Shuai Li

Tencent Hunyuan Team, Ao Liu, Botong Zhou et al. · tencent-ai

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Shiqi Jiang, Ting Ren, Congrui Fu et al.

Screen content (SC) differs from natural scene (NS) with unique characteristics such as noise-free, repetitive patterns, and high contrast. Aiming at addressing the inadequacies of current learned image compression (LIC) methods for SC, we propose an improved two-stage octave convolutional residual blocks (IToRB) for high and low-frequency feature extraction and a cascaded two-stage multi-scale residual blocks (CTMSRB) for improved multi-scale learning and nonlinearity in SC. Additionally, we employ a window-based attention module (WAM) to capture pixel correlations, especially for high contrast regions in the image. We also construct a diverse SC image compression dataset (SDU-SCICD2K) for training, including text, charts, graphics, animation, movie, game and mixture of SC images and NS images. Experimental results show our method, more suited for SC than NS data, outperforms existing LIC methods in rate-distortion performance on SC images. The code is publicly available at https://github.com/SunshineSki/OMR Net.git.

Chang Sun, Hui Yuan, Shuai Li et al.

In point cloud geometry compression, context models usually use the one-hot encoding of node occupancy as the label, and the cross-entropy between the one-hot encoding and the probability distribution predicted by the context model as the loss function. However, this approach has two main weaknesses. First, the differences between contexts of different nodes are not significant, making it difficult for the context model to accurately predict the probability distribution of node occupancy. Second, as the one-hot encoding is not the actual probability distribution of node occupancy, the cross-entropy loss function is inaccurate. To address these problems, we propose a general structure that can enhance existing context models. We introduce the context feature residuals into the context model to amplify the differences between contexts. We also add a multi-layer perception branch, that uses the mean squared error between its output and node occupancy as a loss function to provide accurate gradients in backpropagation. We validate our method by showing that it can improve the performance of an octree-based model (OctAttention) and a voxel-based model (VoxelDNN) on the object point cloud datasets MPEG 8i and MVUB, as well as the LiDAR point cloud dataset SemanticKITTI.

Shuai Li, Huibin Bai, Yanbo Gao et al.

Monocular depth estimation (MDE) has attracted increasing interest in the past few years, owing to its important role in 3D vision. MDE is the estimation of a depth map from a monocular image/video to represent the 3D structure of a scene, which is a highly ill-posed problem. To solve this problem, in this paper, we propose a LiftFormer based on lifting theory topology, for constructing an intermediate subspace that bridges the image color features and depth values, and a subspace that enhances the depth prediction around edges. MDE is formulated by transforming the depth value prediction problem into depth-oriented geometric representation (DGR) subspace feature representation, thus bridging the learning from color values to geometric depth values. A DGR subspace is constructed based on frame theory by using linearly dependent vectors in accordance with depth bins to provide a redundant and robust representation. The image spatial features are transformed into the DGR subspace, where these features correspond directly to the depth values. Moreover, considering that edges usually present sharp changes in a depth map and tend to be erroneously predicted, an edge-aware representation (ER) subspace is constructed, where depth features are transformed and further used to enhance the local features around edges. The experimental results demonstrate that our LiftFormer achieves state-of-the-art performance on widely used datasets, and an ablation study validates the effectiveness of both proposed lifting modules in our LiftFormer.

Xiangyu Chen, Ruiwen Zhen, Shuai Li et al.

Image restoration aims to restore high-quality images from degraded counterparts and has seen significant advancements through deep learning techniques. The technique has been widely applied to mobile devices for tasks such as mobile photography. Given the resource limitations on mobile devices, such as memory constraints and runtime requirements, the efficiency of models during deployment becomes paramount. Nevertheless, most previous works have primarily concentrated on analyzing the efficiency of single modules and improving them individually. This paper examines the efficiency across different layers. We propose a roadmap that can be applied to further accelerate image restoration models prior to deployment while simultaneously increasing PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index). The roadmap first increases the model capacity by adding more parameters to partial convolutions on FLOPs non-sensitive layers. Then, it applies partial depthwise convolution coupled with decoupling upsampling/downsampling layers to accelerate the model speed. Extensive experiments demonstrate that our approach decreases runtime by up to 13% and reduces the number of parameters by up to 23%, while increasing PSNR and SSIM on several image restoration datasets. Source Code of our method is available at \href{https://github.com/xiangyu8/MOFA}{https://github.com/xiangyu8/MOFA}.

Yanbo Gao, Huibin Bai, Huasong Zhou et al.

Self-supervised monocular depth estimation (MDE) has received increasing interests in the last few years. The objects in the scene, including the object size and relationship among different objects, are the main clues to extract the scene structure. However, previous works lack the explicit handling of the changing sizes of the object due to the change of its depth. Especially in a monocular video, the size of the same object is continuously changed, resulting in size and depth ambiguity. To address this problem, we propose a Depth-converted-Scale Convolution (DcSConv) enhanced monocular depth estimation framework, by incorporating the prior relationship between the object depth and object scale to extract features from appropriate scales of the convolution receptive field. The proposed DcSConv focuses on the adaptive scale of the convolution filter instead of the local deformation of its shape. It establishes that the scale of the convolution filter matters no less (or even more in the evaluated task) than its local deformation. Moreover, a Depth-converted-Scale aware Fusion (DcS-F) is developed to adaptively fuse the DcSConv features and the conventional convolution features. Our DcSConv enhanced monocular depth estimation framework can be applied on top of existing CNN based methods as a plug-and-play module to enhance the conventional convolution block. Extensive experiments with different baselines have been conducted on the KITTI benchmark and our method achieves the best results with an improvement up to 11.6% in terms of SqRel reduction. Ablation study also validates the effectiveness of each proposed module.

