Jiasheng Tang

Jiazheng Xing, Hangjie Yuan, Lingling Cai et al.

Connector-based video unified models have demonstrated strong capability in instruction-grounded video synthesis, but integrating a large high-fidelity generator into the unified training loop is computationally prohibitive, limiting achievable visual quality. We therefore propose Lumos-Nexus, a training-efficient unified video generation framework that facilitates the development of strong reasoning-driven generation capabilities while significantly enhancing visual fidelity. Lumos-Nexus adopts a two-stage design: 1) During training, only a lightweight generator is aligned with the understanding block to learn to take in reasoning-driven semantic control. 2) During inference, we introduce Unified Progressive Frequency Bridging (UPFB) to progressively hand off generation to a high-capacity pretrained generator in the shared latent space, enabling coarse-to-fine refinement and producing high-fidelity videos without compromising reasoning quality. To fill the gap in reasoning-driven video generation benchmarks, we introduce VR-Bench, which assesses a model's capability to translate inferred intent into coherent and semantically aligned video content. Extensive experiments demonstrate that Lumos-Nexus achieves substantial gains in visual realism and temporal coherence on VBench, while exhibiting strong reasoning-based generative performance on VR-Bench. Code and models are available at https://jiazheng-xing.github.io/nexus-lumos-home/.

Yifan Pu, Yizeng Han, Zhiwei Tang et al.

Diffusion distillation has dramatically accelerated class-conditional image synthesis, but its applicability to open-ended text-to-image (T2I) generation is still unclear. We present the first systematic study that adapts and compares state-of-the-art distillation techniques on a strong T2I teacher model, FLUX.1-lite. By casting existing methods into a unified framework, we identify the key obstacles that arise when moving from discrete class labels to free-form language prompts. Beyond a thorough methodological analysis, we offer practical guidelines on input scaling, network architecture, and hyperparameters, accompanied by an open-source implementation and pretrained student models. Our findings establish a solid foundation for deploying fast, high-fidelity, and resource-efficient diffusion generators in real-world T2I applications. Code is available on github.com/alibaba-damo-academy/T2I-Distill.

Zhiwei Tang, Jiasheng Tang, Hao Luo et al.

Diffusion models have emerged as state-of-the-art generative models for image generation. However, sampling from diffusion models is usually time-consuming due to the inherent autoregressive nature of their sampling process. In this work, we propose a novel approach that accelerates the sampling of diffusion models by parallelizing the autoregressive process. Specifically, we reformulate the sampling process as solving a system of triangular nonlinear equations through fixed-point iteration. With this innovative formulation, we explore several systematic techniques to further reduce the iteration steps required by the solving process. Applying these techniques, we introduce ParaTAA, a universal and training-free parallel sampling algorithm that can leverage extra computational and memory resources to increase the sampling speed. Our experiments demonstrate that ParaTAA can decrease the inference steps required by common sequential sampling algorithms such as DDIM and DDPM by a factor of 4$\sim$14 times. Notably, when applying ParaTAA with 100 steps DDIM for Stable Diffusion, a widely-used text-to-image diffusion model, it can produce the same images as the sequential sampling in only 7 inference steps. The code is available at https://github.com/TZW1998/ParaTAA-Diffusion.

Hangjie Yuan, Weihua Chen, Jun Cen et al.

Autoregressive large language models (LLMs) have unified a vast range of language tasks, inspiring preliminary efforts in autoregressive video generation. Existing autoregressive video generators either diverge from standard LLM architectures, depend on bulky external text encoders, or incur prohibitive latency due to next-token decoding. In this paper, we introduce Lumos-1, an autoregressive video generator that retains the LLM architecture with minimal architectural modifications. To inject spatiotemporal correlations in LLMs, we identify the efficacy of incorporating 3D RoPE and diagnose its imbalanced frequency spectrum ranges. Therefore, we propose MM-RoPE, a RoPE scheme that preserves the original textual RoPE while providing comprehensive frequency spectra and scaled 3D positions for modeling multimodal spatiotemporal data. Moreover, Lumos-1 resorts to a token dependency strategy that obeys intra-frame bidirectionality and inter-frame temporal causality. Based on this dependency strategy, we identify the issue of frame-wise loss imbalance caused by spatial information redundancy and solve it by proposing Autoregressive Discrete Diffusion Forcing (AR-DF). AR-DF introduces temporal tube masking during training with a compatible inference-time masking policy to avoid quality degradation. By using memory-efficient training techniques, we pre-train Lumos-1 on only 48 GPUs, achieving performance comparable to EMU3 on GenEval, COSMOS-Video2World on VBench-I2V, and OpenSoraPlan on VBench-T2V. Code and models are available at https://github.com/alibaba-damo-academy/Lumos.

