Xianbiao Qi

Yukun Huang, Jianan Wang, Yukai Shi et al. · microsoft-research

Text-to-image diffusion models pre-trained on billions of image-text pairs have recently enabled 3D content creation by optimizing a randomly initialized differentiable 3D representation with score distillation. However, the optimization process suffers slow convergence and the resultant 3D models often exhibit two limitations: (a) quality concerns such as missing attributes and distorted shape and texture; (b) extremely low diversity comparing to text-guided image synthesis. In this paper, we show that the conflict between the 3D optimization process and uniform timestep sampling in score distillation is the main reason for these limitations. To resolve this conflict, we propose to prioritize timestep sampling with monotonically non-increasing functions, which aligns the 3D optimization process with the sampling process of diffusion model. Extensive experiments show that our simple redesign significantly improves 3D content creation with faster convergence, better quality and diversity.

Tianhe Ren, Shilong Liu, Feng Li et al.

The DEtection TRansformer (DETR) algorithm has received considerable attention in the research community and is gradually emerging as a mainstream approach for object detection and other perception tasks. However, the current field lacks a unified and comprehensive benchmark specifically tailored for DETR-based models. To address this issue, we develop a unified, highly modular, and lightweight codebase called detrex, which supports a majority of the mainstream DETR-based instance recognition algorithms, covering various fundamental tasks, including object detection, segmentation, and pose estimation. We conduct extensive experiments under detrex and perform a comprehensive benchmark for DETR-based models. Moreover, we enhance the performance of detection transformers through the refinement of training hyper-parameters, providing strong baselines for supported algorithms.We hope that detrex could offer research communities a standardized and unified platform to evaluate and compare different DETR-based models while fostering a deeper understanding and driving advancements in DETR-based instance recognition. Our code is available at https://github.com/IDEA-Research/detrex. The project is currently being actively developed. We encourage the community to use detrex codebase for further development and contributions.

Xianbiao Qi, Jianan Wang, Yihao Chen et al.

We present a Lipschitz continuous Transformer, called LipsFormer, to pursue training stability both theoretically and empirically for Transformer-based models. In contrast to previous practical tricks that address training instability by learning rate warmup, layer normalization, attention formulation, and weight initialization, we show that Lipschitz continuity is a more essential property to ensure training stability. In LipsFormer, we replace unstable Transformer component modules with Lipschitz continuous counterparts: CenterNorm instead of LayerNorm, spectral initialization instead of Xavier initialization, scaled cosine similarity attention instead of dot-product attention, and weighted residual shortcut. We prove that these introduced modules are Lipschitz continuous and derive an upper bound on the Lipschitz constant of LipsFormer. Our experiments show that LipsFormer allows stable training of deep Transformer architectures without the need of careful learning rate tuning such as warmup, yielding a faster convergence and better generalization. As a result, on the ImageNet 1K dataset, LipsFormer-Swin-Tiny based on Swin Transformer training for 300 epochs can obtain 82.7\% without any learning rate warmup. Moreover, LipsFormer-CSwin-Tiny, based on CSwin, training for 300 epochs achieves a top-1 accuracy of 83.5\% with 4.7G FLOPs and 24M parameters. The code will be released at \url{https://github.com/IDEA-Research/LipsFormer}.

Yihao Chen, Xianbiao Qi, Jianan Wang et al.

We propose DisCo-CLIP, a distributed memory-efficient CLIP training approach, to reduce the memory consumption of contrastive loss when training contrastive learning models. Our approach decomposes the contrastive loss and its gradient computation into two parts, one to calculate the intra-GPU gradients and the other to compute the inter-GPU gradients. According to our decomposition, only the intra-GPU gradients are computed on the current GPU, while the inter-GPU gradients are collected via all_reduce from other GPUs instead of being repeatedly computed on every GPU. In this way, we can reduce the GPU memory consumption of contrastive loss computation from $\bigO(B^2)$ to $\bigO(\frac{B^2}{N})$, where $B$ and $N$ are the batch size and the number of GPUs used for training. Such a distributed solution is mathematically equivalent to the original non-distributed contrastive loss computation, without sacrificing any computation accuracy. It is particularly efficient for large-batch CLIP training. For instance, DisCo-CLIP can enable contrastive training of a ViT-B/32 model with a batch size of 32K or 196K using 8 or 64 A100 40GB GPUs, compared with the original CLIP solution which requires 128 A100 40GB GPUs to train a ViT-B/32 model with a batch size of 32K. The code will be released at https://github.com/IDEA-Research/DisCo-CLIP

He Cao, Jianan Wang, Tianhe Ren et al.

Score-based diffusion models have captured widespread attention and funded fast progress of recent vision generative tasks. In this paper, we focus on diffusion model backbone which has been much neglected before. We systematically explore vision Transformers as diffusion learners for various generative tasks. With our improvements the performance of vanilla ViT-based backbone (IU-ViT) is boosted to be on par with traditional U-Net-based methods. We further provide a hypothesis on the implication of disentangling the generative backbone as an encoder-decoder structure and show proof-of-concept experiments verifying the effectiveness of a stronger encoder for generative tasks with ASymmetriC ENcoder Decoder (ASCEND). Our improvements achieve competitive results on CIFAR-10, CelebA, LSUN, CUB Bird and large-resolution text-to-image tasks. To the best of our knowledge, we are the first to successfully train a single diffusion model on text-to-image task beyond 64x64 resolution. We hope this will motivate people to rethink the modeling choices and the training pipelines for diffusion-based generative models.

Bojia Zi, Xianbiao Qi, Lingzhi Wang et al.

In this paper, we present Delta-LoRA, which is a novel parameter-efficient approach to fine-tune large language models (LLMs). In contrast to LoRA and other low-rank adaptation methods such as AdaLoRA, Delta-LoRA not only updates the low-rank matrices $\bA$ and $\bB$, but also propagate the learning to the pre-trained weights $\bW$ via updates utilizing the delta of the product of two low-rank matrices ($\bA^{(t+1)}\bB^{(t+1)} - \bA^{(t)}\bB^{(t)}$). Such a strategy effectively addresses the limitation that the incremental update of low-rank matrices is inadequate for learning representations capable for downstream tasks. Moreover, as the update of $\bW$ does not need to compute the gradients of $\bW$ and store their momentums, Delta-LoRA shares comparable memory requirements and computational costs with LoRA. Extensive experiments show that Delta-LoRA significantly outperforms existing low-rank adaptation methods. We further support these results with comprehensive analyses that underscore the effectiveness of Delta-LoRA.

