Linze Li

Andrey Moskalenko, Alexey Bryncev, Ivan Kosmynin et al.

This paper presents an overview of the NTIRE 2026 Challenge on Video Saliency Prediction. The goal of the challenge participants was to develop automatic saliency map prediction methods for the provided video sequences. The novel dataset of 2,000 diverse videos with an open license was prepared for this challenge. The fixations and corresponding saliency maps were collected using crowdsourced mouse tracking and contain viewing data from over 5,000 assessors. Evaluation was performed on a subset of 800 test videos using generally accepted quality metrics. The challenge attracted over 20 teams making submissions, and 7 teams passed the final phase with code review. All data used in this challenge is made publicly available - https://github.com/msu-video-group/NTIRE26_Saliency_Prediction.

Jiajun Liang, Linze Li, Zhaodong Bing et al.

Self-distillation exploits non-uniform soft supervision from itself during training and improves performance without any runtime cost. However, the overhead during training is often overlooked, and yet reducing time and memory overhead during training is increasingly important in the giant models' era. This paper proposes an efficient self-distillation method named Zipf's Label Smoothing (Zipf's LS), which uses the on-the-fly prediction of a network to generate soft supervision that conforms to Zipf distribution without using any contrastive samples or auxiliary parameters. Our idea comes from an empirical observation that when the network is duly trained the output values of a network's final softmax layer, after sorting by the magnitude and averaged across samples, should follow a distribution reminiscent to Zipf's Law in the word frequency statistics of natural languages. By enforcing this property on the sample level and throughout the whole training period, we find that the prediction accuracy can be greatly improved. Using ResNet50 on the INAT21 fine-grained classification dataset, our technique achieves +3.61% accuracy gain compared to the vanilla baseline, and 0.88% more gain against the previous label smoothing or self-distillation strategies. The implementation is publicly available at https://github.com/megvii-research/zipfls.

Rongchang Li, Zhenhua Feng, Tianyang Xu et al.

Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring so-called compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variations between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://github.com/RongchangLi/ZSCAR_C2C.

Chengxing Xie, Xiaoming Zhang, Linze Li et al.

Efficient and lightweight single-image super-resolution (SISR) has achieved remarkable performance in recent years. One effective approach is the use of large kernel designs, which have been shown to improve the performance of SISR models while reducing their computational requirements. However, current state-of-the-art (SOTA) models still face problems such as high computational costs. To address these issues, we propose the Large Kernel Distillation Network (LKDN) in this paper. Our approach simplifies the model structure and introduces more efficient attention modules to reduce computational costs while also improving performance. Specifically, we employ the reparameterization technique to enhance model performance without adding extra cost. We also introduce a new optimizer from other tasks to SISR, which improves training speed and performance. Our experimental results demonstrate that LKDN outperforms existing lightweight SR methods and achieves SOTA performance.

Shurong Yang, Huadong Li, Juhao Wu et al.

Diffusion models have demonstrated superior performance in the field of portrait animation. However, current approaches relied on either visual or audio modality to control character movements, failing to exploit the potential of mixed-modal control. This challenge arises from the difficulty in balancing the weak control strength of audio modality and the strong control strength of visual modality. To address this issue, we introduce MegActor-$Σ$: a mixed-modal conditional diffusion transformer (DiT), which can flexibly inject audio and visual modality control signals into portrait animation. Specifically, we make substantial advancements over its predecessor, MegActor, by leveraging the promising model structure of DiT and integrating audio and visual conditions through advanced modules within the DiT framework. To further achieve flexible combinations of mixed-modal control signals, we propose a ``Modality Decoupling Control" training strategy to balance the control strength between visual and audio modalities, along with the ``Amplitude Adjustment" inference strategy to freely regulate the motion amplitude of each modality. Finally, to facilitate extensive studies in this field, we design several dataset evaluation metrics to filter out public datasets and solely use this filtered dataset to train MegActor-$Σ$. Extensive experiments demonstrate the superiority of our approach in generating vivid portrait animations, outperforming previous methods trained on private dataset.

Pengcheng Shao, Tianyang Xu, Zhangyong Tang et al.