Zeyu Zhang, Shuning Chang, Yuanyu He et al.

Generating minute-long videos is a critical step toward developing world models, providing a foundation for realistic extended scenes and advanced AI simulators. The emerging semi-autoregressive (block diffusion) paradigm integrates the strengths of diffusion and autoregressive models, enabling arbitrary-length video generation and improving inference efficiency through KV caching and parallel sampling. However, it yet faces two enduring challenges: (i) KV-cache-induced long-horizon error accumulation, and (ii) the lack of fine-grained long-video benchmarks and coherence-aware metrics. To overcome these limitations, we propose BlockVid, a novel block diffusion framework equipped with semantic-aware sparse KV cache, an effective training strategy called Block Forcing, and dedicated chunk-wise noise scheduling and shuffling to reduce error propagation and enhance temporal consistency. We further introduce LV-Bench, a fine-grained benchmark for minute-long videos, complete with new metrics evaluating long-range coherence. Extensive experiments on VBench and LV-Bench demonstrate that BlockVid consistently outperforms existing methods in generating high-quality, coherent minute-long videos. In particular, it achieves a 22.2% improvement on VDE Subject and a 19.4% improvement on VDE Clarity in LV-Bench over the state of the art approaches. Project website: https://ziplab.co/BlockVid. Inferix (Code): https://github.com/alibaba-damo-academy/Inferix.

Jiefeng Li, Siyuan Bian, Qi Liu et al.

With the progress of 3D human pose and shape estimation, state-of-the-art methods can either be robust to occlusions or obtain pixel-aligned accuracy in non-occlusion cases. However, they cannot obtain robustness and mesh-image alignment at the same time. In this work, we present NIKI (Neural Inverse Kinematics with Invertible Neural Network), which models bi-directional errors to improve the robustness to occlusions and obtain pixel-aligned accuracy. NIKI can learn from both the forward and inverse processes with invertible networks. In the inverse process, the model separates the error from the plausible 3D pose manifold for a robust 3D human pose estimation. In the forward process, we enforce the zero-error boundary conditions to improve the sensitivity to reliable joint positions for better mesh-image alignment. Furthermore, NIKI emulates the analytical inverse kinematics algorithms with the twist-and-swing decomposition for better interpretability. Experiments on standard and occlusion-specific benchmarks demonstrate the effectiveness of NIKI, where we exhibit robust and well-aligned results simultaneously. Code is available at https://github.com/Jeff-sjtu/NIKI

Chong Liu, Yuqi Zhang, Hao Luo et al.

Multi-Target Multi-Camera Tracking has a wide range of applications and is the basis for many advanced inferences and predictions. This paper describes our solution to the Track 3 multi-camera vehicle tracking task in 2021 AI City Challenge (AICITY21). This paper proposes a multi-target multi-camera vehicle tracking framework guided by the crossroad zones. The framework includes: (1) Use mature detection and vehicle re-identification models to extract targets and appearance features. (2) Use modified JDETracker (without detection module) to track single-camera vehicles and generate single-camera tracklets. (3) According to the characteristics of the crossroad, the Tracklet Filter Strategy and the Direction Based Temporal Mask are proposed. (4) Propose Sub-clustering in Adjacent Cameras for multi-camera tracklets matching. Through the above techniques, our method obtained an IDF1 score of 0.8095, ranking first on the leaderboard. The code have released: https://github.com/LCFractal/AIC21-MTMC.

Fei Du, Bo Xu, Jiasheng Tang et al.