Chengxu Liu, Xuan Wang, Yuanting Fan et al.

Under-display camera (UDC) systems are the foundation of full-screen display devices in which the lens mounts under the display. The pixel array of light-emitting diodes used for display diffracts and attenuates incident light, causing various degradations as the light intensity changes. Unlike general video restoration which recovers video by treating different degradation factors equally, video restoration for UDC systems is more challenging that concerns removing diverse degradation over time while preserving temporal consistency. In this paper, we introduce a novel video restoration network, called D$^2$RNet, specifically designed for UDC systems. It employs a set of Decoupling Attention Modules (DAM) that effectively separate the various video degradation factors. More specifically, a soft mask generation function is proposed to formulate each frame into flare and haze based on the diffraction arising from incident light of different intensities, followed by the proposed flare and haze removal components that leverage long- and short-term feature learning to handle the respective degradations. Such a design offers an targeted and effective solution to eliminating various types of degradation in UDC systems. We further extend our design into multi-scale to overcome the scale-changing of degradation that often occur in long-range videos. To demonstrate the superiority of D$^2$RNet, we propose a large-scale UDC video benchmark by gathering HDR videos and generating realistically degraded videos using the point spread function measured by a commercial UDC system. Extensive quantitative and qualitative evaluations demonstrate the superiority of D$^2$RNet compared to other state-of-the-art video restoration and UDC image restoration methods. Code is available at https://github.com/ChengxuLiu/DDRNet.git

Dengyan Luo, Yanping Xiang, Hu Wang et al.

The detection of moving infrared dim-small targets has been a challenging and prevalent research topic. The current state-of-the-art methods are mainly based on ConvLSTM to aggregate information from adjacent frames to facilitate the detection of the current frame. However, these methods implicitly utilize motion information only in the training stage and fail to explicitly explore motion compensation, resulting in poor performance in the case of a video sequence including large motion. In this paper, we propose a Deformable Feature Alignment and Refinement (DFAR) method based on deformable convolution to explicitly use motion context in both the training and inference stages. Specifically, a Temporal Deformable Alignment (TDA) module based on the designed Dilated Convolution Attention Fusion (DCAF) block is developed to explicitly align the adjacent frames with the current frame at the feature level. Then, the feature refinement module adaptively fuses the aligned features and further aggregates useful spatio-temporal information by means of the proposed Attention-guided Deformable Fusion (AGDF) block. In addition, to improve the alignment of adjacent frames with the current frame, we extend the traditional loss function by introducing a new motion compensation loss. Extensive experimental results demonstrate that the proposed DFAR method achieves the state-of-the-art performance on two benchmark datasets including DAUB and IRDST.

HongXin Xie, JianDe Sun, Yi Shao et al.

Molecular odor prediction has great potential across diverse fields such as chemistry, pharmaceuticals, and environmental science, enabling the rapid design of new materials and enhancing environmental monitoring. However, current methods face two main challenges: First, existing models struggle with non-smooth objective functions and the complexity of mixed feature dimensions; Second, datasets suffer from severe label imbalance, which hampers model training, particularly in learning minority class labels. To address these issues, we introduce a novel feature mapping method and a molecular ensemble optimization loss function. By incorporating feature importance learning and frequency modulation, our model adaptively adjusts the contribution of each feature, efficiently capturing the intricate relationship between molecular structures and odor descriptors. Our feature mapping preserves feature independence while enhancing the model's efficiency in utilizing molecular features through frequency modulation. Furthermore, the proposed loss function dynamically adjusts label weights, improves structural consistency, and strengthens label correlations, effectively addressing data imbalance and label co-occurrence challenges. Experimental results show that our method significantly can improves the accuracy of molecular odor prediction across various deep learning models, demonstrating its promising potential in molecular structure representation and chemoinformatics.

Chuankun Li, Shuai Li, Yanbo Gao et al.

