Existing low-light enhancement methods typically rely on fitting data mappings (pixel-wise mappings through fully supervised methods or distribution-wise mappings through weakly supervised or self-supervised methods). However, their performance is heavily dependent on specific scenes and fails to adequately model the intrinsic prior of natural images, resulting in poor generalization. To tackle this challenge, we leverage the strengths of powerful generative diffusion models, conditioned on a thoughtfully designed prior, and propose a novel zero-reference low-light enhancement framework that gets rid of dependence on the distribution of low-light images. In detail, we address the most fundamental core by proposing an illumination-invariant prior derived from the theory of physical light transfer, bridging the gap between normal and low-light domains, and enabling zero-shot enhancement without the need for low-light-specific training. A prior-to-image restoration framework is built upon generative diffusion models, pre-trained on normal-light data. During inference, the framework extracts the illumination-invariant prior from low-light inputs and maps them back to high-quality images, naturally for low-light enhancement. Additionally, such intrinsic properties of illumination-invariant prior open up opportunities for distilling diffusion models into compact CNN-based networks. We propose a novel prior-injected distillation paradigm incorporating intensity, frequency, and gradient domain-augmented regularization comprehensively. This distillation framework not only reduces computational costs but also maintains high fidelity and perceptual quality in enhanced outputs, making it more efficient and practical for real-world applications. The approach further extends seamlessly to handle over-exposure scenarios, demonstrating its versatility in addressing complex lighting conditions. Extensive experiments demonstrate the superiority of our framework in various scenarios, as well as its strong interpretability, robustness, and efficiency.

A low-light image taken in a dark scene usually suffers from severe distortions, which does not accurately characterize the ambient lighting. Long exposure is an accustomed way to capture more supplementary light and alleviate the degradation, but sometimes it induces other distortions, e.g., blurriness. To address this issue, we propose a new paradigm that introduces additional captured ambient guidance, i.e., a long-exposure image to steer the low-light enhancement. In practice, this long-exposure image can be obtained conveniently, but usually suffers from blurriness and misalignment. To effectively extract and fuse information from degraded and misaligned low-light and guidance image pairs, we propose a Long Exposure Compensation Network (LECNet). Adaptive Band Regression is introduced to disentangle the image into multi-scale representations and coarse-to-fine aggregate them with an attention mechanism. For stable image-guidance registration and artifact suppression, we propose a Bounded Cross-domain Deformable Alignment to warp the guidance based on extracted feature pyramids step by step. To integrate knowledge about the degradation into our LECNet for better fidelity, a dual learned back projection is enforced between the predicted result and the paired inputs in illumination and texture detail consistency, serving the model training for both offline training and online sample-adaptive finetuning. For training and evaluation of this new paradigm, we build a dataset with both synthetic and real-captured image triplets of long/short exposure pairs and extra blurry guidance. The experimental evaluation demonstrates the significance of our new paradigm, as well as the superiority of our LECNet and its usability in the real world.

Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity. When quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (**V**ideo \& **I**mage **Di**ffusion **T**ransformer **Q**uantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory optimization and a 1.4-1.7x end-to-end latency speedup.

Long-Context Transformer Models (LCTMs) are vital for real-world applications but suffer high computational costs due to attention’s quadratic complexity. Block-sparse attention mitigates this by focusing computation on critical regions, yet existing methods struggle with balancing accuracy and efficiency due to costly block importance measurements. In this paper, we introduce XAttention, a plug-and-play framework that dramatically accelerates long-context inference in Transformers models using sparse attention. XAttention’s key innovation is the insight that the sum of antidiagonal values (i.e., from the lower-left to upper-right) in the attention matrix provides a powerful proxy for block importance. This allows for precise identification and pruning of non-essential blocks, resulting in high sparsity and dramatically accelerated inference. Across comprehensive evaluations on demanding long-context benchmarks—including RULER and LongBench for language, VideoMME for video understanding, and VBench for video generation—XAttention achieves accuracy comparable to full attention while delivering substantial computational gains. We demonstrate up to 13.5x acceleration in attention computation. These results underscore XAttention’s ability to unlock the practical potential of block sparse attention, paving the way for scalable and efficient deployment of LCTMs in real-world applications.

Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. To further enhance the efficiency of attention computation compared to SageAttention while maintaining precision, we propose SageAttention2, which utilizes significantly faster 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, we propose to quantize matrixes $(Q, K)$ to INT4 in a hardware-friendly thread-level granularity and quantize matrixes $(\widetilde P, V)$ to FP8. Second, we propose a method to smooth $Q$, enhancing the accuracy of INT4 $QK^\top$. Third, we propose a two-level accumulation strategy for $\widetilde PV$ to enhance the accuracy of FP8 $\widetilde PV$. The operations per second (OPS) of SageAttention2 surpass FlashAttention2 and xformers by about 3x and 4.5x, respectively. Moreover, SageAttention2 matches the speed of FlashAttention3(fp8) on the Hopper GPUs, while delivering much higher accuracy. Comprehensive experiments confirm that our approach incurs negligible end-to-end metrics loss across diverse models, including those for language, image, and video generation.

An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics.

Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at both training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight critical tokens; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85\% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53$\times$ with no drop in diffusion loss. Retrofitting the open-source Wan2.1-1.3B model speeds up attention time by 6$\times$ and lowers end-to-end generation time from 31s to 18s with comparable quality, while for the 14B model, end-to-end generation time is reduced from 1274s to 576s. Furthermore, we introduce a preliminary study of Sparse-Distill, the first method to enable sparse attention and distillation concurrently, achieving 50.9x speed up for Wan-1.3B while maintaining quality. These results establish trainable sparse attention as a practical alternative to full attention and a key enabler for further scaling of video diffusion models. Code is available at https://github.com/hao-ai-lab/FastVideo.

The efficiency of attention is important due to its quadratic time complexity. We enhance the efficiency of attention through two key contributions: First, we leverage the new $\texttt{FP4}$ Tensor Cores in Blackwell GPUs to accelerate attention computation. Our implementation achieves $\textbf{1038}$ $\texttt{TOPS}$ on $\texttt{RTX5090}$, which is a $\textbf{5}\times$ speedup over the fastest FlashAttention on $\texttt{RTX5090}$. Experiments show that our $\texttt{FP4}$ attention can accelerate inference of various models in a plug-and-play way. Second, we pioneer low-bit attention to training tasks. Existing low-bit attention works like FlashAttention3 and SageAttention focus only on inference. However, the efficiency of training large models is also important. To explore whether low-bit attention can be effectively applied to training tasks, we design an accurate and efficient $\texttt{8-bit}$ attention for both forward and backward propagation. Experiments indicate that $\texttt{8-bit}$ attention achieves lossless performance in fine-tuning tasks but exhibits slower convergence in pretraining tasks. The code is available at https://github.com/thu-ml/SageAttention.

As the Large Language Model (LLM) becomes increasingly important in various domains, the performance of LLM inference is crucial to massive LLM applications. However, centering around the computational efficiency and the memory utilization, the following challenges remain unsolved in achieving high-throughput LLM inference: (1) Synchronous partial softmax update. The softmax operation requires a synchronous update operation among each partial softmax result, leading to ~20% overheads for the attention computation in LLMs. (2) Under-utilized computation of flat GEMM. The shape of matrices performing GEMM in LLM inference tends to be flat, leading to under-utilized computation and 50% performance loss after padding zeros in previous designs (e.g., cuBLAS, CUTLASS, etc.). (3) Memory redundancy caused by activations. Dynamic allocation of activations during inference leads to redundant storage of useless variables, bringing 22% more memory consumption.We presentFlashDecoding++Next, a high-throughput inference engine supporting mainstream LLMs and hardware backends. To tackle the above challenges,FlashDecoding++Nextcreatively proposes: (1) Asynchronous softmax with unified maximum.FlashDecoding++Nextintroduces a unified maximum technique for different partial softmax computations to avoid synchronization. Based on this, a fine-grained pipelining is proposed, leading to 1.18× and 1.14× for theprefillanddecodephases in LLM inference, respectively. (2) Flat GEMM optimization with double buffering.FlashDecoding++Nextpoints out that flat GEMMs with different shapes face varied bottlenecks. Then, techniques like double buffering are introduced, resulting in up to 52% speedup for the flat GEMM operation. (3) Buffer reusing and unified memory management.FlashDecoding++Nextreuses the pre-allocated activation buffers throughout the inference process to remove redundancy. Based on that, we unify the management of different types of storage to further exploit the reusing opportunity. The memory optimization enables up to 1.57× longer sequence to be processed.FlashDecoding++Nextdemonstrates remarkable throughput improvement, delivering up to 68.88× higher throughput compared to the HuggingFace [1] implementation. On average,FlashDecoding++Nextachieves 1.25× and 1.46× higher throughput compared to vLLM [2] and TensorRT-LLM [3] on mainstream LLMs.

