scratch for compression. To our knowledge, this is the first compression-oriented video diffusion model. We realize this model directly in pixel space with a global-to-local hierarchy that recovers fine spatio-temporal details, enabling high-quality generative reconstruction from compressed representations. To accelerate inference, we distill the multi-step model into one step using distribution matching distillation (DMD). Applying DMD directly, however, drives the student toward motion-stalled reconstructions. We trace this to a teacher-side guidance failure: once student-induced perturbations leave the frozen teacher's training region, its guidance can become misleading, causing DMD updates to reinforce rather than correct the student drift. To break the resulting feedback loop, we propose Adaptive Score Distillation, which gates DMD updates according to their alignment with the ground-truth direction, enabling high-quality reconstruction with coherent motion. Experimental results show that GenVC achieves state-of-the-art perceptual quality at ultra-low bitrates, with average bitrate savings of 62.5% at matched LPIPS and 71.3% at matched FID over GLVC. Unlike prior codecs that inherit billion-scale pretrained backbones, our diffusion model has only 478.0M parameters and decodes 1080p video in a single step at 15.1 fps on an A100 GPU.

Explore related subjects

eess.IVcs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Naifu Xue, Zhaoyang Jia, Haosen Li, Zihan Zheng, Jiahao Li, Bin Li, Xiaoyi Zhang, Qi Meng, Yuan Zhang, Yan Lu. 2026-07-24. Generative Video Compression with Adaptive Score Distillation. https://arxiv.org/abs/2607.22772

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

KELP: K-space-conditioned Estimation of Learned Sampling Patterns for Scan-Adaptive Multi-Coil MRI

Deep learning techniques have gained considerable attention for accelerating MRI acquisition while maintaining image quality. In this work, we present a convolutional neural network (CNN)-based framework for predicting scan-adaptive undersampling patterns directly from low-frequency multi-coil $k$-space data. Unlike approaches that optimize sampling patterns during training or rely on nearest-neighbor search at inference, our method is trained using precomputed scan-adaptive optimized masks as supervised labels and predicts a scan-specific sampling pattern in a single forward pass. The training procedure alternates between optimizing the sampling network and a reconstruction network, allowing the learned masks to adapt to reconstruction performance. Experiments on the fastMRI multi-coil knee dataset demonstrate competitive reconstruction quality compared with existing population- and scan-adaptive sampling approaches at $4\times$ and $8\times$ acceleration factors. In addition, the proposed method enables efficient scan-specific mask prediction with inference cost independent of the training dictionary size.

eess.IV↗
arXiv2026

LaminoDiff: Generative Computed Laminography via Near-Isotropic Spectral Supervision and Anisotropic Geometry

Computed Laminography (CL) is widely used for nondestructive inspection of extended planar objects, but its restricted angular coverage produces an anisotropic point spread function, missing-cone spectral incompleteness, and severe aliasing with interlayer leakage. This paper presents LaminoDiff, a physics-constrained diffusion framework for CL reconstruction. During training, a CT-derived near-isotropic supervision target is reconstructed from full-angle Computed Tomography (CT) projections physically degraded to match CL detector noise and focal-spot blur while retaining full angular coverage; it is withheld from inference. At inference, the reverse process uses only the CL observation. An Anisotropic Representation (AR) constructs depth-aware channels from three adjacent slices: a neighbor average, a center-neighbor residual, and the retained current slice, followed by directional in-plane feature extraction. Experiments on simulated and real multilayer printed circuit board data, including ball grid array and high-frequency stub samples, show that LaminoDiff improves artifact suppression, edge preservation, and depth stratification over analytic Feldkamp--Davis--Kress and representative learning-based baselines.

eess.IV↗
arXiv2026

IViT: A Novel Interpretable Visual Transformer for Skin Disease Detection

The clinical diagnosis of skin diseases is susceptible to interference from inter-class similarity of skin lesions, and over-reliance on clinicians'experience easily leads to subjective bias. Although existing deep learning aided diagnosis methods achieve competitive accuracy, they suffer from the black-box opacity of Vision Transformer (ViT) and poor adaptability to medical few-shot scenarios. Moreover, mainstream explainable algorithms generally face the bottleneck of significant accuracy degradation when improving interpretability. This paper proposes an interpretable ViT (IViT) constrained by Quadratic Programming (QP). The introduced pre-trained transfer learning adapts to few-shot feature extraction. A discrete QP feature selection framework is constructed to screen generic and discriminative features consistent with clinical diagnostic logic. A multi-objective loss function is designed to reduce feature redundancy and optimize activation distribution while preserving classification performance. Experimental results on six standard skin disease datasets show that IViT achieves an accuracy of 93.80%, only 0.21% lower than the baseline, with feature redundancy reduced by 29.5%. Its core activation regions are consistent with clinically concerned lesion areas. The proposed model balances accuracy and interpretability, providing a reliable solution for the clinical deployment of few-shot intelligent skin disease diagnosis.

eess.IV↗