Source-linked AI summary
NTIRE 2026 Challenge on Bitstream-Corrupted Video Restoration: Methods and Results
Wenbin Zou, Tianyi Liu, Kejun Wu, Huiping Zhuang, Zongwei Wu, Zhuyun Zhou, Radu Timofte, Kim-Hui Yap, Lap-Pui Chau, Yi Wang, Shiqi Zhou, Xiaodi Shi, Yuxiang Chen, Yilian Zhong, Shibo Yin, Yushun Fang, Xilei Zhu, Yahui Wang, Chen Lu, Zhitao Wang, Lifa Ha, Hengyu Man, Xiaopeng Fan, Priyansh Singh, Sidharth, Krrish Dev, Soham Kakkar, Vinit Jakhetiya, Ovais Iqbal Shah, Wei Zhou, Linfeng Li, Qi Xu, Zhenyang Liu, Kepeng Xu, Tong Qiao, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi
TL;DR
Bitstream corruption produces severe, irregular video degradation that existing restoration assumptions do not fully capture. The paper organizes the NTIRE 2026 BSCVR challenge as a common benchmark with datasets, evaluation protocols, participating methods, and final analyses. MGTV-AI led objective metrics, while RedMediaTech achieved the best LPIPS, and the comparison highlighted remaining difficulty in semantic details and temporal stability.
Problem
Bitstream-corrupted video recovery remains difficult because realistic corruption is irregular, codec-dependent, and insufficiently represented by conventional restoration or simplified concealment assumptions.
Method
The paper reports a challenge benchmark comprising datasets, evaluation protocols, participating restoration methods, and analyses of final results and technical trends.
Results
MGTV-AI achieved the highest objective restoration performance, while RedMediaTech obtained the best overall LPIPS score of 0.0852.
Takeaways & Limitations
The challenge indicates that combining spatiotemporal attention, generative models, and foundation-model priors is a promising direction for bitstream-corrupted restoration.
Takeaways & Limitations
Restoration remains challenged by semantic details such as faces and text and by temporal coherence in large-motion scenes.
Abstract
from arXiv · showhide
This paper reports on the NTIRE 2026 Challenge on Bitstream-Corrupted Video Restoration (BSCVR). The challenge aims to advance research on recovering visually coherent videos from corrupted bitstreams, whose decoding often produces severe spatial-temporal artifacts and content distortion. Built upon recent progress in bitstream-corrupted video recovery, the challenge provides a common benchmark for evaluating restoration methods under realistic corruption settings. We describe the dataset, evaluation protocol, and participating methods, and summarize the final results and main technical trends. The challenge highlights the difficulty of this emerging task and provides useful insights for future research on robust video restoration under practical bitstream corruption.
1. Introduction
Bitstream-corrupted video recovery addresses severe, irregular spatial-temporal degradation that conventional restoration and simplified concealment assumptions do not adequately represent. The challenge establishes a common benchmark to evaluate methods under realistic corruption and summarize emerging technical trends.
- Motivation: Bitstream corruption can cause severe spatial-temporal degradation, content distortion, and temporal inconsistency after decoding.These failures affect reliability in surveillance, streaming, communication, and visual analytics.
- Motivation: Unlike conventional restoration, bitstream-corrupted recovery faces irregular, non-stationary degradation shaped by codec behavior and inter-frame prediction.Traditional methods typically assume stable degradation priors, while inpainting and concealment methods often use simplified masks.
- Research Context: Recent benchmark and recovery studies established bitstream-corrupted video recovery as a dedicated research problem with realistic corruption distinct from manually simulated masks.Subsequent work explored enhanced recovery and blind settings.
- Challenge Aim: The BSCVR challenge provides a common testbed for reconstruction fidelity, perceptual quality, and robustness under diverse realistic corruption patterns.It brings together restoration, inpainting, generative enhancement, and codec-aware recovery researchers.
- Report Scope: The report presents the challenge setup, datasets, evaluation protocol, participating methods, final results, and observed technical trends.The challenge is positioned as a way to advance methods and understand the evolving research direction.
2. Related Works
Related work shows that conventional restoration, error concealment, and inpainting rely on relatively stable or handcrafted corruption assumptions. Dedicated bitstream-corrupted recovery benchmarks and methods address more realistic artifacts, but robust generalization remains unresolved.
- Traditional Methods: Conventional video restoration models degradations such as blur, low resolution, noise, and compression artifacts with relatively stable priors.Representative methods use temporal alignment, feature fusion, and transformer-based modeling.
- Traditional Methods: Error concealment and inpainting methods commonly estimate missing content through interpolation, motion compensation, hybrid strategies, or simulated corruption patterns.These approaches generally rely on handcrafted assumptions about missing regions.
- Dedicated Recovery: The BSCV benchmark introduced a dedicated task and recovery framework using residual visual cues within corrupted regions plus neighboring spatial-temporal context.It showed that realistic bitstream corruption differs fundamentally from conventional simulated patterns.
- Dedicated Recovery: Later studies explored visual foundation-model guidance and diffusion priors for corruption localization, feature-aware completion, and more practical recovery systems.These works demonstrate the feasibility of incorporating external priors and knowledge.