Youze Huang, Penghui Ruan, Bojia Zi et al.

Recent advances in diffusion models have brought remarkable progress in image and video editing, yet some tasks remain underexplored. In this paper, we introduce a new task, Object Retexture, which transfers local textures from a reference object to a target object in images or videos. To perform this task, a straightforward solution is to use ControlNet conditioned on the source structure and the reference texture. However, this approach suffers from limited controllability for two reasons: conditioning on the raw reference image introduces unwanted structural information, and it fails to disentangle the visual texture and structure information of the source. To address this problem, we propose Refaçade, a method that consists of two key designs to achieve precise and controllable texture transfer in both images and videos. First, we employ a texture remover trained on paired textured/untextured 3D mesh renderings to remove appearance information while preserving the geometry and motion of source videos. Second, we disrupt the reference global layout using a jigsaw permutation, encouraging the model to focus on local texture statistics rather than the global layout of the object. Extensive experiments demonstrate superior visual quality, precise editing, and controllability, outperforming strong baselines in both quantitative and human evaluations. Code is available at https://github.com/fishZe233/Refacade.

Yukai Shi, Jianan Wang, He Cao et al.

In this paper, we present TOSS, which introduces text to the task of novel view synthesis (NVS) from just a single RGB image. While Zero-1-to-3 has demonstrated impressive zero-shot open-set NVS capability, it treats NVS as a pure image-to-image translation problem. This approach suffers from the challengingly under-constrained nature of single-view NVS: the process lacks means of explicit user control and often results in implausible NVS generations. To address this limitation, TOSS uses text as high-level semantic information to constrain the NVS solution space. TOSS fine-tunes text-to-image Stable Diffusion pre-trained on large-scale text-image pairs and introduces modules specifically tailored to image and camera pose conditioning, as well as dedicated training for pose correctness and preservation of fine details. Comprehensive experiments are conducted with results showing that our proposed TOSS outperforms Zero-1-to-3 with more plausible, controllable and multiview-consistent NVS results. We further support these results with comprehensive ablations that underscore the effectiveness and potential of the introduced semantic guidance and architecture design.

Xianbiao Qi, Marco Chen, Jiaquan Ye et al.

The Muon optimizer has recently attracted considerable attention for its strong empirical performance and use of orthogonalized updates on matrix-shaped parameters, yet its underlying mechanisms and relationship to adaptive optimizers such as Adam remain insufficiently understood. In this work, we aim to address these questions through a unified spectral perspective. Specifically, we view Muon as the p = 0 endpoint of a family of spectral transformations of the form U \boldsymbolΣ^{p} V' , and consider additional variants with p = 1/2 , p = 1/4 , and p = 1 . These transformations are applied to both first-moment updates, as in momentum SGD, and to root-mean-square (RMS) normalized gradient updates as in Adam. To enable efficient computation, we develop a coupled Newton iteration that avoids explicit singular value decomposition. Across controlled experiments, we find that RMS-normalized updates yield more stable optimization than first-moment updates. Moreover, while spectral compression provides strong stabilization benefits under first-moment updates, the Muon update (p = 0) does not consistently outperform Adam. These results suggest that Muon is best understood as an effective form of spectral normalization, but not a universally superior optimization method. Our source code will be released at https://github.com/Ocram7/BeyondMuon.

Shufa Wei, Xiaolong Xu, Xianbiao Qi et al.

Large Language Models (LLMs) have demonstrated exceptional capabilities across various natural language processing tasks. Yet, many of these advanced LLMs are tailored for broad, general-purpose applications. In this technical report, we introduce AcademicGPT, designed specifically to empower academic research. AcademicGPT is a continual training model derived from LLaMA2-70B. Our training corpus mainly consists of academic papers, thesis, content from some academic domain, high-quality Chinese data and others. While it may not be extensive in data scale, AcademicGPT marks our initial venture into a domain-specific GPT tailored for research area. We evaluate AcademicGPT on several established public benchmarks such as MMLU and CEval, as well as on some specialized academic benchmarks like PubMedQA, SCIEval, and our newly-created ComputerScienceQA, to demonstrate its ability from general knowledge ability, to Chinese ability, and to academic ability. Building upon AcademicGPT's foundation model, we also developed several applications catered to the academic area, including General Academic Question Answering, AI-assisted Paper Reading, Paper Review, and AI-assisted Title and Abstract Generation.

Xianbiao Qi, Jianan Wang, Lei Zhang

This article provides a comprehensive understanding of optimization in deep learning, with a primary focus on the challenges of gradient vanishing and gradient exploding, which normally lead to diminished model representational ability and training instability, respectively. We analyze these two challenges through several strategic measures, including the improvement of gradient flow and the imposition of constraints on a network's Lipschitz constant. To help understand the current optimization methodologies, we categorize them into two classes: explicit optimization and implicit optimization. Explicit optimization methods involve direct manipulation of optimizer parameters, including weight, gradient, learning rate, and weight decay. Implicit optimization methods, by contrast, focus on improving the overall landscape of a network by enhancing its modules, such as residual shortcuts, normalization methods, attention mechanisms, and activations. In this article, we provide an in-depth analysis of these two optimization classes and undertake a thorough examination of the Jacobian matrices and the Lipschitz constants of many widely used deep learning modules, highlighting existing issues as well as potential improvements. Moreover, we also conduct a series of analytical experiments to substantiate our theoretical discussions. This article does not aim to propose a new optimizer or network. Rather, our intention is to present a comprehensive understanding of optimization in deep learning. We hope that this article will assist readers in gaining a deeper insight in this field and encourages the development of more robust, efficient, and high-performing models.

Xuantong Liu, Shaozhe Hao, Xianbiao Qi et al.

