NTIRE 2026 Challenge on Bitstream-Corrupted Video Restoration: Methods and Results

• Bitstream corruption during transmission, storage, or decoding can cause severe spatial-temporal artifacts, content distortion, and temporal inconsistency that affect video applications.The paper identifies impacts on user experience and the reliability of surveillance, streaming, communication, and visual analytics.Source: §1 · p. 1
• Unlike denoising, deblurring, or compression artifact reduction, bitstream corruption is irregular, codec-dependent, and can propagate across frames, challenging stable degradation priors and simplified masks.Corrupted regions may contain mixed residual information and complex artifact patterns.Source: §1 · p. 1
• Prior work established bitstream-corrupted video recovery as a dedicated problem, introducing the BSCV benchmark and baseline [32] and exploring enhanced recovery and blind settings.The passage notes that corruption decoded from bitstreams differs substantially from manually simulated masks and that recovery remains challenging in diversity, quality, and deployment.Source: §1 · p. 1
• The NTIRE 2026 BSCVR Challenge provides a common testbed to benchmark reconstruction fidelity, perceptual quality, and robustness under diverse realistic corruption patterns.It brings together researchers in video restoration, inpainting, generative enhancement, and codec-aware recovery to advance methods and understand technical trends.Source: §1 · pp. 1–2
• The report presents the challenge setup, datasets, and evaluation protocol, then summarizes participating methods and analyzes final results and technical trends.Source: §1 · p. 2

2.1. Traditional Methods

• Conventional restoration methods target relatively stable degradations, while error concealment and inpainting often assume handcrafted or simulated missing regions unlike realistic bitstream damage.Bitstream corruption can propagate irregularly over time and leave partially preserved but misleading content, limiting direct use of these methods.Source: §2.1 · pp. 2–3

2.2. Bitstream-Corrupted Video Recovery

• Liu et al. introduced BSCV, the first large-scale benchmark, and a recovery framework using residual visual cues and neighboring spatial-temporal context [32].The work established that realistic bitstream corruption differs fundamentally from conventional mask-based simulation in error concealment and video inpainting.Source: §2.2 · p. 3
• Later work explored visual foundation models and diffusion priors to guide recovery, including corruption localization and completion of corruption-aware features.These approaches integrate external priors and knowledge toward more practical and deployable recovery systems.Source: §2.2 · p. 3
• Bitstream-corrupted video restoration remains unsolved across diverse corruption patterns, practical robustness, and generalization to different contents and artifact forms.A standard challenge benchmark can systematically evaluate methods, compare technical designs, and identify promising future research directions.Source: §2.2 · p. 3

3.1. Datasets, Toolbox and Evaluation

• The benchmark provides 3,471 HD corrupted training videos with GT and masks, 50 fully released validation clips, a hidden test set, a baseline toolbox, and RGB PSNR/SSIM ranking.Training videos come from the BSCV simulation pipeline [32] and masks mark corrupted regions; scores average corrupted frames within each video, then videos across the dataset, before normalized PSNR and SSIM are weighted for final ranking.Source: §3.1 · pp. 3, 5

3.2. Challenge Phases

• The development phase lets participants use released validation pairs and a leaderboard, while the test phase requires output videos, code, and method factsheets submitted to a test server.Source: §3.2 · p. 5

3.3. Challenge Results

• 33.6423 dB PSNR and 0.9334 SSIM made MGTV-AI the quantitative leader, while RedMediaTech recorded the best overall LPIPS at 0.0852.Seven teams from 153 registered participants entered the final test phase.Source: §3.3 · p. 5
• MGTV-AI and RedMediaTech best demonstrate restoration of sharp edges and high-frequency details, while other teams suppress artifacts but often blur textures; semantic detail recovery and temporal stability remain difficult.The comparison covers four video scenarios, including cases where bitstream data is almost entirely lost.Source: Figure 2 · p. 4, §3.3 · p. 5
• Three of six submitted solutions build on B2SCVR, and half use PEFT methods such as LoRA [19] or MoE-LoRA with external priors including SAM2 [45], DINO [37], and Qwen-Image VAE [63].The reported trend combines foundation-model priors with parameter-efficient adaptation.Source: §3.3 · p. 5