- Open Challenges: Bitstream-corrupted restoration remains difficult because methods must handle diverse artifacts, practical conditions, and generalization across corrupted contents.A standard challenge benchmark can support systematic evaluation and comparison of technical designs.
3. NTIRE 2026 Challenge
The NTIRE 2026 BSCVR challenge supplies datasets, evaluation resources, and a test framework for realistic bitstream-corrupted video restoration. Final results show distinct strengths in objective fidelity and perceptual quality, alongside broad adoption of foundation models and parameter-efficient adaptation.
- Challenge Setup: The challenge introduces datasets, a toolbox, two evaluation phases, and common trends across submitted solutions.The development phase uses released validation pairs, while the test phase evaluates submissions on corrupted videos.
- Datasets: 3,471 HD bitstream-corrupted training videos include corrupted sequences, ground-truth sequences, and per-frame binary masks.The videos contain decoding artifacts with non-uniform spatio-temporal distributions.
- Evaluation: Seven teams participated in the final test phase after 153 registrations, submitting results, codes, and factsheets.The final ranking uses average PSNR on the hidden test set.
- Results: 33.6423 dB overall PSNR and 0.9334 overall SSIM were achieved by MGTV-AI, while RedMediaTech obtained the best overall LPIPS of 0.0852.These results separate the strongest objective restoration scores from the strongest perceptual-quality score.
- Results: Visual comparisons found MGTV-AI and RedMediaTech strongest overall, while other methods often removed blockiness but retained softness or lost fine textures.Semantic details and temporal stability remain difficult when original bitstream information is almost entirely lost.
- Technical Trends: Three of six submitted solutions used B2SCVR, and half used PEFT methods such as LoRA or MoE-LoRA with external semantic or structural priors.The reported trend combines visual foundation models with parameter-efficient adaptation.
4. Challenge Teams and Methods
Participating teams use diverse restoration designs, ranging from multi-stage and diffusion-based systems to semantic-prior, boundary-refinement, and mask-guided pipelines. Common strategies include staged optimization, foundation-model priors, and explicit handling of corrupted regions and boundaries.
- MGTV-AI: MGTV-AI uses a three-stage framework that progressively completes masked regions, processes later-stage outputs, and fuses predictions with the input mask.The method trains separate stages with different resolutions, losses, and optimization schedules, then uses later-stage fusion for the final result.
- RedMediaTech: RedMediaTech builds a single-step Wan2.1 DiT restoration model, replacing its original VAE with Qwen-Image VAE for large motion and complex temporal variations.Its two-stage training balances perceptual quality using MSE and LPIPS before fine-tuning with MSE for PSNR and SSIM.
- bighit: bighit combines a semantic memory bank with router-guided MoE-LoRA experts to retrieve reliable context and adapt restoration to heterogeneous corruption patterns.A boundary refinement head predicts residuals and blending weights, while an optional second stage suppresses residual artifacts and improves boundary consistency.
- weichow: weichow adapts B2SCVR with mask-guided multi-resolution compositing that preserves uncorrupted content while reconstructing damaged regions.The pipeline fine-tunes on challenge data and uses masks to combine restored corrupted regions with original pixels.
- holding: holding modifies B2SCVR with mask-aware suppression, target-centric cross-frame attention, and boundary-aware seam refinement in a single-model pipeline.These modules target corruption leakage, unstable temporal aggregation, and visible transitions near mask boundaries.
5. Conclusion
The challenge establishes standardized evaluation for severe bitstream-corruption artifacts while revealing a split between objective reconstruction and perceptual realism. Current systems still struggle with fine-detail recovery and temporal coherence in extreme damage.
- MGTV-AI leads PSNR and SSIM using a three-stage progressive repair framework for structural reconstruction.The result is attributed to multi-stage fusion.
- RedMediaTech achieves the best LPIPS score by combining Wan2.1 diffusion with Qwen Image VAE in a two-stage training strategy.The approach generates realistic textures and details in severely damaged segments.
- Visual foundation models with parameter-efficient fine-tuning have become mainstream, with SAM2 providing semantic priors and LoRA improving generalization while controlling computational costs.
- Extreme corruption still causes edge softening, loss of fine features such as faces and text, and difficulty maintaining temporal coherence during large motion.
- These limitations motivate more robust and deployable restoration systems that combine spatiotemporal attention with generative models.
RedMediaTech
RedMediaTech is listed as an affiliation entry associated with Xiaohongshu INC.
- RedMediaTech is associated with Xiaohongshu INC.
bighit
The section identifies a two-stage bitstream-corrupted video restoration method using semantic memory and a mixture of LoRA experts, with authors affiliated with the University of Illinois Urbana-Champaign.
- The method is titled Two-Stage Bitstream-Corrupted Video Restoration via Semantic Memory and Mixture of LoRA Experts.
- The listed authors are affiliated with the University of Illinois Urbana-Champaign.
Vroom
The section lists an affiliation with the Indian Institute of Technology Jammu.
- The listed affiliation is the Indian Institute of Technology Jammu.
weichow
The listed affiliations include the National University of Singapore, Shanghai Jiao Tong University, and Xidian University.
- The affiliations include the National University of Singapore and Shanghai Jiao Tong University.
- Tong Qiao is affiliated with Xidian University.