The success of autoregressive (AR) language models in text generation has inspired the computer vision community to adopt Large Language Models (LLMs) for image generation. However, considering the essential differences between text and image modalities, the design space of language models for image generation remains underexplored. We observe that image tokens exhibit greater randomness compared to text tokens, which presents challenges when training with token prediction. Nevertheless, AR models demonstrate their potential by effectively learning patterns even from a seemingly suboptimal optimization problem. Our analysis also reveals that while all models successfully grasp the importance of local information in image generation, smaller models struggle to capture the global context. In contrast, larger models showcase improved capabilities in this area, helping to explain the performance gains achieved when scaling up model size. We further elucidate the design space of language models for vision generation, including tokenizer choice, model choice, model scalability, vocabulary design, and sampling strategy through extensive comparative experiments. Our work is the first to analyze the optimization behavior of language models in vision generation, and we believe it can inspire more effective designs when applying LMs to other domains. Finally, our elucidated language model for image generation, termed as ELM, achieves state-of-the-art performance on the ImageNet 256*256 benchmark. The code is available at https://github.com/Pepperlll/LMforImageGeneration.git.

Wei He, Shangzhi Zhang, Chun-Guang Li et al.

Spectral clustering, as a popular tool for data clustering, requires an eigen-decomposition step on a given affinity to obtain the spectral embedding. Nevertheless, such a step suffers from the lack of generalizability and scalability. Moreover, the obtained spectral embeddings can hardly provide a good approximation to the ground-truth partition and thus a k-means step is adopted to quantize the embedding. In this paper, we propose a simple yet effective scalable and generalizable approach, called Neural Normalized Cut (NeuNcut), to learn the clustering membership for spectral clustering directly. In NeuNcut, we properly reparameterize the unknown cluster membership via a neural network, and train the neural network via stochastic gradient descent with a properly relaxed normalized cut loss. As a result, our NeuNcut enjoys a desired generalization ability to directly infer clustering membership for out-of-sample unseen data and hence brings us an efficient way to handle clustering task with ultra large-scale data. We conduct extensive experiments on both synthetic data and benchmark datasets and experimental results validate the effectiveness and the superiority of our approach. Our code is available at: https://github.com/hewei98/NeuNcut.

Xianghan Meng, Zhiyuan Huang, Wei He et al.

Subspace clustering is a classical unsupervised learning task, built on a basic assumption that high-dimensional data can be approximated by a union of subspaces (UoS). Nevertheless, the real-world data are often deviating from the UoS assumption. To address this challenge, state-of-the-art deep subspace clustering algorithms attempt to jointly learn UoS representations and self-expressive coefficients. However, the general framework of the existing algorithms suffers from a catastrophic feature collapse and lacks a theoretical guarantee to learn desired UoS representation. In this paper, we present a Principled fRamewOrk for Deep Subspace Clustering (PRO-DSC), which is designed to learn structured representations and self-expressive coefficients in a unified manner. Specifically, in PRO-DSC, we incorporate an effective regularization on the learned representations into the self-expressive model, prove that the regularized self-expressive model is able to prevent feature space collapse, and demonstrate that the learned optimal representations under certain condition lie on a union of orthogonal subspaces. Moreover, we provide a scalable and efficient approach to implement our PRO-DSC and conduct extensive experiments to verify our theoretical findings and demonstrate the superior performance of our proposed deep subspace clustering approach. The code is available at https://github.com/mengxianghan123/PRO-DSC.

Wei He, Xianghan Meng, Zhiyuan Huang et al.

Generalized Category Discovery (GCD) aims to identify both known and unknown categories, with only partial labels given for the known categories, posing a challenging open-set recognition problem. State-of-the-art approaches for GCD task are usually built on multi-modality representation learning, which is heavily dependent upon inter-modality alignment. However, few of them cast a proper intra-modality alignment to generate a desired underlying structure of representation distributions. In this paper, we propose a novel and effective multi-modal representation learning framework for GCD via Semi-Supervised Rate Reduction, called SSR$^2$-GCD, to learn cross-modality representations with desired structural properties based on emphasizing to properly align intra-modality relationships. Moreover, to boost knowledge transfer, we integrate prompt candidates by leveraging the inter-modal alignment offered by Vision Language Models. We conduct extensive experiments on generic and fine-grained benchmark datasets demonstrating superior performance of our approach.

Yaxuan Huang, Xili Dai, Jianan Wang et al.

Room layout estimation from multiple-perspective images is poorly investigated due to the complexities that emerge from multi-view geometry, which requires muti-step solutions such as camera intrinsic and extrinsic estimation, image matching, and triangulation. However, in 3D reconstruction, the advancement of recent 3D foundation models such as DUSt3R has shifted the paradigm from the traditional multi-step structure-from-motion process to an end-to-end single-step approach. To this end, we introduce Plane-DUSt3R, a novel method for multi-view room layout estimation leveraging the 3D foundation model DUSt3R. Plane-DUSt3R incorporates the DUSt3R framework and fine-tunes on a room layout dataset (Structure3D) with a modified objective to estimate structural planes. By generating uniform and parsimonious results, Plane-DUSt3R enables room layout estimation with only a single post-processing step and 2D detection results. Unlike previous methods that rely on single-perspective or panorama image, Plane-DUSt3R extends the setting to handle multiple-perspective images. Moreover, it offers a streamlined, end-to-end solution that simplifies the process and reduces error accumulation. Experimental results demonstrate that Plane-DUSt3R not only outperforms state-of-the-art methods on the synthetic dataset but also proves robust and effective on in the wild data with different image styles such as cartoon. Our code is available at: https://github.com/justacar/Plane-DUSt3R

Marco Chen, Xianbiao Qi, Yelin He et al.

In this work, we revisit Transformer optimization through the lens of second-order geometry and establish a direct connection between architectural design, activation scale, the Hessian matrix, and the maximum tolerable learning rate. We introduce a simple normalization strategy, termed SimpleNorm, which stabilizes intermediate activation scales by construction. Then, by analyzing the Hessian of the loss with respect to network activations, we theoretically show that SimpleNorm significantly reduces the spectral norm of the Hessian, thereby permitting larger stable learning rates. We validate our theoretical findings through extensive experiments on large GPT models at parameter scales 1B, 1.4B, 7B and 8B. Empirically, SimpleGPT, our SimpleNorm-based network, tolerates learning rates 3$\times$-10$\times$ larger than standard convention, consistently demonstrates strong optimization stability, and achieves substantially better performance than well-established baselines. Specifically, when training 7B-scale models for 60K steps, SimpleGPT achieves a training loss that is 0.08 lower than that of LLaMA2 with QKNorm, reducing the loss from 2.290 to 2.208. Our source code will be released at https://github.com/Ocram7/SimpleGPT.

