Zhantao Yang

Zhantao Yang, Ruili Feng, Han Zhang et al. · tsinghua

Diffusion models, which employ stochastic differential equations to sample images through integrals, have emerged as a dominant class of generative models. However, the rationality of the diffusion process itself receives limited attention, leaving the question of whether the problem is well-posed and well-conditioned. In this paper, we explore a perplexing tendency of diffusion models: they often display the infinite Lipschitz property of the network with respect to time variable near the zero point. We provide theoretical proofs to illustrate the presence of infinite Lipschitz constants and empirical results to confirm it. The Lipschitz singularities pose a threat to the stability and accuracy during both the training and inference processes of diffusion models. Therefore, the mitigation of Lipschitz singularities holds great potential for enhancing the performance of diffusion models. To address this challenge, we propose a novel approach, dubbed E-TSDM, which alleviates the Lipschitz singularities of the diffusion model near the zero point of timesteps. Remarkably, our technique yields a substantial improvement in performance. Moreover, as a byproduct of our method, we achieve a dramatic reduction in the Fréchet Inception Distance of acceleration methods relying on network Lipschitz, including DDIM and DPM-Solver, by over 33%. Extensive experiments on diverse datasets validate our theory and method. Our work may advance the understanding of the general diffusion process, and also provide insights for the design of diffusion models.

Han Zhang, Ruili Feng, Zhantao Yang et al.

Diffusion models, which learn to reverse a signal destruction process to generate new data, typically require the signal at each step to have the same dimension. We argue that, considering the spatial redundancy in image signals, there is no need to maintain a high dimensionality in the evolution process, especially in the early generation phase. To this end, we make a theoretical generalization of the forward diffusion process via signal decomposition. Concretely, we manage to decompose an image into multiple orthogonal components and control the attenuation of each component when perturbing the image. That way, along with the noise strength increasing, we are able to diminish those inconsequential components and thus use a lower-dimensional signal to represent the source, barely losing information. Such a reformulation allows to vary dimensions in both training and inference of diffusion models. Extensive experiments on a range of datasets suggest that our approach substantially reduces the computational cost and achieves on-par or even better synthesis performance compared to baseline methods. We also show that our strategy facilitates high-resolution image synthesis and improves FID of diffusion model trained on FFHQ at $1024\times1024$ resolution from 52.40 to 10.46. Code and models will be made publicly available.

Zhantao Yang, Ruili Feng, Keyu Yan et al.

Advancements in large Vision-Language Models have brought precise, accurate image captioning, vital for advancing multi-modal image understanding and processing. Yet these captions often carry lengthy, intertwined contexts that are difficult to parse and frequently overlook essential cues, posing a great barrier for models like GroundingDINO and SDXL, which lack the strong text encoding and syntax analysis needed to fully leverage dense captions. To address this, we propose BACON, a prompting method that breaks down VLM-generated captions into disentangled, structured elements such as objects, relationships, styles, and themes. This approach not only minimizes confusion from handling complex contexts but also allows for efficient transfer into a JSON dictionary, enabling models without linguistic processing capabilities to easily access key information. We annotated 100,000 image-caption pairs using BACON with GPT-4V and trained an LLaVA captioner on this dataset, enabling it to produce BACON-style captions without relying on costly GPT-4V. Evaluations of overall quality, precision, and recall-as well as user studies-demonstrate that the resulting caption model consistently outperforms other SOTA VLM models in generating high-quality captions. Besides, we show that BACON-style captions exhibit better clarity when applied to various models, enabling them to accomplish previously unattainable tasks or surpass existing SOTA solutions without training. For example, BACON-style captions help GroundingDINO achieve 1.51x higher recall scores on open-vocabulary object detection tasks compared to leading methods.

Ximing Xing, Ziteng Xue, Zhenxi Li et al.