We extend the classical tracking-by-detection paradigm to this tracking-any-object task. Solid detection results are first extracted from TAO dataset. Some state-of-the-art techniques like \textbf{BA}lanced-\textbf{G}roup \textbf{S}oftmax (\textbf{BAGS}\cite{li2020overcoming}) and DetectoRS\cite{qiao2020detectors} are integrated during detection. Then we learned appearance features to represent any object by training feature learning networks. We ensemble several models for improving detection and feature representation. Simple linking strategies with most similar appearance features and tracklet-level post association module are finally applied to generate final tracking results. Our method is submitted as \textbf{AOA} on the challenge website. Code is available at https://github.com/feiaxyt/Winner_ECCV20_TAO.

Wangbo Zhao, Jiasheng Tang, Yizeng Han et al. · stanford

Existing parameter-efficient fine-tuning (PEFT) methods have achieved significant success on vision transformers (ViTs) adaptation by improving parameter efficiency. However, the exploration of enhancing inference efficiency during adaptation remains underexplored. This limits the broader application of pre-trained ViT models, especially when the model is computationally extensive. In this paper, we propose Dynamic Tuning (DyT), a novel approach to improve both parameter and inference efficiency for ViT adaptation. Specifically, besides using the lightweight adapter modules, we propose a token dispatcher to distinguish informative tokens from less important ones, allowing the latter to dynamically skip the original block, thereby reducing the redundant computation during inference. Additionally, we explore multiple design variants to find the best practice of DyT. Finally, inspired by the mixture-of-experts (MoE) mechanism, we introduce an enhanced adapter to further boost the adaptation performance. We validate DyT across various tasks, including image/video recognition and semantic segmentation. For instance, DyT achieves superior performance compared to existing PEFT methods while evoking only 71% of their FLOPs on the VTAB-1K benchmark.

Wangbo Zhao, Yizeng Han, Jiasheng Tang et al.

Vision-language models (VLMs) have shown remarkable success across various multi-modal tasks, yet large VLMs encounter significant efficiency challenges due to processing numerous visual tokens. A promising approach to accelerating large VLM inference is using partial information, such as attention maps from specific layers, to assess token importance and prune less essential tokens. However, our study reveals three key insights: (i) Partial attention information is insufficient for accurately identifying critical visual tokens, resulting in suboptimal performance, especially at low token retention ratios; (ii) Global attention information, such as the attention map aggregated across all layers, more effectively preserves essential tokens and maintains comparable performance under aggressive pruning. However, the attention maps from all layers requires a full inference pass, which increases computational load and is therefore impractical in existing methods; and (iii) The global attention map aggregated from a small VLM closely resembles that of a large VLM, suggesting an efficient alternative. Based on these findings, we introduce a \textbf{training-free} method, \underline{\textbf{S}}mall VLM \underline{\textbf{G}}uidance for accelerating \underline{\textbf{L}}arge VLMs (\textbf{SGL}). Specifically, we employ the attention map aggregated from a small VLM to guide visual token pruning in a large VLM. Additionally, an early exiting mechanism is developed to fully use the small VLM's predictions, dynamically invoking the larger VLM only when necessary, yielding a superior trade-off between accuracy and computation. Extensive evaluations across 11 benchmarks demonstrate the effectiveness and generalizability of SGL, achieving up to 91\% pruning ratio for visual tokens while retaining competitive performance.

Kejia Zhang, Keda Tao, Jiasheng Tang et al.

Large vision-language models (LVMs) extend large language models (LLMs) with visual perception capabilities, enabling them to process and interpret visual information. A major challenge compromising their reliability is object hallucination that LVMs may generate plausible but factually inaccurate information. We propose a novel visual adversarial perturbation (VAP) method to mitigate this hallucination issue. VAP alleviates LVM hallucination by applying strategically optimized visual noise without altering the base model. Our approach formulates hallucination suppression as an optimization problem, leveraging adversarial strategies to generate beneficial visual perturbations that enhance the model's factual grounding and reduce parametric knowledge bias. Extensive experimental results demonstrate that our method consistently reduces object hallucinations across 8 state-of-the-art LVMs, validating its efficacy across diverse evaluations.