Shilong Liu, Feng Li, Hao Zhang et al.

We present in this paper a novel query formulation using dynamic anchor boxes for DETR (DEtection TRansformer) and offer a deeper understanding of the role of queries in DETR. This new formulation directly uses box coordinates as queries in Transformer decoders and dynamically updates them layer-by-layer. Using box coordinates not only helps using explicit positional priors to improve the query-to-feature similarity and eliminate the slow training convergence issue in DETR, but also allows us to modulate the positional attention map using the box width and height information. Such a design makes it clear that queries in DETR can be implemented as performing soft ROI pooling layer-by-layer in a cascade manner. As a result, it leads to the best performance on MS-COCO benchmark among the DETR-like detection models under the same setting, e.g., AP 45.7\% using ResNet50-DC5 as backbone trained in 50 epochs. We also conducted extensive experiments to confirm our analysis and verify the effectiveness of our methods. Code is available at \url{https://github.com/SlongLiu/DAB-DETR}.

Yihao Chen, Xin Tang, Xianbiao Qi et al.

Graph Neural Networks (GNNs) have attracted considerable attention and have emerged as a new promising paradigm to process graph-structured data. GNNs are usually stacked to multiple layers and the node representations in each layer are computed through propagating and aggregating the neighboring node features with respect to the graph. By stacking to multiple layers, GNNs are able to capture the long-range dependencies among the data on the graph and thus bring performance improvements. To train a GNN with multiple layers effectively, some normalization techniques (e.g., node-wise normalization, batch-wise normalization) are necessary. However, the normalization techniques for GNNs are highly task-relevant and different application tasks prefer to different normalization techniques, which is hard to know in advance. To tackle this deficiency, in this paper, we propose to learn graph normalization by optimizing a weighted combination of normalization techniques at four different levels, including node-wise normalization, adjacency-wise normalization, graph-wise normalization, and batch-wise normalization, in which the adjacency-wise normalization and the graph-wise normalization are newly proposed in this paper to take into account the local structure and the global structure on the graph, respectively. By learning the optimal weights, we are able to automatically select a single best or a best combination of multiple normalizations for a specific task. We conduct extensive experiments on benchmark datasets for different tasks, including node classification, link prediction, graph classification and graph regression, and confirm that the learned graph normalization leads to competitive results and that the learned weights suggest the appropriate normalization techniques for the specific task. Source code is released here https://github.com/cyh1112/GraphNormalization.

Wenwen Yu, Ning Lu, Xianbiao Qi et al.

Computer vision with state-of-the-art deep learning models has achieved huge success in the field of Optical Character Recognition (OCR) including text detection and recognition tasks recently. However, Key Information Extraction (KIE) from documents as the downstream task of OCR, having a large number of use scenarios in real-world, remains a challenge because documents not only have textual features extracting from OCR systems but also have semantic visual features that are not fully exploited and play a critical role in KIE. Too little work has been devoted to efficiently make full use of both textual and visual features of the documents. In this paper, we introduce PICK, a framework that is effective and robust in handling complex documents layout for KIE by combining graph learning with graph convolution operation, yielding a richer semantic representation containing the textual and visual features and global layout without ambiguity. Extensive experiments on real-world datasets have been conducted to show that our method outperforms baselines methods by significant margins. Our code is available at https://github.com/wenwenyu/PICK-pytorch.

Ning Lu, Wenwen Yu, Xianbiao Qi et al.

Attention-based scene text recognizers have gained huge success, which leverages a more compact intermediate representation to learn 1d- or 2d- attention by a RNN-based encoder-decoder architecture. However, such methods suffer from attention-drift problem because high similarity among encoded features leads to attention confusion under the RNN-based local attention mechanism. Moreover, RNN-based methods have low efficiency due to poor parallelization. To overcome these problems, we propose the MASTER, a self-attention based scene text recognizer that (1) not only encodes the input-output attention but also learns self-attention which encodes feature-feature and target-target relationships inside the encoder and decoder and (2) learns a more powerful and robust intermediate representation to spatial distortion, and (3) owns a great training efficiency because of high training parallelization and a high-speed inference because of an efficient memory-cache mechanism. Extensive experiments on various benchmarks demonstrate the superior performance of our MASTER on both regular and irregular scene text. Pytorch code can be found at https://github.com/wenwenyu/MASTER-pytorch, and Tensorflow code can be found at https://github.com/jiangxiluning/MASTER-TF.

Bojia Zi, Shihao Zhao, Xianbiao Qi et al.

Recent advancements in video generation have been remarkable, yet many existing methods struggle with issues of consistency and poor text-video alignment. Moreover, the field lacks effective techniques for text-guided video inpainting, a stark contrast to the well-explored domain of text-guided image inpainting. To this end, this paper proposes a novel text-guided video inpainting model that achieves better consistency, controllability and compatibility. Specifically, we introduce a simple but efficient motion capture module to preserve motion consistency, and design an instance-aware region selection instead of a random region selection to obtain better textual controllability, and utilize a novel strategy to inject some personalized models into our CoCoCo model and thus obtain better model compatibility. Extensive experiments show that our model can generate high-quality video clips. Meanwhile, our model shows better motion consistency, textual controllability and model compatibility. More details are shown in [cococozibojia.github.io](cococozibojia.github.io).

Bojia Zi, Penghui Ruan, Marco Chen et al.