There is currently strong interest in improving visual object tracking by augmenting the RGB modality with the output of a visual event camera that is particularly informative about the scene motion. However, existing approaches perform event feature extraction for RGB-E tracking using traditional appearance models, which have been optimised for RGB only tracking, without adapting it for the intrinsic characteristics of the event data. To address this problem, we propose an Event backbone (Pooler), designed to obtain a high-quality feature representation that is cognisant of the innate characteristics of the event data, namely its sparsity. In particular, Multi-Scale Pooling is introduced to capture all the motion feature trends within event data through the utilisation of diverse pooling kernel sizes. The association between the derived RGB and event representations is established by an innovative module performing adaptive Mutually Guided Fusion (MGF). Extensive experimental results show that our method significantly outperforms state-of-the-art trackers on two widely used RGB-E tracking datasets, including VisEvent and COESOT, where the precision and success rates on COESOT are improved by 4.9% and 5.2%, respectively. Our code will be available at https://github.com/SSSpc333/TENet.

Linze Li, Zong-Wei Hong, Shen Zhang et al.

While Flow Matching theoretically guarantees constant-velocity trajectories, we identify a critical breakdown in high-dimensional practice: the Velocity Deficit. We show that the MSE objective systematically underestimates velocity magnitude, causing generated samples to fail to reach the data manifold-a phenomenon we term Integration Lag. To rectify this, we propose Initial Energy Injection, instantiated via two complementary methods: the training-based Magnitude-Aware Flow Matching (MAFM) and the training-free Scale Schedule Corrector (SSC). Both are grounded in our discovery of a crucial asymmetry: velocity contraction causes harmful kinetic stagnation at the trajectory's start, yet acts as a beneficial denoising mechanism at its end. Empirically, SSC yields significant efficiency gains with zero retraining and just one line of code. On ImageNet-1k (256x256), it improves FID by 44.6% (from 13.68 to 7.58) and achieves a 5x speedup, enabling a 50-step generator (FID 7.58) to beat a 250-step baseline (FID 8.65). Furthermore, our methods generalize to Text-to-Image tasks and high-resolution generation, improving FID on MS-COCO by ~22%.

Xiaole Zhao, Linze Li, Chengxing Xie et al.

Transformer-based deep models for single image super-resolution (SISR) have greatly improved the performance of lightweight SISR tasks in recent years. However, they often suffer from heavy computational burden and slow inference due to the complex calculation of multi-head self-attention (MSA), seriously hindering their practical application and deployment. In this work, we present an efficient SR model to mitigate the dilemma between model efficiency and SR performance, which is dubbed Entropy Attention and Receptive Field Augmentation network (EARFA), and composed of a novel entropy attention (EA) and a shifting large kernel attention (SLKA). From the perspective of information theory, EA increases the entropy of intermediate features conditioned on a Gaussian distribution, providing more informative input for subsequent reasoning. On the other hand, SLKA extends the receptive field of SR models with the assistance of channel shifting, which also favors to boost the diversity of hierarchical features. Since the implementation of EA and SLKA does not involve complex computations (such as extensive matrix multiplications), the proposed method can achieve faster nonlinear inference than Transformer-based SR models while maintaining better SR performance. Extensive experiments show that the proposed model can significantly reduce the delay of model inference while achieving the SR performance comparable with other advanced models.

Yongqi Ding, Kunshan Yang, Linze Li et al.

Although the temporal spike dynamics of spiking neural networks (SNNs) enable low-power temporal pattern capture capabilities, they also incur inherent inconsistencies that severely compromise representation. In this paper, we perform dual consistency optimization via Stable Spike to mitigate this problem, thereby improving the recognition performance of SNNs. With the hardware-friendly ``AND" bit operation, we efficiently decouple the stable spike skeleton from the multi-timestep spike maps, thereby capturing critical semantics while reducing inconsistencies from variable noise spikes. Enforcing the unstable spike maps to converge to the stable spike skeleton significantly improves the inherent consistency across timesteps. Furthermore, we inject amplitude-aware spike noise into the stable spike skeleton to diversify the representations while preserving consistent semantics. The SNN is encouraged to produce perturbation-consistent predictions, thereby contributing to generalization. Extensive experiments across multiple architectures and datasets validate the effectiveness and versatility of our method. In particular, our method significantly advances neuromorphic object recognition under ultra-low latency, improving accuracy by up to 8.33\%. This will help unlock the full power consumption and speed potential of SNNs.

Chengxing Xie, Xiaoming Zhang, Linze Li et al.