Recent large language models have shifted SVG generation from differentiable rendering optimization to autoregressive program synthesis. However, existing approaches still rely on generic byte-level tokenization inherited from natural language processing, which poorly reflects the geometric structure of vector graphics. Numerical coordinates are fragmented into discrete symbols, destroying spatial relationships and introducing severe token redundancy, often leading to coordinate hallucination and inefficient long-sequence generation. To address these challenges, we propose HiVG, a hierarchical SVG tokenization framework tailored for autoregressive vector graphics generation. HiVG decomposes raw SVG strings into structured \textit{atomic tokens} and further compresses executable command--parameter groups into geometry-constrained \textit{segment tokens}, substantially improving sequence efficiency while preserving syntactic validity. To further mitigate spatial mismatch, we introduce a Hierarchical Mean--Noise (HMN) initialization strategy that injects numerical ordering signals and semantic priors into new token embeddings. Combined with a curriculum training paradigm that progressively increases program complexity, HiVG enables more stable learning of executable SVG programs. Extensive experiments on both text-to-SVG and image-to-SVG tasks demonstrate improved generation fidelity, spatial consistency, and sequence efficiency compared with conventional tokenization schemes.

Siyu Cao, Hangting Chen, Peng Chen et al.

We present HunyuanImage 3.0, a native multimodal model that unifies multimodal understanding and generation within an autoregressive framework, with its image generation module publicly available. The achievement of HunyuanImage 3.0 relies on several key components, including meticulous data curation, advanced architecture design, a native Chain-of-Thoughts schema, progressive model pre-training, aggressive model post-training, and an efficient infrastructure that enables large-scale training and inference. With these advancements, we successfully trained a Mixture-of-Experts (MoE) model comprising over 80 billion parameters in total, with 13 billion parameters activated per token during inference, making it the largest and most powerful open-source image generative model to date. We conducted extensive experiments and the results of automatic and human evaluation of text-image alignment and visual quality demonstrate that HunyuanImage 3.0 rivals previous state-of-the-art models. By releasing the code and weights of HunyuanImage 3.0, we aim to enable the community to explore new ideas with a state-of-the-art foundation model, fostering a dynamic and vibrant multimodal ecosystem. All open source assets are publicly available at https://github.com/Tencent-Hunyuan/HunyuanImage-3.0

Zhangquan Chen, Jiale Tao, Ruihuang Li et al.

While humans perceive the world through diverse modalities that operate synergistically to support a holistic understanding of their surroundings, existing omnivideo models still face substantial challenges on audio-visual understanding tasks. In this paper, we propose OmniVideo-R1, a novel reinforced framework that improves mixed-modality reasoning. OmniVideo-R1 empowers models to "think with omnimodal cues" by two key strategies: (1) query-intensive grounding based on self-supervised learning paradigms; and (2) modality-attentive fusion built upon contrastive learning paradigms. Extensive experiments on multiple benchmarks demonstrate that OmniVideo-R1 consistently outperforms strong baselines, highlighting its effectiveness and robust generalization capabilities.

Chen Li, Zhantao Yang, Fangyi Chen et al.

Long-horizon reasoning requires deciding not only what actions to take, but how deeply to commit before the next observation. We formalize this as \emph{commitment depth}: the number of primitive actions executed open-loop between replans. Commitment depth induces a trade-off between replanning cost and compounding execution error, yet most existing long-horizon systems fix it as a hand-designed scalar. In this work, we instead treat commitment depth as a learnable, state-conditioned variable of the policy itself. We instantiate this within a model-native vision--language policy that jointly predicts both what to execute and for how long. Across Sliding Puzzle and Sokoban, the resulting adaptive policy Pareto-dominates every non-degenerate fixed-depth baseline, achieving up to 12.5 percentage points higher solve rate while using approximately 25\% fewer primitive actions per episode. Despite using a 7B backbone, our method outperforms GPT-5.5 and Claude Sonnet on both tasks, while every tested open-weight vision--language model achieves 0\% zero-shot success. We further present a theoretical analysis showing that, under the standard commitment-depth surrogate, state-conditioned commitment strictly dominates any fixed depth whenever the locally optimal depth varies across states.