Recent advancements in video generation have spurred the development of video editing techniques, which can be divided into inversion-based and end-to-end methods. However, current video editing methods still suffer from several challenges. Inversion-based methods, though training-free and flexible, are time-consuming during inference, struggle with fine-grained editing instructions, and produce artifacts and jitter. On the other hand, end-to-end methods, which rely on edited video pairs for training, offer faster inference speeds but often produce poor editing results due to a lack of high-quality training video pairs. In this paper, to close the gap in end-to-end methods, we introduce Señorita-2M, a high-quality video editing dataset. Señorita-2M consists of approximately 2 millions of video editing pairs. It is built by crafting four high-quality, specialized video editing models, each crafted and trained by our team to achieve state-of-the-art editing results. We also propose a filtering pipeline to eliminate poorly edited video pairs. Furthermore, we explore common video editing architectures to identify the most effective structure based on current pre-trained generative model. Extensive experiments show that our dataset can help to yield remarkably high-quality video editing results. More details are available at https://senorita-2m-dataset.github.io.

Bojia Zi, Weixuan Peng, Xianbiao Qi et al.

Recent advances in video diffusion models have driven rapid progress in video editing techniques. However, video object removal, a critical subtask of video editing, remains challenging due to issues such as hallucinated objects and visual artifacts. Furthermore, existing methods often rely on computationally expensive sampling procedures and classifier-free guidance (CFG), resulting in slow inference. To address these limitations, we propose MiniMax-Remover, a novel two-stage video object removal approach. Motivated by the observation that text condition is not best suited for this task, we simplify the pretrained video generation model by removing textual input and cross-attention layers, resulting in a more lightweight and efficient model architecture in the first stage. In the second stage, we distilled our remover on successful videos produced by the stage-1 model and curated by human annotators, using a minimax optimization strategy to further improve editing quality and inference speed. Specifically, the inner maximization identifies adversarial input noise ("bad noise") that makes failure removals, while the outer minimization step trains the model to generate high-quality removal results even under such challenging conditions. As a result, our method achieves a state-of-the-art video object removal results with as few as 6 sampling steps and doesn't rely on CFG, significantly improving inference efficiency. Extensive experiments demonstrate the effectiveness and superiority of MiniMax-Remover compared to existing methods. Codes and Videos are available at: https://minimax-remover.github.io.

Shaozhe Hao, Xuantong Liu, Xianbiao Qi et al.

We introduce BiGR, a novel conditional image generation model using compact binary latent codes for generative training, focusing on enhancing both generation and representation capabilities. BiGR is the first conditional generative model that unifies generation and discrimination within the same framework. BiGR features a binary tokenizer, a masked modeling mechanism, and a binary transcoder for binary code prediction. Additionally, we introduce a novel entropy-ordered sampling method to enable efficient image generation. Extensive experiments validate BiGR's superior performance in generation quality, as measured by FID-50k, and representation capabilities, as evidenced by linear-probe accuracy. Moreover, BiGR showcases zero-shot generalization across various vision tasks, enabling applications such as image inpainting, outpainting, editing, interpolation, and enrichment, without the need for structural modifications. Our findings suggest that BiGR unifies generative and discriminative tasks effectively, paving the way for further advancements in the field. We further enable BiGR to perform text-to-image generation, showcasing its potential for broader applications.

Xianbiao Qi, Yelin He, Jiaquan Ye et al.

Scaling Transformer to a large scale without using some technical tricks such as learning rate warump and using an obviously lower learning rate is an extremely challenging task, and is increasingly gaining more attention. In this paper, we provide a theoretical analysis for the process of training Transformer and reveal the rationale behind the model crash phenomenon in the training process, termed \textit{spectral energy concentration} of ${\bW_q}^{\top} \bW_k$, which is the reason for a malignant entropy collapse, where ${\bW_q}$ and $\bW_k$ are the projection matrices for the query and the key in Transformer, respectively. To remedy this problem, motivated by \textit{Weyl's Inequality}, we present a novel optimization strategy, \ie, making the weight updating in successive steps smooth -- if the ratio $\frac{σ_{1}(\nabla \bW_t)}{σ_{1}(\bW_{t-1})}$ is larger than a threshold, we will automatically bound the learning rate to a weighted multiple of $\frac{σ_{1}(\bW_{t-1})}{σ_{1}(\nabla \bW_t)}$, where $\nabla \bW_t$ is the updating quantity in step $t$. Such an optimization strategy can prevent spectral energy concentration to only a few directions, and thus can avoid malignant entropy collapse which will trigger the model crash. We conduct extensive experiments using ViT, Swin-Transformer and GPT, showing that our optimization strategy can effectively and stably train these Transformers without using learning rate warmup.

Penghui Ruan, Bojia Zi, Xianbiao Qi et al.

Object-level manipulation, relocating or reorienting objects in images or videos while preserving scene realism, is central to film post-production, AR, and creative editing. Yet existing methods struggle to jointly achieve three core goals: background preservation, geometric consistency under viewpoint shifts, and user-controllable transformations. Geometry-based approaches offer precise control but require explicit 3D reconstruction and generalize poorly; diffusion-based methods generalize better but lack fine-grained geometric control. We present Ctrl&Shift, an end-to-end diffusion framework to achieve geometry-consistent object manipulation without explicit 3D representations. Our key insight is to decompose manipulation into two stages, object removal and reference-guided inpainting under explicit camera pose control, and encode both within a unified diffusion process. To enable precise, disentangled control, we design a multi-task, multi-stage training strategy that separates background, identity, and pose signals across tasks. To improve generalization, we introduce a scalable real-world dataset construction pipeline that generates paired image and video samples with estimated relative camera poses. Extensive experiments demonstrate that Ctrl&Shift achieves state-of-the-art results in fidelity, viewpoint consistency, and controllability. To our knowledge, this is the first framework to unify fine-grained geometric control and real-world generalization for object manipulation, without relying on any explicit 3D modeling.

Linchun Wu, Qin Zou, Xianbiao Qi et al.