Akide Liu, Zeyu Zhang, Zhexin Li et al.

Diffusion generative models have become the standard for producing high-quality, coherent video content, yet their slow inference speeds and high computational demands hinder practical deployment. Although both quantization and sparsity can independently accelerate inference while maintaining generation quality, naively combining these techniques in existing training-free approaches leads to significant performance degradation due to the lack of joint optimization. We introduce FPSAttention, a novel training-aware co-design of FP8 quantization and sparsity for video generation, with a focus on the 3D bi-directional attention mechanism. Our approach features three key innovations: 1) A unified 3D tile-wise granularity that simultaneously supports both quantization and sparsity; 2) A denoising step-aware strategy that adapts to the noise schedule, addressing the strong correlation between quantization/sparsity errors and denoising steps; 3) A native, hardware-friendly kernel that leverages FlashAttention and is implemented with optimized Hopper architecture features for highly efficient execution. Trained on Wan2.1's 1.3B and 14B models and evaluated on the VBench benchmark, FPSAttention achieves a 7.09x kernel speedup for attention operations and a 4.96x end-to-end speedup for video generation compared to the BF16 baseline at 720p resolution-without sacrificing generation quality.

Wangbo Zhao, Yizeng Han, Jiasheng Tang et al.

Diffusion Transformer (DiT), an emerging diffusion model for visual generation, has demonstrated superior performance but suffers from substantial computational costs. Our investigations reveal that these costs primarily stem from the \emph{static} inference paradigm, which inevitably introduces redundant computation in certain \emph{diffusion timesteps} and \emph{spatial regions}. To overcome this inefficiency, we propose \textbf{Dy}namic \textbf{Di}ffusion \textbf{T}ransformer (DyDiT), an architecture that \emph{dynamically} adjusts its computation along both \emph{timestep} and \emph{spatial} dimensions. Specifically, we introduce a \emph{Timestep-wise Dynamic Width} (TDW) approach that adapts model width conditioned on the generation timesteps. In addition, we design a \emph{Spatial-wise Dynamic Token} (SDT) strategy to avoid redundant computation at unnecessary spatial locations. TDW and SDT can be seamlessly integrated into DiT and significantly accelerates the generation process. Building on these designs, we further enhance DyDiT in three key aspects. First, DyDiT is integrated seamlessly with flow matching-based generation, enhancing its versatility. Furthermore, we enhance DyDiT to tackle more complex visual generation tasks, including video generation and text-to-image generation, thereby broadening its real-world applications. Finally, to address the high cost of full fine-tuning and democratize technology access, we investigate the feasibility of training DyDiT in a parameter-efficient manner and introduce timestep-based dynamic LoRA (TD-LoRA). Extensive experiments on diverse visual generation models, including DiT, SiT, Latte, and FLUX, demonstrate the effectiveness of DyDiT.

Siyuan Bian, Jiefeng Li, Jiasheng Tang et al.

Accurate human shape recovery from a monocular RGB image is a challenging task because humans come in different shapes and sizes and wear different clothes. In this paper, we propose ShapeBoost, a new human shape recovery framework that achieves pixel-level alignment even for rare body shapes and high accuracy for people wearing different types of clothes. Unlike previous approaches that rely on the use of PCA-based shape coefficients, we adopt a new human shape parameterization that decomposes the human shape into bone lengths and the mean width of each part slice. This part-based parameterization technique achieves a balance between flexibility and validity using a semi-analytical shape reconstruction algorithm. Based on this new parameterization, a clothing-preserving data augmentation module is proposed to generate realistic images with diverse body shapes and accurate annotations. Experimental results show that our method outperforms other state-of-the-art methods in diverse body shape situations as well as in varied clothing situations.

Wangbo Zhao, Yizeng Han, Zhiwei Tang et al.