Although the recent learning-based image and video coding techniques achieve rapid development, the signal fidelity-driven target in these methods leads to the divergence to a highly effective and efficient coding framework for both human and machine. In this paper, we aim to address the issue by making use of the power of generative models to bridge the gap between full fidelity (for human vision) and high discrimination (for machine vision). Therefore, relying on existing pretrained generative adversarial networks (GAN), we build a GAN inversion framework that projects the image into a low-dimensional natural image manifold. In this manifold, the feature is highly discriminative and also encodes the appearance information of the image, named aslatent code. Taking a variational bit-rate constraint with a hyperprior model to model/suppress the entropy of image manifold code, our method is capable of fulfilling the needs of both machine and human visions at very low bit-rates. To improve the visual quality of image reconstruction, we further proposemultiple latent codesandscalable inversion. The former gets several latent codes in the inversion, while the latter additionally compresses and transmits a shallow compact feature to support visual reconstruction. Experimental results demonstrate the superiority of our method in both human vision tasks,i.e. image reconstruction, and machine vision tasks, including semantic parsing and attribute prediction.

Enhancing low-light videos in a supervised style presents a set of challenges, including limited data diversity, misalignment, and the domain gap introduced through the dataset construction pipeline. Our paper tackles these challenges by constructing a self-learned enhancement approach that gets rid of the reliance on any external training data. The challenge of self-supervised learning lies in fitting high-quality signal representations solely from input signals. Our work designs a bottleneck neural representation mechanism that extracts those signals. More in detail, we encode the frame-wise representation with a compact deep embedding and utilize a neural network to parameterize the video-level manifold consistently. Then, an entropy constraint is applied to the enhanced results based on the adjacent spatial-temporal context to filter out the degraded visual signals, e.g. noise and frame inconsistency. Last, a novel Chromatic Retinex decomposition is proposed to effectively align the reflectance distribution temporally. It benefits the entropy control on different components of each frame and facilitates noise-to-noise training, successfully suppressing the temporal flicker. Extensive experiments demonstrate the robustness and superior effectiveness of our proposed method. Our project is publicly available at: https://huangerbai.github.io/SLBNR/.

As an emerging research practice leveraging recent advanced AI techniques, e.g. deep models based prediction and generation, Video Coding for Machines (VCM) is committed to bridging to an extent separate research tracks of video/image compression and feature compression, and attempts to optimize compactness and efficiency jointly from a unified perspective of high accuracy machine vision and full fidelity human vision. With the rapid advances of deep feature representation and visual data compression in mind, in this paper, we summarize VCM methodology and philosophy based on existing academia and industrial efforts. The development of VCM follows a general rate-distortion optimization, and the categorization of key modules or techniques is established including feature-assisted coding, scalable coding, intermediate feature compression/optimization, and machine vision targeted codec, from broader perspectives of vision tasks, analytics resources, etc. From previous works, it is demonstrated that, although existing works attempt to reveal the nature of scalable representation in bits when dealing with machine and human vision tasks, there remains a rare study in the generality of low bit rate representation, and accordingly how to support a variety of visual analytic tasks. Therefore, we investigate a novel visual information compression for the analytics taxonomy problem to strengthen the capability of compact visual representations extracted from multiple tasks for visual analytics. A new perspective of task relationships versus compression is revisited. By keeping in mind the transferability among different machine vision tasks (e.g. high-level semantic and mid-level geometry-related), we aim to support multiple tasks jointly at low bit rates. In particular, to narrow the dimensionality gap between neural network generated features extracted from pixels and a variety of machine vision features/labels (e.g. scene class, segmentation labels), a codebook hyperprior is designed to compress the neural network-generated features. As demonstrated in our experiments, this new hyperprior model is expected to improve feature compression efficiency by estimating the signal entropy more accurately, which enables further investigation of the granularity of abstracting compact features among different tasks.

Photography in low-light conditions suffers from dense noise and insufficient light. Flash photography, introducing extra light sources, performs better at suppressing noise and revealing details, while being interrupted by unnatural ambient illumination. This paper offers an analysis of the pros and cons to utilize low-light and flash images for enhancement, which inspires us to design a unified sample-adaptive CNN to capture diverse focuses from different inputs in a complementary way. Specifically, a Flash Compensated Dynamic Filtering Network is proposed to utilize the revealed details of flash images to compensate for fine structure reconstruction in low-light enhancement. To adaptively fuse information from misaligned low-light and flash image pairs, our network is designed with three distinctive features. Firstly, we adopt a layer-wise regression strategy, where results are predicted from the single input first and then fused to sufficiently leverage complementary information. Secondly, we employ a sample-adaptive mechanism, where each pixel is estimated with its distinctive parameters augmented by weighted residual connections. Finally, we utilize a coarse-to-fine architecture, where features are extracted by diversified receptive fields to utilize hierarchical contextual information. Experimental results demonstrate that the three design principles lead to the significant superiority of the proposed method over state-of-the-art methods.