Visual anomaly inspection is critical in manufacturing, yet hampered by the scarcity of real anomaly samples for training robust detectors. Synthetic data generation presents a viable strategy for data augmentation; however, current methods remain constrained by two principal limitations: 1) the generation of anomalies that are structurally inconsistent with the normal background, and 2) the presence of undesirable feature entanglement between synthesized images and their corresponding annotation masks, which undermines the perceptual realism of the output. This paper introduces Double Helix Diffusion (DH-Diff), a novel cross-domain generative framework designed to simultaneously synthesize high-fidelity anomaly images and their pixel-level annotation masks, explicitly addressing these challenges. DH-Diff employs a unique architecture inspired by a double helix, cycling through distinct modules for feature separation, connection, and merging. Specifically, a domain-decoupled attention mechanism mitigates feature entanglement by enhancing image and annotation features independently, and meanwhile a semantic score map alignment module ensures structural authenticity by coherently integrating anomaly foregrounds. DH-Diff offers flexible control via text prompts and optional graphical guidance. Extensive experiments demonstrate that DH-Diff significantly outperforms state-of-the-art methods in diversity and authenticity, leading to significant improvements in downstream anomaly detection performance.

Dongquan Yang, Yifan Yang, Xiaotian Yu et al.

Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.

Xianbiao Qi, Marco Chen, Wenjie Xiao et al.

Transformers have become the de facto backbone of modern deep learning, yet their training typically demands an advanced optimizer with adaptive learning rate like AdamW, rather than a momentum SGDW (mSGDW). Previous works show that it is mainly due to a heavy-tailed distribution of the gradients. In this paper, we introduce a Deeply Normalized Transformer (DNT), which is meticulously engineered to overcome this limitation enabling seamless training with vanilla mSGDW while yielding comparable performance to the Transformers trained via AdamW. To be specific, in DNT, we strategically integrate normalization techniques at proper positions in the Transformers to effectively modulate the Jacobian matrices of each layer, balance the influence of weights, activations, and their interactions, and thus enable the distributions of gradients concentrated. We provide both theoretical justifications of the normalization technique used in our DNT and extensive empirical evaluation on two popular Transformer architectures to validate that: a) DNT outperforms its counterparts (\ie, ViT and GPT), and b) DNT can be effectively trained with vanilla mSGDW.

Jiahao Lin, Weixuan Peng, Bojia Zi et al.

Recent advances in deep generative models have led to significant progress in video generation, yet the fidelity of AI-generated videos remains limited. Synthesized content often exhibits visual artifacts such as temporally inconsistent motion, physically implausible trajectories, unnatural object deformations, and local blurring that undermine realism and user trust. Accurate detection and spatial localization of these artifacts are crucial for both automated quality control and for guiding the development of improved generative models. However, the research community currently lacks a comprehensive benchmark specifically designed for artifact localization in AI generated videos. Existing datasets either restrict themselves to video or frame level detection or lack the fine-grained spatial annotations necessary for evaluating localization methods. To address this gap, we introduce BrokenVideos, a benchmark dataset of 3,254 AI-generated videos with meticulously annotated, pixel-level masks highlighting regions of visual corruption. Each annotation is validated through detailed human inspection to ensure high quality ground truth. Our experiments show that training state of the art artifact detection models and multi modal large language models (MLLMs) on BrokenVideos significantly improves their ability to localize corrupted regions. Through extensive evaluation, we demonstrate that BrokenVideos establishes a critical foundation for benchmarking and advancing research on artifact localization in generative video models. The dataset is available at: https://broken-video-detection-datetsets.github.io/Broken-Video-Detection-Datasets.github.io/.

Yukun Huang, Jianan Wang, Ailing Zeng et al.

We present DreamWaltz, a novel framework for generating and animating complex 3D avatars given text guidance and parametric human body prior. While recent methods have shown encouraging results for text-to-3D generation of common objects, creating high-quality and animatable 3D avatars remains challenging. To create high-quality 3D avatars, DreamWaltz proposes 3D-consistent occlusion-aware Score Distillation Sampling (SDS) to optimize implicit neural representations with canonical poses. It provides view-aligned supervision via 3D-aware skeleton conditioning which enables complex avatar generation without artifacts and multiple faces. For animation, our method learns an animatable 3D avatar representation from abundant image priors of diffusion model conditioned on various poses, which could animate complex non-rigged avatars given arbitrary poses without retraining. Extensive evaluations demonstrate that DreamWaltz is an effective and robust approach for creating 3D avatars that can take on complex shapes and appearances as well as novel poses for animation. The proposed framework further enables the creation of complex scenes with diverse compositions, including avatar-avatar, avatar-object and avatar-scene interactions. See https://dreamwaltz3d.github.io/ for more vivid 3D avatar and animation results.

Yuxiang Zhong, Xianbiao Qi, Shanjun Li et al.

In this technical report, we present our 1st place solution for the ICDAR 2021 competition on mathematical formula detection (MFD). The MFD task has three key challenges including a large scale span, large variation of the ratio between height and width, and rich character set and mathematical expressions. Considering these challenges, we used Generalized Focal Loss (GFL), an anchor-free method, instead of the anchor-based method, and prove the Adaptive Training Sampling Strategy (ATSS) and proper Feature Pyramid Network (FPN) can well solve the important issue of scale variation. Meanwhile, we also found some tricks, e.g., Deformable Convolution Network (DCN), SyncBN, and Weighted Box Fusion (WBF), were effective in MFD task. Our proposed method ranked 1st in the final 15 teams.

Yixuan Qiao, Hao Chen, Jun Wang et al.

TextVQA requires models to read and reason about text in images to answer questions about them. Specifically, models need to incorporate a new modality of text present in the images and reason over it to answer TextVQA questions. In this challenge, we use generative model T5 for TextVQA task. Based on pre-trained checkpoint T5-3B from HuggingFace repository, two other pre-training tasks including masked language modeling(MLM) and relative position prediction(RPP) are designed to better align object feature and scene text. In the stage of pre-training, encoder is dedicate to handle the fusion among multiple modalities: question text, object text labels, scene text labels, object visual features, scene visual features. After that decoder generates the text sequence step-by-step, cross entropy loss is required by default. We use a large-scale scene text dataset in pre-training and then fine-tune the T5-3B with the TextVQA dataset only.

Jiaquan Ye, Xianbiao Qi, Yelin He et al.