Diffusion Transformers (DiTs) excel at visual generation yet remain hampered by slow sampling. Existing training-free accelerators - step reduction, feature caching, and sparse attention - enhance inference speed but typically rely on a uniform heuristic or a manually designed adaptive strategy for all images, leaving quality on the table. Alternatively, dynamic neural networks offer per-image adaptive acceleration, but their high fine-tuning costs limit broader applicability. To address these limitations, we introduce RAPID3: Tri-Level Reinforced Acceleration Policies for Diffusion Transformers, a framework that delivers image-wise acceleration with zero updates to the base generator. Specifically, three lightweight policy heads - Step-Skip, Cache-Reuse, and Sparse-Attention - observe the current denoising state and independently decide their corresponding speed-up at each timestep. All policy parameters are trained online via Group Relative Policy Optimization (GRPO) while the generator remains frozen. Meanwhile, an adversarially learned discriminator augments the reward signal, discouraging reward hacking by boosting returns only when generated samples stay close to the original model's distribution. Across state-of-the-art DiT backbones, including Stable Diffusion 3 and FLUX, RAPID3 achieves nearly 3x faster sampling with competitive generation quality.

Inferix Team, Tianyu Feng, Yizeng Han et al.

World models serve as core simulators for fields such as agentic AI, embodied AI, and gaming, capable of generating long, physically realistic, and interactive high-quality videos. Moreover, scaling these models could unlock emergent capabilities in visual perception, understanding, and reasoning, paving the way for a new paradigm that moves beyond current LLM-centric vision foundation models. A key breakthrough empowering them is the semi-autoregressive (block-diffusion) decoding paradigm, which merges the strengths of diffusion and autoregressive methods by generating video tokens in block-applying diffusion within each block while conditioning on previous ones, resulting in more coherent and stable video sequences. Crucially, it overcomes limitations of standard video diffusion by reintroducing LLM-style KV Cache management, enabling efficient, variable-length, and high-quality generation. Therefore, Inferix is specifically designed as a next-generation inference engine to enable immersive world synthesis through optimized semi-autoregressive decoding processes. This dedicated focus on world simulation distinctly sets it apart from systems engineered for high-concurrency scenarios (like vLLM or SGLang) and from classic video diffusion models (such as xDiTs). Inferix further enhances its offering with interactive video streaming and profiling, enabling real-time interaction and realistic simulation to accurately model world dynamics. Additionally, it supports efficient benchmarking through seamless integration of LV-Bench, a new fine-grained evaluation benchmark tailored for minute-long video generation scenarios. We hope the community will work together to advance Inferix and foster world model exploration.

Kejia Zhang, Keda Tao, Zhiming Luo et al.

Multimodal large language models (MLLMs) enable vision-language reasoning, yet often generate plausible outputs that are factually incorrect or visually ungrounded, thereby compromising their reliability. Direct preference optimization (DPO) is a common strategy for correcting hallucinations by aligning model outputs with human preferences. Existing DPO strategies typically treat hallucination-related preferences as fixed targets, relying on static supervision signals during training. This approach tends to overfit to superficial linguistic cues in preference data, leading to distributional rigidity and spurious correlations that impair grounding in causally relevant visual information. To overcome this limitation, we propose TARS, a token-adaptive preference strategy that reformulates DPO as a min-max optimization problem. TARS maximizes token-level distributional shifts under semantic constraints to simulate alignment uncertainty, and simultaneously minimizes the expected preference loss under these controlled perturbations. This joint objective preserves causal grounding while mitigating overfitting to preference patterns, thereby reducing hallucinations in multimodal reasoning. We evaluate TARS on multiple hallucination benchmarks and find consistently strong performance. Using only 4.8k preference samples and no expert feedback, TARS reduces hallucination rates from 26.4% to 13.2% and decreases cognition value from 2.5 to 0.4. It outperforms standard DPO and matches GPT-4o on several key metrics.

Weitao Du, Shuning Chang, Jiasheng Tang et al.

In this work, we propose a novel generative learning paradigm, K-Flow, an algorithm that flows along the $K$-amplitude. Here, $k$ is a scaling parameter that organizes frequency bands (or projected coefficients), and amplitude describes the norm of such projected coefficients. By incorporating the $K$-amplitude decomposition, K-Flow enables flow matching across the scaling parameter as time. We discuss three venues and six properties of K-Flow, from theoretical foundations, energy and temporal dynamics, and practical applications, respectively. Specifically, from the practical usage perspective, K-Flow allows steerable generation by controlling the information at different scales. To demonstrate the effectiveness of K-Flow, we conduct experiments on unconditional image generation, class-conditional image generation, and molecule assembly generation. Additionally, we conduct three ablation studies to demonstrate how K-Flow steers scaling parameter to effectively control the resolution of image generation.

