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

Abstract: 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.

• The paper introduces a novel benchmark and challenge for restoring videos corrupted by bitstream errors using a large-scale dataset and precise binary masks.
• It evaluates multi-stage restoration pipelines that combine deep-learning architectures with foundation models and PEFT, quantified by PSNR, SSIM, and LPIPS.
• Methods leveraging visual priors and modular fusion demonstrate effective artifact suppression, yet challenges remain in recovering semantic details and temporal consistency.

NTIRE 2026 Challenge on Bitstream-Corrupted Video Restoration: Summary and Technical Analysis

Problem Definition and Benchmark Setup

Bitstream-corrupted video restoration (BSCVR) presents a distinct and highly challenging regime compared to conventional restoration tasks such as denoising, deblurring, or artifact reduction. Real-world video corruption arises from packet loss, bit errors, or segment damage during transmission, storage, or decoding. The resulting spatial-temporal artifacts are irregular, non-stationary, and strongly codec-dependent (Figure 1). The NTIRE 2026 Challenge establishes a standardized benchmark with a large-scale dataset (BSCV), offering corrupted sequences, ground truth frames, and binary masks precisely indicating degraded regions. Evaluation employs PSNR and SSIM for fidelity, alongside LPIPS for perceptual quality.

Figure 1: Realistic bitstream corruption patterns include block, color, duplication, misalignment, texture loss, and trailing artifacts, diverging from conventional simulated masks.

Methods: Architectural Trends and Innovations

The seven finalist teams converge on several prominent architectural trends, notably the widespread adoption of B2SCVR [liu2025towards] as the baseline, integration of visual foundation models for semantic priors, and deployment of parameter-efficient fine-tuning (PEFT) to manage computational overhead.

• MGTV-AI implements a three-stage pipeline: local inpainting using an optimized BSCVR-P (with ProPainter), temporal refinement via BasicVSR++, and spatial enhancement using NAFNet. The ensemble and multi-resolution strategy improve fine-grained structure recovery and edge sharpness (Figure 2).
• RedMediaTech leverages the Wan2.1 Diffusion Transformer, replacing its native VAE with Qwen-Image VAE, and utilizes a two-stage loss function (MSE+LPIPS, then MSE-only) to optimize perceptual and distortion metrics. The approach demonstrates the efficacy of generative priors and high-capacity latent representations in extreme corruption cases (Figure 3).
• bighit proposes a two-stage architecture with a semantic memory retrieval branch and a router-guided mixture-of-LoRA-experts (MoE-LoRA), enabling dynamic adaptation to heterogeneous corruption. Structural fusions (SAM2, DINOv3) and boundary refinement (NAFNet-style enhancer) ensure spatial-temporal consistency and artifact suppression (Figure 4).
• Vroom integrates frozen SAM2 with LoRA adaptation for spatio-temporal attention refinement. Boundary refinement heads (morphological, residual) mitigate seam effects, while a bi-directional reverse TTA further enhances temporal stability (Figure 5).
• weichow employs a mask-guided multi-resolution compositing pipeline, preserving lossless uncorrupted pixels and restoring degraded regions with B2SCVR. Frozen foundation model priors ensure robust feature extraction and computational efficiency (Figure weichow).
• holding introduces three lightweight modules: mask-aware gated suppression (M1), target-centric cross-frame attention (CA), and boundary-aware seam refinement (M2), directly targeting feature leakage and temporal inconsistency at mask boundaries (Figure 6).
• NTR uses sliding window processing, morphological mask dilation, bidirectional optical flow (SPyNet), and SwinIR-guided feature fusion. Multi-stage GAN training (with PSNR refinement) ensures robust artifact elimination and preserves structural details.

  Figure 7: Comparative results across teams reveal sharpness and artifact suppression; MGTV-AI and RedMediaTech exhibit superior semantic detail recovery, mid-tier teams reduce blockiness but often miss fine textures and temporal consistency.

  

  Figure 2: MGTV-AI’s three-stage framework: local completion (BSCVR-P), global refinement (BasicVSR++), spatial enhancement (NAFNet), ensemble for structural fidelity.

  

  Figure 3: RedMediaTech: single-step Wan2.1 DiT with two-stage training and Qwen-Image VAE swap, enhancing motion robustness and artifact resilience.

  

  Figure 4: bighit: semantic memory retrieval and MoE-LoRA adaptation jointly handle diverse corruption, followed by boundary-artifact refinement.

  

  Figure 5: Vroom: B2SCVR backbone with SAM2 prior, LoRA for attention, and specialized boundary refinement modules.

(Figure weichow)

Figure weichow: weichow pipeline: lossless mask-guided compositing preserves original pixels, B2SCVR restores corrupted regions via semantic priors.

Figure 6: holding: B2SCVR backbone with M1 (corruption suppression), CA (cross-frame aggregation), and M2 (boundary seam refinement) modules.

Quantitative Results and Comparative Evaluation

Among the seven entries, MGTV-AI achieves the highest quantitative scores: PSNR 33.642 dB and SSIM 0.9334, outperforming others in structural fidelity. RedMediaTech attains the best perceptual quality (LPIPS 0.0852), illustrating the benefit of diffusion-based generative priors. Mid-tier solutions (bighit, Vroom, weichow, holding) successfully suppress block-level artifacts, but commonly exhibit softness and incomplete recovery of semantic details, especially under severe corruption or rapid motion. Temporal stability and precise text/face restoration remain unsolved bottlenecks for all contenders.

Theoretical and Practical Implications

The challenge demonstrates the inadequacy of conventional restoration and inpainting methods under realistic bitstream corruption—residual information, non-uniform masks, and misleading artifacts necessitate task-specific pipelines. Visual foundation models (SAM2, DINOv3, Qwen-Image VAE) and PEFT approaches (LoRA, MoE-LoRA) significantly enhance generalization and efficiency. Multistage fusion and compositing strategies enable modular design, separating lossless content from synthesized restoration. Diffusion models and transformer-based architectures provide robust generative priors for hallucinating missing data, but the accurate recovery of semantic content and maintenance of long-range temporal coherence are unsolved, suggesting further research is needed in spatial-temporal semantic modeling, advanced compositing strategies, and codec-aware restoration.

Future Directions

Advancements in foundation model integration, parameter-efficient tuning, and generative pipelines will drive further progress. Robustness across codecs, scalability to high-resolution and long-form content, and real-time deployment must be targeted. Approaches combining foundation models’ semantic segmentation with temporal transformers, or hybrid diffusion-transformer pipelines, may offer increased fidelity and stability. Addressing domain transfer, uncertainty quantification, and human-in-the-loop restoration will be critical for practical adoption in streaming and surveillance.

Conclusion

The NTIRE 2026 Challenge on Bitstream-Corrupted Video Restoration (2604.06945) establishes a rigorous benchmark and technical baseline for tackling real-world severe video degradation. Leading methods demonstrate the importance of multi-stage modular pipelines, foundation model priors, and PEFT strategies. While state-of-the-art solutions effectively suppress structural artifacts and hallucinate missing regions, semantic detail recovery and temporal stability remain open problems. Further developments are expected in foundation-model-driven, codec-aware restoration architectures, accelerating the deployment of robust AI systems in video communication, surveillance, and content creation domains.