Han Zhang, Jinghong Mao, Shangwen Zhu et al.

Diffusion reconstruction plays a critical role in various applications such as image editing, restoration, and style transfer. In theory, the reconstruction should be simple - it just inverts and regenerates images by numerically solving the Probability Flow-Ordinary Differential Equation (PF-ODE). Yet in practice, noticeable reconstruction errors have been observed, which cannot be well explained by numerical errors. In this work, we identify a deeper intrinsic property in the PF-ODE generation process, the instability, that can further amplify the reconstruction errors. The root of this instability lies in the sparsity inherent in the generation distribution, which means that the probability is concentrated on scattered and small regions while the vast majority remains almost empty. To demonstrate the existence of instability and its amplification on reconstruction error, we conduct experiments on both toy numerical examples and popular open-sourced diffusion models. Furthermore, based on the characteristics of image data, we theoretically prove that the instability's probability converges to one as the data dimensionality increases. Our findings highlight the inherent challenges in diffusion-based reconstruction and can offer insights for future improvements.

Ruili Feng, Han Zhang, Zhantao Yang et al.

We present The Matrix, the first foundational realistic world simulator capable of generating continuous 720p high-fidelity real-scene video streams with real-time, responsive control in both first- and third-person perspectives, enabling immersive exploration of richly dynamic environments. Trained on limited supervised data from AAA games like Forza Horizon 5 and Cyberpunk 2077, complemented by large-scale unsupervised footage from real-world settings like Tokyo streets, The Matrix allows users to traverse diverse terrains -- deserts, grasslands, water bodies, and urban landscapes -- in continuous, uncut hour-long sequences. Operating at 16 FPS, the system supports real-time interactivity and demonstrates zero-shot generalization, translating virtual game environments to real-world contexts where collecting continuous movement data is often infeasible. For example, The Matrix can simulate a BMW X3 driving through an office setting--an environment present in neither gaming data nor real-world sources. This approach showcases the potential of AAA game data to advance robust world models, bridging the gap between simulations and real-world applications in scenarios with limited data.

Xiangwei Shen, Zhimin Li, Zhantao Yang et al.

Recent studies have demonstrated the effectiveness of directly aligning diffusion models with human preferences using differentiable reward. However, they exhibit two primary challenges: (1) they rely on multistep denoising with gradient computation for reward scoring, which is computationally expensive, thus restricting optimization to only a few diffusion steps; (2) they often need continuous offline adaptation of reward models in order to achieve desired aesthetic quality, such as photorealism or precise lighting effects. To address the limitation of multistep denoising, we propose Direct-Align, a method that predefines a noise prior to effectively recover original images from any time steps via interpolation, leveraging the equation that diffusion states are interpolations between noise and target images, which effectively avoids over-optimization in late timesteps. Furthermore, we introduce Semantic Relative Preference Optimization (SRPO), in which rewards are formulated as text-conditioned signals. This approach enables online adjustment of rewards in response to positive and negative prompt augmentation, thereby reducing the reliance on offline reward fine-tuning. By fine-tuning the FLUX model with optimized denoising and online reward adjustment, we improve its human-evaluated realism and aesthetic quality by over 3x.

Yukun Wang, Ruihuang Li, Jiale Tao et al.

Video fundamentally intertwines two crucial axes: the dynamic content of a scene and the camera motion through which it is observed. However, existing generation models often entangle these factors, limiting independent control. In this work, we introduce OmniCamera, a unified framework designed to explicitly disentangle and command these two dimensions. This compositional approach enables flexible video generation by allowing arbitrary pairings of camera and content conditions, unlocking unprecedented creative control. To overcome the fundamental challenges of modality conflict and data scarcity inherent in such a system, we present two key innovations. First, we construct OmniCAM, a novel hybrid dataset combining curated real-world videos with synthetic data that provides diverse paired examples for robust multi-task learning. Second, we propose a Dual-level Curriculum Co-Training strategy that mitigates modality interference and synergistically learns from diverse data sources. This strategy operates on two levels: first, it progressively introduces control modalities by difficulties (condition-level), and second, trains for precise control on synthetic data before adapting to real data for photorealism (data-level). As a result, OmniCamera achieves state-of-the-art performance, enabling flexible control for complex camera movements while maintaining superior visual quality.