Machine intelligence systems have been increasingly widely deployed in real-world circumstances, while the conventional human-vision oriented video coding schemes are inefficient to be embedded in large-scale systems and further support a wide range of applications. There have been urgent demands for a new generation of compression framework to efficiently encodes visual data, where the compression and analytics for machine vision and human perception can be jointly optimized. To this end, we propose a novel visual compression framework to provide visual contents with different granularity for both human and machine vision tasks collaboratively. The proposed scalable compression framework maintains the critical semantic information in a basic layer, so that it is capable of supporting the accurate machine vision analysis under a tight bit-rate constraint. It is scalable to provide visual representations of different granularity to support various kinds of tasks, including video reconstruction that serves human vision examination. Experimental results on the human-centered videos have demonstrated the promising functionality of scalable visual coding with improved efficiency for high-performance machine analysis and human perception.

Low light conditions not only degrade human visual experience, but also reduce the performance of downstream machine analytics. Although many works have been designed for low-light enhancement or domain adaptive machine analytics, the former considers less on high-level vision, while the latter neglects the potential of image-level signal adjustment. How to restore underexposed images/videos from the perspective of machine vision has long been overlooked. In this paper, we are the first to propose a learnable illumination enhancement model for high-level vision. Inspired by real camera response functions, we assume that the illumination enhancement function should be a concave curve, and propose to satisfy this concavity through discrete integral. With the intention of adapting illumination from the perspective of machine vision without task-specific annotated data, we design an asymmetric cross-domain self-supervised training strategy. Our model architecture and training designs mutually benefit each other, forming a powerful unsupervised normal-to-low light adaptation framework. Comprehensive experiments demonstrate that our method surpasses existing low-light enhancement and adaptation methods and shows superior generalization on various low-light vision tasks, including classification, detection, action recognition, and optical flow estimation. All of our data, code, and results will be available online upon publication of the paper.

In this paper, we make the first benchmark effort to elaborate on the superiority of using RAW images in the low light enhancement and develop a novel alternative route to utilize RAW images in a more flexible and practical way. Inspired by a full consideration on the typical image processing pipeline, we are inspired to develop a new evaluation framework, Factorized Enhancement Model (FEM), which decomposes the properties of RAW images into measurable factors and provides a tool for exploring how properties of RAW images affect the enhancement performance empirically. The empirical benchmark results show that the Linearity of data and Exposure Time recorded in meta-data play the most critical role, which brings distinct performance gains in various measures over the approaches taking the sRGB images as input. With the insights obtained from the benchmark results in mind, a RAW-guiding Exposure Enhancement Network (REENet) is developed, which makes trade-offs between the advantages and inaccessibility of RAW images in real applications in a way of using RAW images only in the training phase. REENet projects sRGB images into linear RAW domains to apply constraints with corresponding RAW images to reduce the difficulty of modeling training. After that, in the testing phase, our REENet does not rely on RAW images. Experimental results demonstrate not only the superiority of REENet to state-of-the-art sRGB-based methods and but also the effectiveness of the RAW guidance and all components.

Due to the absence of a desirable objective for low-light image enhancement, previous data-driven methods may provide undesirable enhanced results including amplified noise, degraded contrast and biased colors. In this work, inspired by Retinex theory, we design an end-to-end signal prior-guided layer separation and data-driven mapping network with layer-specified constraints for single-image low-light enhancement. A Sparse Gradient Minimization sub-Network (SGM-Net) is constructed to remove the low-amplitude structures and preserve major edge information, which facilitates extracting paired illumination maps of low/normal-light images. After the learned decomposition, two sub-networks (Enhance-Net and Restore-Net) are utilized to predict the enhanced illumination and reflectance maps, respectively, which helps stretch the contrast of the illumination map and remove intensive noise in the reflectance map. The effects of all these configured constraints, including the signal structure regularization and losses, combine together reciprocally, which leads to good reconstruction results in overall visual quality. The evaluation on both synthetic and real images, particularly on those containing intensive noise, compression artifacts and their interleaved artifacts, shows the effectiveness of our novel models, which significantly outperforms the state-of-the-art methods.