Image super-resolution (SR) has significantly advanced through the adoption of Transformer architectures. However, conventional techniques aimed at enlarging the self-attention window to capture broader contexts come with inherent drawbacks, especially the significantly increased computational demands. Moreover, the feature perception within a fixed-size window of existing models restricts the effective receptive field (ERF) and the intermediate feature diversity. We demonstrate that a flexible integration of attention across diverse spatial extents can yield significant performance enhancements. In line with this insight, we introduce Multi-Range Attention Transformer (MAT) for SR tasks. MAT leverages the computational advantages inherent in dilation operation, in conjunction with self-attention mechanism, to facilitate both multi-range attention (MA) and sparse multi-range attention (SMA), enabling efficient capture of both regional and sparse global features. Combined with local feature extraction, MAT adeptly capture dependencies across various spatial ranges, improving the diversity and efficacy of its feature representations. We also introduce the MSConvStar module, which augments the model's ability for multi-range representation learning. Comprehensive experiments show that our MAT exhibits superior performance to existing state-of-the-art SR models with remarkable efficiency (~3.3 faster than SRFormer-light).

Shen Zhang, Siyuan Liang, Yaning Tan et al.

Diffusion transformers (DiTs) struggle to generate images at resolutions higher than their training resolutions. The primary obstacle is that the explicit positional encodings(PE), such as RoPE, need extrapolating to unseen positions which degrades performance when the inference resolution differs from training. In this paper, We propose a Length-Extrapolatable Diffusion Transformer~(LEDiT) to overcome this limitation. LEDiT needs no explicit PEs, thereby avoiding PE extrapolation. The key innovation of LEDiT lies in the use of causal attention. We demonstrate that causal attention can implicitly encode global positional information and show that such information facilitates extrapolation. We further introduce a locality enhancement module, which captures fine-grained local information to complement the global coarse-grained position information encoded by causal attention. Experimental results on both conditional and text-to-image generation tasks demonstrate that LEDiT supports up to 4x resolution scaling (e.g., from 256x256 to 512x512), achieving better image quality compared to the state-of-the-art length extrapolation methods. We believe that LEDiT marks a departure from the standard RoPE-based methods and offers a promising insight into length extrapolation. Project page: https://shenzhang2145.github.io/ledit/

Zhaodong Bing, Linze Li, Jiajun Liang

Knowledge distillation (KD) in transformers often faces challenges due to misalignment in the number of attention heads between teacher and student models. Existing methods either require identical head counts or introduce projectors to bridge dimensional gaps, limiting flexibility and efficiency. We propose Squeezing-Heads Distillation (SHD), a novel approach that enables seamless knowledge transfer between models with varying head counts by compressing multi-head attention maps via efficient linear approximation. Unlike prior work, SHD eliminates alignment barriers without additional parameters or architectural modifications. Our method dynamically approximates the combined effect of multiple teacher heads into fewer student heads, preserving fine-grained attention patterns while reducing redundancy. Experiments across language (LLaMA, GPT) and vision (DiT, MDT) generative and vision (DeiT) discriminative tasks demonstrate SHD's effectiveness: it outperforms logit-based and feature-alignment KD baselines, achieving state-of-the-art results in image classification, image generation language fine-tuning, and language pre-training. The key innovations of flexible head compression, projector-free design, and linear-time complexity make SHD a versatile and scalable solution for distilling modern transformers. This work bridges a critical gap in KD, enabling efficient deployment of compact models without compromising performance.

Linze Li, Sunqi Fan, Hengjun Pu et al.

Over recent years, diffusion models have facilitated significant advancements in video generation. Yet, the creation of face-related videos still confronts issues such as low facial fidelity, lack of frame consistency, limited editability and uncontrollable human poses. To address these challenges, we introduce a facial animation generation method that enhances both face identity fidelity and editing capabilities while ensuring frame consistency. This approach incorporates the concept of an anchor frame to counteract the degradation of generative ability in original text-to-image models when incorporating a motion module. We propose two strategies towards this objective: training-free and training-based anchor frame methods. Our method's efficacy has been validated on multiple representative DreamBooth and LoRA models, delivering substantial improvements over the original outcomes in terms of facial fidelity, text-to-image editability, and video motion. Moreover, we introduce conditional control using a 3D parametric face model to capture accurate facial movements and expressions. This solution augments the creative possibilities for facial animation generation through the integration of multiple control signals. For additional samples, please visit https://paper-faac.github.io/.