Zhantao Yang, Han Zhang, Fangyi Chen et al.

Knowledge Graph (KG) is playing an increasingly important role in various AI systems. For e-commerce, an efficient and low-cost automated knowledge graph construction method is the foundation of enabling various successful downstream applications. In this paper, we propose a novel method for constructing structured product knowledge graphs from raw product images. The method cooperatively leverages recent advances in the vision-language model (VLM) and large language model (LLM), fully automating the process and allowing timely graph updates. We also present a human-annotated e-commerce product dataset for benchmarking product property extraction in knowledge graph construction. Our method outperforms our baseline in all metrics and evaluated properties, demonstrating its effectiveness and bright usage potential.

Zhantao Yang, Huangji Wang, Ruili Feng et al.

Generating captions for long and complex videos is both critical and challenging, with significant implications for the growing fields of text-to-video generation and multi-modal understanding. One key challenge in long video captioning is accurately recognizing the same individuals who appear in different frames, which we refer to as the ID-Matching problem. Few prior works have focused on this important issue. Those that have, usually suffer from limited generalization and depend on point-wise matching, which limits their overall effectiveness. In this paper, unlike previous approaches, we build upon LVLMs to leverage their powerful priors. We aim to unlock the inherent ID-Matching capabilities within LVLMs themselves to enhance the ID-Matching performance of captions. Specifically, we first introduce a new benchmark for assessing the ID-Matching capabilities of video captions. Using this benchmark, we investigate LVLMs containing GPT-4o, revealing key insights that the performance of ID-Matching can be improved through two methods: 1) enhancing the usage of image information and 2) increasing the quantity of information of individual descriptions. Based on these insights, we propose a novel video captioning method called Recognizing Identities for Captioning Effectively (RICE). Extensive experiments including assessments of caption quality and ID-Matching performance, demonstrate the superiority of our approach. Notably, when implemented on GPT-4o, our RICE improves the precision of ID-Matching from 50% to 90% and improves the recall of ID-Matching from 15% to 80% compared to baseline. RICE makes it possible to continuously track different individuals in the captions of long videos.

Chen Li, Zhantao Yang, Han Zhang et al.

Vision-Language-Action (VLA) models show promise in embodied reasoning, yet remain far from true generalists-they often require task-specific fine-tuning, and generalize poorly to unseen tasks. We propose MetaVLA, a unified, backbone-agnostic post-training framework for efficient and scalable alignment. MetaVLA introduces Context-Aware Meta Co-Training, which consolidates diverse target tasks into a single fine-tuning stage while leveraging structurally diverse auxiliary tasks to improve in-domain generalization. Unlike naive multi-task SFT, MetaVLA integrates a lightweight meta-learning mechanism-derived from Attentive Neural Processes-to enable rapid adaptation from diverse contexts with minimal architectural change or inference overhead. On the LIBERO benchmark, MetaVLA with six auxiliary tasks outperforms OpenVLA by up to 8.0% on long-horizon tasks, reduces training steps from 240K to 75K, and cuts GPU time by ~76%. These results show that scalable, low-resource post-training is achievable-paving the way toward general-purpose embodied agents. Code will be available.

Chen Li, Han Zhang, Zhantao Yang et al.