This paper presents our solution for ICDAR 2021 competition on scientific literature parsing taskB: table recognition to HTML. In our method, we divide the table content recognition task into foursub-tasks: table structure recognition, text line detection, text line recognition, and box assignment.Our table structure recognition algorithm is customized based on MASTER [1], a robust image textrecognition algorithm. PSENet [2] is used to detect each text line in the table image. For text linerecognition, our model is also built on MASTER. Finally, in the box assignment phase, we associatedthe text boxes detected by PSENet with the structure item reconstructed by table structure prediction,and fill the recognized content of the text line into the corresponding item. Our proposed methodachieves a 96.84% TEDS score on 9,115 validation samples in the development phase, and a 96.32%TEDS score on 9,064 samples in the final evaluation phase.

Yelin He, Xianbiao Qi, Jiaquan Ye et al.

This paper presents our solution for the ICDAR 2021 Competition on Scientific Table Image Recognition to LaTeX. This competition has two sub-tasks: Table Structure Reconstruction (TSR) and Table Content Reconstruction (TCR). We treat both sub-tasks as two individual image-to-sequence recognition problems. We leverage our previously proposed algorithm MASTER \cite{lu2019master}, which is originally proposed for scene text recognition. We optimize the MASTER model from several perspectives: network structure, optimizer, normalization method, pre-trained model, resolution of input image, data augmentation, and model ensemble. Our method achieves 0.7444 Exact Match and 0.8765 Exact Match @95\% on the TSR task, and obtains 0.5586 Exact Match and 0.7386 Exact Match 95\% on the TCR task.

Bingcong Li, Xin Tang, Xianbiao Qi et al.

Recently, inspired by Transformer, self-attention-based scene text recognition approaches have achieved outstanding performance. However, we find that the size of model expands rapidly with the lexicon increasing. Specifically, the number of parameters for softmax classification layer and output embedding layer are proportional to the vocabulary size. It hinders the development of a lightweight text recognition model especially applied for Chinese and multiple languages. Thus, we propose a lightweight scene text recognition model named Hamming OCR. In this model, a novel Hamming classifier, which adopts locality sensitive hashing (LSH) algorithm to encode each character, is proposed to replace the softmax regression and the generated LSH code is directly employed to replace the output embedding. We also present a simplified transformer decoder to reduce the number of parameters by removing the feed-forward network and using cross-layer parameter sharing technique. Compared with traditional methods, the number of parameters in both classification and embedding layers is independent on the size of vocabulary, which significantly reduces the storage requirement without loss of accuracy. Experimental results on several datasets, including four public benchmaks and a Chinese text dataset synthesized by SynthText with more than 20,000 characters, shows that Hamming OCR achieves competitive results.

Di Wu, Yihao Chen, Xianbiao Qi et al.

How can we effectively utilise the 2D monocular image information for recovering the 6D pose (6-DoF) of the visual objects? Deep learning has shown to be effective for robust and real-time monocular pose estimation. Oftentimes, the network learns to regress the 6-DoF pose using a naive loss function. However, due to a lack of geometrical scene understanding from the directly regressed pose estimation, there are misalignments between the rendered mesh from the 3D object and the 2D instance segmentation result, e.g., bounding boxes and masks prediction. This paper bridges the gap between 2D mask generation and 3D location prediction via a differentiable neural mesh renderer. We utilise the overlay between the accurate mask prediction and less accurate mesh prediction to iteratively optimise the direct regressed 6D pose information with a focus on translation estimation. By leveraging geometry, we demonstrate that our technique significantly improves direct regression performance on the difficult task of translation estimation and achieve the state of the art results on Peking University/Baidu - Autonomous Driving dataset and the ApolloScape 3D Car Instance dataset. The code can be found at \url{https://bit.ly/2IRihfU}.

Junjian Zhang, Chun-Guang Li, Chong You et al.

Subspace clustering methods based on data self-expression have become very popular for learning from data that lie in a union of low-dimensional linear subspaces. However, the applicability of subspace clustering has been limited because practical visual data in raw form do not necessarily lie in such linear subspaces. On the other hand, while Convolutional Neural Network (ConvNet) has been demonstrated to be a powerful tool for extracting discriminative features from visual data, training such a ConvNet usually requires a large amount of labeled data, which are unavailable in subspace clustering applications. To achieve simultaneous feature learning and subspace clustering, we propose an end-to-end trainable framework, called Self-Supervised Convolutional Subspace Clustering Network (S$^2$ConvSCN), that combines a ConvNet module (for feature learning), a self-expression module (for subspace clustering) and a spectral clustering module (for self-supervision) into a joint optimization framework. Particularly, we introduce a dual self-supervision that exploits the output of spectral clustering to supervise the training of the feature learning module (via a classification loss) and the self-expression module (via a spectral clustering loss). Our experiments on four benchmark datasets show the effectiveness of the dual self-supervision and demonstrate superior performance of our proposed approach.

Wangmeng Xiang, Jianqiang Huang, Xianbiao Qi et al.

Person re-identification (Person ReID) is a challenging task due to the large variations in camera viewpoint, lighting, resolution, and human pose. Recently, with the advancement of deep learning technologies, the performance of Person ReID has been improved swiftly. Feature extraction and feature matching are two crucial components in the training and deployment stages of Person ReID. However, many existing Person ReID methods have measure inconsistency between the training stage and the deployment stage, and they couple magnitude and orientation information of feature vectors in feature representation. Meanwhile, traditional triplet loss methods focus on samples within a mini-batch and lack knowledge of global feature distribution. To address these issues, we propose a novel homocentric hypersphere embedding scheme to decouple magnitude and orientation information for both feature and weight vectors, and reformulate classification loss and triplet loss to their angular versions and combine them into an angular discriminative loss. We evaluate our proposed method extensively on the widely used Person ReID benchmarks, including Market1501, CUHK03 and DukeMTMC-ReID. Our method demonstrates leading performance on all datasets.

Xianbiao Qi, Lei Zhang

Owe to the rapid development of deep neural network (DNN) techniques and the emergence of large scale face databases, face recognition has achieved a great success in recent years. During the training process of DNN, the face features and classification vectors to be learned will interact with each other, while the distribution of face features will largely affect the convergence status of network and the face similarity computing in test stage. In this work, we formulate jointly the learning of face features and classification vectors, and propose a simple yet effective centralized coordinate learning (CCL) method, which enforces the features to be dispersedly spanned in the coordinate space while ensuring the classification vectors to lie on a hypersphere. An adaptive angular margin is further proposed to enhance the discrimination capability of face features. Extensive experiments are conducted on six face benchmarks, including those have large age gap and hard negative samples. Trained only on the small-scale CASIA Webface dataset with 460K face images from about 10K subjects, our CCL model demonstrates high effectiveness and generality, showing consistently competitive performance across all the six benchmark databases.

