arXiv subjects

Zixiang Zhao

Publications and source records attributed to Zixiang Zhao.

At least 19 recordsLinked to original sources

arXiv2026

5% > 100%: Flatness Preference is All You Need for Multimodal Parameter-Efficient Fine-Tuning

Parameter-Efficient Fine-Tuning (PEFT) methods provide a streamlined and efficient tool for adapting large models to domain-specific multimodal downstream tasks. Although these methods proved their tangible effects in practice, their principal aspects remain under-explored. Therefore we remain curious about the underlying generalization mechanisms in various PEFT methods and how they can be further enhanced. In this paper, we reveal the flatness preference widely present in various PEFTs, where a small fraction of sharp dimensions dominates the generalization of PEFT. This finding suggests an appealing possibility: we may be satisfied with a better generalization by merely attending to this small fraction of sharp dimensions instead of all of them. Furthermore, we propose Flatness Preference Optimization (FlatPO) to flatten these key sharpness dimensions, leading various PEFTs toward better generalization. Extensive experiments demonstrate the effectiveness of our findings and the proposed method. Code is available at https://github.com/Can-Lin/FlatPO.

cs.CV↗

arXiv2026

Stitched Value Model for Diffusion Alignment

For practical use, diffusion- or flow-based generative models must be aligned with task-specific rewards, such as prompt fidelity or aesthetic preference. That alignment is challenging because the reward is defined for clean output images, but the alignment procedure requires value function estimates at noisy intermediate latents. Existing methods resort to Tweedie-style or Monte Carlo approximations, trading off estimator bias against computational cost: Tweedie estimates are efficient but biased, while Monte Carlo estimates are more accurate but require expensive rollouts. A natural alternative would be a learned value function, but it remains an open question how to effectively train a strong and general value model specifically for noisy latents. Here, we propose StitchVM, a model stitching framework that efficiently transfers reward models pretrained for clean images to the noisy latent regime. StitchVM starts from an existing, truncated pixel-space reward model and attaches a frozen diffusion backbone to it as its head. From the pixel-space model, the resulting hybrid retains a carefully pretrained, robust reward capability; from the diffusion backbone, it inherits its native ability to handle noisy latents. The stitching procedure is exceptionally lightweight, e.g., stitching and finetuning CLIP ViT-L and SD 3.5 Medium takes only 10 GPU-hours. By lifting powerful pixel-space reward models to latent space, StitchVM opens up a new style of diffusion alignment: instead of rough, yet costly per-sample approximation of the value function, the correct function for the actual, noisy latents is constructed once and then amortized over many samples and iterations. We show that this approach yields improvements across a broad range of downstream steering and post-training methods: DPS becomes $3.2\times$ faster while halving peak GPU memory, and DiffusionNFT becomes $2.3\times$ faster.

cs.CV↗

arXiv2026

A Faster Path to Continual Learning

Continual Learning (CL) aims to train neural networks on a dynamic stream of tasks without forgetting previously learned knowledge. Among optimization-based approaches, C-Flat has emerged as a promising solution due to its plug-and-play nature and its ability to encourage uniformly low-loss regions for both new and old tasks. However, C-Flat requires three additional gradient computations per iteration, imposing substantial overhead on the optimization process. In this work, we propose C-Flat Turbo, a faster yet stronger optimizer that significantly reduces the training cost. We show that the gradients associated with first-order flatness contain direction-invariant components relative to the proxy-model gradients, enabling us to skip redundant gradient computations in the perturbed ascent steps. Moreover, we observe that these flatness-promoting gradients progressively stabilize across tasks, which motivates a linear scheduling strategy with an adaptive trigger to allocate larger turbo steps for later tasks. Experiments show that C-Flat Turbo is 1.0$\times$ to 1.25$\times$ faster than C-Flat across a wide range of CL methods, while achieving comparable or even improved accuracy.

cs.LG↗

arXiv2026

FastVMT: Eliminating Redundancy in Video Motion Transfer

Video motion transfer aims to synthesize videos by generating visual content according to a text prompt while transferring the motion pattern observed in a reference video. Recent methods predominantly use the Diffusion Transformer (DiT) architecture. To achieve satisfactory runtime, several methods attempt to accelerate the computations in the DiT, but fail to address structural sources of inefficiency. In this work, we identify and remove two types of computational redundancy in earlier work: motion redundancy arises because the generic DiT architecture does not reflect the fact that frame-to-frame motion is small and smooth; gradient redundancy occurs if one ignores that gradients change slowly along the diffusion trajectory. To mitigate motion redundancy, we mask the corresponding attention layers to a local neighborhood such that interaction weights are not computed unnecessarily distant image regions. To exploit gradient redundancy, we design an optimization scheme that reuses gradients from previous diffusion steps and skips unwarranted gradient computations. On average, FastVMT achieves a 3.43x speedup without degrading the visual fidelity or the temporal consistency of the generated videos.