Shuning Chang, Pichao Wang, Jiasheng Tang et al.

Diffusion Transformers (DiT) are renowned for their impressive generative performance; however, they are significantly constrained by considerable computational costs due to the quadratic complexity in self-attention and the extensive sampling steps required. While advancements have been made in expediting the sampling process, the underlying architectural inefficiencies within DiT remain underexplored. We introduce SparseDiT, a novel framework that implements token sparsification across spatial and temporal dimensions to enhance computational efficiency while preserving generative quality. Spatially, SparseDiT employs a tri-segment architecture that allocates token density based on feature requirements at each layer: Poolingformer in the bottom layers for efficient global feature extraction, Sparse-Dense Token Modules (SDTM) in the middle layers to balance global context with local detail, and dense tokens in the top layers to refine high-frequency details. Temporally, SparseDiT dynamically modulates token density across denoising stages, progressively increasing token count as finer details emerge in later timesteps. This synergy between SparseDiT spatially adaptive architecture and its temporal pruning strategy enables a unified framework that balances efficiency and fidelity throughout the generation process. Our experiments demonstrate SparseDiT effectiveness, achieving a 55% reduction in FLOPs and a 175% improvement in inference speed on DiT-XL with similar FID score on 512x512 ImageNet, a 56% reduction in FLOPs across video generation datasets, and a 69% improvement in inference speed on PixArt-$α$ on text-to-image generation task with a 0.24 FID score decrease. SparseDiT provides a scalable solution for high-quality diffusion-based generation compatible with sampling optimization techniques.

Qi Qian, Jiasheng Tang, Hao Li et al.

Distance metric learning (DML) has been studied extensively in the past decades for its superior performance with distance-based algorithms. Most of the existing methods propose to learn a distance metric with pairwise or triplet constraints. However, the number of constraints is quadratic or even cubic in the number of the original examples, which makes it challenging for DML to handle the large-scale data set. Besides, the real-world data may contain various uncertainty, especially for the image data. The uncertainty can mislead the learning procedure and cause the performance degradation. By investigating the image data, we find that the original data can be observed from a small set of clean latent examples with different distortions. In this work, we propose the margin preserving metric learning framework to learn the distance metric and latent examples simultaneously. By leveraging the ideal properties of latent examples, the training efficiency can be improved significantly while the learned metric also becomes robust to the uncertainty in the original data. Furthermore, we can show that the metric is learned from latent examples only, but it can preserve the large margin property even for the original data. The empirical study on the benchmark image data sets demonstrates the efficacy and efficiency of the proposed method.

Qi Qian, Shenghuo Zhu, Jiasheng Tang et al.

In this work, we study the problem of learning a single model for multiple domains. Unlike the conventional machine learning scenario where each domain can have the corresponding model, multiple domains (i.e., applications/users) may share the same machine learning model due to maintenance loads in cloud computing services. For example, a digit-recognition model should be applicable to hand-written digits, house numbers, car plates, etc. Therefore, an ideal model for cloud computing has to perform well at each applicable domain. To address this new challenge from cloud computing, we develop a framework of robust optimization over multiple domains. In lieu of minimizing the empirical risk, we aim to learn a model optimized to the adversarial distribution over multiple domains. Hence, we propose to learn the model and the adversarial distribution simultaneously with the stochastic algorithm for efficiency. Theoretically, we analyze the convergence rate for convex and non-convex models. To our best knowledge, we first study the convergence rate of learning a robust non-convex model with a practical algorithm. Furthermore, we demonstrate that the robustness of the framework and the convergence rate can be further enhanced by appropriate regularizers over the adversarial distribution. The empirical study on real-world fine-grained visual categorization and digits recognition tasks verifies the effectiveness and efficiency of the proposed framework.