Xiaopeng Hong, Xianbiao Qi, Guoying Zhao et al.

Fisher vector (FV) has become a popular image representation. One notable underlying assumption of the FV framework is that local descriptors are well decorrelated within each cluster so that the covariance matrix for each Gaussian can be simplified to be diagonal. Though the FV usually relies on the Principal Component Analysis (PCA) to decorrelate local features, the PCA is applied to the entire training data and hence it only diagonalizes the \textit{universal} covariance matrix, rather than those w.r.t. the local components. As a result, the local decorrelation assumption is usually not supported in practice. To relax this assumption, this paper proposes a completed model of the Fisher vector, which is termed as the Completed Fisher vector (CFV). The CFV is a more general framework of the FV, since it encodes not only the variances but also the correlations of the whitened local descriptors. The CFV thus leads to improved discriminative power. We take the task of material categorization as an example and experimentally show that: 1) the CFV outperforms the FV under all parameter settings; 2) the CFV is robust to the changes in the number of components in the mixture; 3) even with a relatively small visual vocabulary the CFV still works well on two challenging datasets.

Xianbiao Qi, Guoying Zhao, Jie Chen et al.

\textit{Indirect Immunofluorescence Imaging of Human Epithelial Type 2} (HEp-2) cells is an effective way to identify the presence of Anti-Nuclear Antibody (ANA). Most existing works on HEp-2 cell classification mainly focus on feature extraction, feature encoding and classifier design. Very few efforts have been devoted to study the importance of the pre-processing techniques. In this paper, we analyze the importance of the pre-processing, and investigate the role of Gaussian Scale Space (GSS) theory as a pre-processing approach for the HEp-2 cell classification task. We validate the GSS pre-processing under the Local Binary Pattern (LBP) and the Bag-of-Words (BoW) frameworks. Under the BoW framework, the introduced pre-processing approach, using only one Local Orientation Adaptive Descriptor (LOAD), achieved superior performance on the Executable Thematic on Pattern Recognition Techniques for Indirect Immunofluorescence (ET-PRT-IIF) image analysis. Our system, using only one feature, outperformed the winner of the ICPR 2014 contest that combined four types of features. Meanwhile, the proposed pre-processing method is not restricted to this work; it can be generalized to many existing works.

Xianbiao Qi, Guoying Zhao, Linlin Shen et al.

In this paper, we propose a novel local feature, called Local Orientation Adaptive Descriptor (LOAD), to capture regional texture in an image. In LOAD, we proposed to define point description on an Adaptive Coordinate System (ACS), adopt a binary sequence descriptor to capture relationships between one point and its neighbors and use multi-scale strategy to enhance the discriminative power of the descriptor. The proposed LOAD enjoys not only discriminative power to capture the texture information, but also has strong robustness to illumination variation and image rotation. Extensive experiments on benchmark data sets of texture classification and real-world material recognition show that the proposed LOAD yields the state-of-the-art performance. It is worth to mention that we achieve a 65.4\% classification accuracy-- which is, to the best of our knowledge, the highest record by far --on Flickr Material Database by using a single feature. Moreover, by combining LOAD with the feature extracted by Convolutional Neural Networks (CNN), we obtain significantly better performance than both the LOAD and CNN. This result confirms that the LOAD is complementary to the learning-based features.

Xianbiao Qi, Guoying Zhao, Chun-Guang Li et al.

Indirect Immunofluorescence (IIF) HEp-2 cell image is an effective evidence for diagnosis of autoimmune diseases. Recently computer-aided diagnosis of autoimmune diseases by IIF HEp-2 cell classification has attracted great attention. However the HEp-2 cell classification task is quite challenging due to large intra-class variation and small between-class variation. In this paper we propose an effective and efficient approach for the automatic classification of IIF HEp-2 cell image by fusing multi-resolution texture information and richer shape information. To be specific, we propose to: a) capture the multi-resolution texture information by a novel Pairwise Rotation Invariant Co-occurrence of Local Gabor Binary Pattern (PRICoLGBP) descriptor, b) depict the richer shape information by using an Improved Fisher Vector (IFV) model with RootSIFT features which are sampled from large image patches in multiple scales, and c) combine them properly. We evaluate systematically the proposed approach on the IEEE International Conference on Pattern Recognition (ICPR) 2012, IEEE International Conference on Image Processing (ICIP) 2013 and ICPR 2014 contest data sets. The experimental results for the proposed methods significantly outperform the winners of ICPR 2012 and ICIP 2013 contest, and achieve comparable performance with the winner of the newly released ICPR 2014 contest.

Xianbiao Qi, Chun-Guang Li, Guoying Zhao et al.

Dynamic texture and scene classification are two fundamental problems in understanding natural video content. Extracting robust and effective features is a crucial step towards solving these problems. However the existing approaches suffer from the sensitivity to either varying illumination, or viewpoint changing, or even camera motion, and/or the lack of spatial information. Inspired by the success of deep structures in image classification, we attempt to leverage a deep structure to extract feature for dynamic texture and scene classification. To tackle with the challenges in training a deep structure, we propose to transfer some prior knowledge from image domain to video domain. To be specific, we propose to apply a well-trained Convolutional Neural Network (ConvNet) as a mid-level feature extractor to extract features from each frame, and then form a representation of a video by concatenating the first and the second order statistics over the mid-level features. We term this two-level feature extraction scheme as a Transferred ConvNet Feature (TCoF). Moreover we explore two different implementations of the TCoF scheme, i.e., the \textit{spatial} TCoF and the \textit{temporal} TCoF, in which the mean-removed frames and the difference between two adjacent frames are used as the inputs of the ConvNet, respectively. We evaluate systematically the proposed spatial TCoF and the temporal TCoF schemes on three benchmark data sets, including DynTex, YUPENN, and Maryland, and demonstrate that the proposed approach yields superior performance.