Vision-language models (VLMs) have made significant strides in reasoning, yet they often struggle with complex multimodal tasks and tend to generate overly verbose outputs. A key limitation is their reliance on chain-of-thought (CoT) reasoning, despite many tasks benefiting from alternative topologies like trees or graphs. To address this, we introduce STELAR-Vision, a training framework for topology-aware reasoning. At its core is TopoAug, a synthetic data pipeline that enriches training with diverse topological structures. Using supervised fine-tuning and reinforcement learning, we post-train Qwen2VL models with both accuracy and efficiency in mind. Additionally, we propose Frugal Learning, which reduces output length with minimal accuracy loss. On MATH-V and VLM-S2H, STELAR-Vision improves accuracy by 9.7% over its base model and surpasses the larger Qwen2VL-72B-Instruct by 7.3%. On five out-of-distribution benchmarks, it outperforms Phi-4-Multimodal-Instruct by up to 28.4% and LLaMA-3.2-11B-Vision-Instruct by up to 13.2%, demonstrating strong generalization. Compared to Chain-Only training, our approach achieves 4.3% higher overall accuracy on in-distribution datasets and consistently outperforms across all OOD benchmarks. We have released datasets, and code will be available.

Shangwen Zhu, Qianyu Peng, Yuting Hu et al.

Flow-based generative models have achieved remarkable progress, with classifier-free guidance (CFG) becoming the standard for high-fidelity generation. However, the conventional practice of applying a strong, fixed guidance scale throughout inference is poorly suited for the rapid, few-step sampling required by modern applications. In this work, we uncover the root cause of this conflict: a fundamental sampling instability where the earliest steps are acutely sensitive to guidance. We trace this to a significant spike in the ratio of conditional to unconditional predictions--a spike that we prove to be an inherent property of the training data distribution itself, making it a almost inevitable challenge. Applying a high, static guidance value during this volatile initial phase leads to an exponential amplification of error, degrading image quality. To resolve this, we propose a simple, theoretically grounded, adaptive guidance schedule that automatically dampens the guidance scale at early steps based on the evolving ratio. Our method is lightweight, incurs no inference overhead, and is compatible with standard frameworks. Experiments across state-of-the-art image (SD3.5, Qwen-Image) and video (WAN2.1) models show our approach enables up to 3x faster sampling while maintaining or improving quality, robustness, and semantic alignment. Our findings highlight that adapting guidance to the sampling process, rather than fixing it, is critical for unlocking the full potential of fast, flow-based models.

Jie He, Vincent Theo Willem Kenbeek, Zhantao Yang et al.

We introduce TD-Interpreter, a specialized ML tool that assists engineers in understanding complex timing diagrams (TDs), originating from a third party, during their design and verification process. TD-Interpreter is a visual question-answer environment which allows engineers to input a set of TDs and ask design and verification queries regarding these TDs. We implemented TD-Interpreter with multimodal learning by fine-tuning LLaVA, a lightweight 7B Multimodal Large Language Model (MLLM). To address limited training data availability, we developed a synthetic data generation workflow that aligns visual information with its textual interpretation. Our experimental evaluation demonstrates the usefulness of TD-Interpreter which outperformed untuned GPT-4o by a large margin on the evaluated benchmarks.

Shangwen Zhu, Han Zhang, Zhantao Yang et al.

Text-based diffusion models have made significant breakthroughs in generating high-quality images and videos from textual descriptions. However, the lengthy sampling time of the denoising process remains a significant bottleneck in practical applications. Previous methods either ignore the statistical relationships between adjacent steps or rely on attention or feature similarity between them, which often only works with specific network structures. To address this issue, we discover a new statistical relationship in the transition operator between adjacent steps, focusing on the relationship of the outputs from the network. This relationship does not impose any requirements on the network structure. Based on this observation, we propose a novel training-free acceleration method called LTC-Accel, which uses the identified relationship to estimate the current transition operator based on adjacent steps. Due to no specific assumptions regarding the network structure, LTC-Accel is applicable to almost all diffusion-based methods and orthogonal to almost all existing acceleration techniques, making it easy to combine with them. Experimental results demonstrate that LTC-Accel significantly speeds up sampling in text-to-image and text-to-video synthesis while maintaining competitive sample quality. Specifically, LTC-Accel achieves a speedup of 1.67-fold in Stable Diffusion v2 and a speedup of 1.55-fold in video generation models. When combined with distillation models, LTC-Accel achieves a remarkable 10-fold speedup in video generation, allowing real-time generation of more than 16FPS.