4.1. MGTV-AI

• MGTV-AI uses a three-stage pipeline: mask-region completion, global temporal restoration, then spatial detail refinement with NAFNet.Its first stage replaces BSCVR-P’s optical-flow network with ProPainter’s structure [32] [72]; BasicVSR++ [4] then repairs regions inside and outside the mask, followed by NAFNet [6].Source: §4.1 · pp. 5–6

4.2. RedMediaTech

• RedMediaTech combines Wan2.1 DiT with a two-stage loss strategy and Qwen-Image VAE to address large motion and improve perceptual quality and distortion metrics.The first stage uses MSE and LPIPS; the second replaces the Wan2.1 VAE and fine-tunes with MSE to improve PSNR and SSIM, while processing about four times as many tokens.Source: §4.2 · pp. 6–7, Figure 4 · p. 7

4.3. bighit

• bighit combines DINOv3 semantic-memory retrieval with router-guided MoE-LoRA experts to recover content using historical context and adapt to heterogeneous corruption patterns.Its pipeline adds SAM2-guided structural fusion, bidirectional flow propagation, and temporal aggregation, with an optional NAFNet-style enhancer for residual artifacts and boundary inconsistencies.Source: §4.3 · p. 8

4.4. Vroom

• Vroom augments B2SCVR with SAM2 features and LoRA adaptation, then refines corruption boundaries using a residual-predicting Boundary Refinement Head.The head operates on a mask-derived boundary band and predicts blending weights to soften seams; a separate refiner constrains corrections to masked regions.Source: §4.4 · pp. 9–10

4.5. weichow

• weichow fine-tunes B2SCVR on challenge data and uses mask-guided multi-resolution compositing to preserve undamaged pixels while reconstructing corrupted regions.Its inference pipeline downsamples inputs for restoration, upsamples restored regions, and composites them with original clean pixels.Source: §4.5 · p. 10

4.6. holding

• The holding team reports improved quantitative restoration performance and visual consistency using corruption suppression, target-centric cross-frame attention, and boundary-aware seam refinement in a single-model pipeline.The modules address feature leakage from degraded regions, unstable temporal aggregation, and visible boundary artifacts.Source: §4.6 · p. 11

4.7. NTR

• NTR builds a mask-guided B2SCVR pipeline with bidirectional flow, SwinIR-based feature fusion, deformable alignment, and Temporal Focal Transformer aggregation.Predictions are applied only to masked regions, preserving the original uncorrupted pixels.Source: §4.7 · p. 12

5. Conclusion

• MGTV-AI led on PSNR and SSIM with a three-stage progressive repair framework, while RedMediaTech achieved the best LPIPS score using Wan2.1 and Qwen Image VAE.RedMediaTech used a two-stage training strategy to generate realistic textures and details in severely damaged segments.Source: §5 · p. 12
• Visual foundation models combined with parameter-efficient fine-tuning became mainstream, with teams using SAM2 semantic priors and LoRA to improve generalization while controlling computational costs.The challenge also established standardized evaluation criteria for severe spatiotemporal artifacts and content distortion caused by real-world transmission loss.Source: §5 · p. 12
• Models still struggle with edge softening, reconstructing fine features such as faces and text, and maintaining temporal coherence in large-motion scenes.The challenge validates integrating spatiotemporal attention with generative models and points toward more robust, deployable video-restoration systems.Source: §5 · pp. 12–13

bighit

• The contribution is titled “Two-Stage Bitstream-Corrupted Video Restoration via Semantic Memory and Mixture of LoRA Experts.”Source: bighit · p. 16
• Yaoxin Jiang, Jiajia Liu, and Yaokun Shi are affiliated with the University of Illinois Urbana-Champaign.Source: bighit · p. 17