Unsupervised skeleton based action recognition has achieved remarkable progress recently. Existing unsupervised learning methods suffer from severe overfitting problem, and thus small networks are used, significantly reducing the representation capability. To address this problem, the overfitting mechanism behind the unsupervised learning for skeleton based action recognition is first investigated. It is observed that the skeleton is already a relatively high-level and low-dimension feature, but not in the same manifold as the features for action recognition. Simply applying the existing unsupervised learning method may tend to produce features that discriminate the different samples instead of action classes, resulting in the overfitting problem. To solve this problem, this paper presents an Unsupervised spatial-temporal Feature Enrichment and Fidelity Preservation framework (U-FEFP) to generate rich distributed features that contain all the information of the skeleton sequence. A spatial-temporal feature transformation subnetwork is developed using spatial-temporal graph convolutional network and graph convolutional gate recurrent unit network as the basic feature extraction network. The unsupervised Bootstrap Your Own Latent based learning is used to generate rich distributed features and the unsupervised pretext task based learning is used to preserve the information of the skeleton sequence. The two unsupervised learning ways are collaborated as U-FEFP to produce robust and discriminative representations. Experimental results on three widely used benchmarks, namely NTU-RGB+D-60, NTU-RGB+D-120 and PKU-MMD dataset, demonstrate that the proposed U-FEFP achieves the best performance compared with the state-of-the-art unsupervised learning methods. t-SNE illustrations further validate that U-FEFP can learn more discriminative features for unsupervised skeleton based action recognition.

Chuankun Li, Shuai Li, Yanbo Gao et al.

Pose based hand gesture recognition has been widely studied in the recent years. Compared with full body action recognition, hand gesture involves joints that are more spatially closely distributed with stronger collaboration. This nature requires a different approach from action recognition to capturing the complex spatial features. Many gesture categories, such as "Grab" and "Pinch", have very similar motion or temporal patterns posing a challenge on temporal processing. To address these challenges, this paper proposes a two-stream neural network with one stream being a self-attention based graph convolutional network (SAGCN) extracting the short-term temporal information and hierarchical spatial information, and the other being a residual-connection enhanced bidirectional Independently Recurrent Neural Network (RBi-IndRNN) for extracting long-term temporal information. The self-attention based graph convolutional network has a dynamic self-attention mechanism to adaptively exploit the relationships of all hand joints in addition to the fixed topology and local feature extraction in the GCN. On the other hand, the residual-connection enhanced Bi-IndRNN extends an IndRNN with the capability of bidirectional processing for temporal modelling. The two streams are fused together for recognition. The Dynamic Hand Gesture dataset and First-Person Hand Action dataset are used to validate its effectiveness, and our method achieves state-of-the-art performance.

Chenpeng Zhang, Shuai Li, Mao Ye et al.

Recurrent neural networks (RNNs) are widely used as a memory model for sequence-related problems. Many variants of RNN have been proposed to solve the gradient problems of training RNNs and process long sequences. Although some classical models have been proposed, capturing long-term dependence while responding to short-term changes remains a challenge. To this problem, we propose a new model named Dual Recurrent Neural Networks (DuRNN). The DuRNN consists of two parts to learn the short-term dependence and progressively learn the long-term dependence. The first part is a recurrent neural network with constrained full recurrent connections to deal with short-term dependence in sequence and generate short-term memory. Another part is a recurrent neural network with independent recurrent connections which helps to learn long-term dependence and generate long-term memory. A selection mechanism is added between two parts to help the needed long-term information transfer to the independent neurons. Multiple modules can be stacked to form a multi-layer model for better performance. Our contributions are: 1) a new recurrent model developed based on the divide-and-conquer strategy to learn long and short-term dependence separately, and 2) a selection mechanism to enhance the separating and learning of different temporal scales of dependence. Both theoretical analysis and extensive experiments are conducted to validate the performance of our model, and we also conduct simple visualization experiments and ablation analyses for the model interpretability. Experimental results indicate that the proposed DuRNN model can handle not only very long sequences (over 5000 time steps), but also short sequences very well. Compared with many state-of-the-art RNN models, our model has demonstrated efficient and better performance.

Shuai Li, Dinei Florencio, Wanqing Li et al.

Detecting camouflaged moving foreground objects has been known to be difficult due to the similarity between the foreground objects and the background. Conventional methods cannot distinguish the foreground from background due to the small differences between them and thus suffer from under-detection of the camouflaged foreground objects. In this paper, we present a fusion framework to address this problem in the wavelet domain. We first show that the small differences in the image domain can be highlighted in certain wavelet bands. Then the likelihood of each wavelet coefficient being foreground is estimated by formulating foreground and background models for each wavelet band. The proposed framework effectively aggregates the likelihoods from different wavelet bands based on the characteristics of the wavelet transform. Experimental results demonstrated that the proposed method significantly outperformed existing methods in detecting camouflaged foreground objects. Specifically, the average F-measure for the proposed algorithm was 0.87, compared to 0.71 to 0.8 for the other state-of-the-art methods.

Shuai Li, Wanqing Li, Chris Cook et al.

Pooling is an important component in convolutional neural networks (CNNs) for aggregating features and reducing computational burden. Compared with other components such as convolutional layers and fully connected layers which are completely learned from data, the pooling component is still handcrafted such as max pooling and average pooling. This paper proposes a learnable pooling function using recurrent neural networks (RNN) so that the pooling can be fully adapted to data and other components of the network, leading to an improved performance. Such a network with learnable pooling function is referred to as a fully trainable network (FTN). Experimental results have demonstrated that the proposed RNN-based pooling can well approximate the existing pooling functions and improve the performance of the network. Especially for small networks, the proposed FTN can improve the performance by seven percentage points in terms of error rate on the CIFAR-10 dataset compared with the traditional CNN.