cs.CV↗

arXiv2026

MambaVF: State Space Model for Efficient Video Fusion

Video fusion is a fundamental technique in various video processing tasks. However, existing video fusion methods heavily rely on optical flow estimation and feature warping, resulting in severe computational overhead and limited scalability. This paper presents MambaVF, an efficient video fusion framework based on state space models (SSMs) that performs temporal modeling without explicit motion estimation. First, by reformulating video fusion as a sequential state update process, MambaVF captures long-range temporal dependencies with linear complexity while significantly reducing computation and memory costs. Second, MambaVF proposes a lightweight SSM-based fusion module that replaces conventional flow-guided alignment via a spatio-temporal bidirectional scanning mechanism. This module enables efficient information aggregation across frames. Extensive experiments across multiple benchmarks demonstrate that our MambaVF achieves state-of-the-art performance in multi-exposure, multi-focus, infrared-visible, and medical video fusion tasks. We highlight that MambaVF enjoys high efficiency, reducing up to 92.25% of parameters and 88.79% of computational FLOPs and a 2.1x speedup compared to existing methods. Project page: https://mambavf.github.io

cs.CV↗

arXiv2026

Continual GUI Agents

As digital environments (data distribution) are in flux, with new GUI data arriving over time-introducing new domains or resolutions-agents trained on static environments deteriorate in performance. In this work, we introduce Continual GUI Agents, a new task that requires GUI agents to perform continual learning under shifted domains and resolutions. We find existing methods fail to maintain stable grounding as GUI distributions shift over time, due to the diversity of UI interaction points and regions in fluxing scenarios. To address this, we introduce GUI-Anchoring in Flux (GUI-AiF), a new reinforcement fine-tuning framework that stabilizes continual learning through two novel rewards: Anchoring Point Reward in Flux (APR-iF) and Anchoring Region Reward in Flux (ARR-iF). These rewards guide the agents to align with shifting interaction points and regions, mitigating the tendency of existing reward strategies to over-adapt to static grounding cues (e.g., fixed coordinates or element scales). Extensive experiments show GUI-AiF surpasses state-of-the-art baselines. Our work establishes the first continual learning framework for GUI agents, revealing the untapped potential of reinforcement fine-tuning for continual GUI Agents.

cs.LG↗

arXiv2025

ArtiWorld: LLM-Driven Articulation of 3D Objects in Scenes

Building interactive simulators and scalable robot-learning environments requires a large number of articulated assets. However, most existing 3D assets in simulation are rigid, and manually converting them into articulated objects is extremely labor- and cost-intensive. This raises a natural question: can we automatically identify articulable objects in a scene and convert them into articulated assets directly? In this paper, we present ArtiWorld, a scene-aware pipeline that localizes candidate articulable objects from textual scene descriptions and reconstructs executable URDF models that preserve the original geometry. At the core of this pipeline is Arti4URDF, which leverages 3D point cloud, prior knowledge of a large language model (LLM), and a URDF-oriented prompt design to rapidly convert rigid objects into interactive URDF-based articulated objects while maintaining their 3D shape. We evaluate ArtiWorld at three levels: 3D simulated objects, full 3D simulated scenes, and real-world scan scenes. Across all three settings, our method consistently outperforms existing approaches and achieves state-of-the-art performance, while preserving object geometry and correctly capturing object interactivity to produce usable URDF-based articulated models. This provides a practical path toward building interactive, robot-ready simulation environments directly from existing 3D assets. Code and data will be released.

cs.CV↗

arXiv2025

Understanding the atomically precise evolution of the miscibility of newly prepared face-centered cubic W-Cu nanoalloys and its asymmetry

According to classical Miedema theory, reducing crystals to the order of nanometer sizes might greatly modulate the mixing enthalpy of elements, thus enabling the invention of a lot of new bulk-immiscible alloys. Although numerous alloys with higher mixing enthalpies remain unexplored, this strategy is approaching its limit, as reflected by the critical diameter of recent alloys of 1.8 nm, which corresponds to ~150 atoms and hardly provides a crystalline order. Future development requires not only even smaller atomic-scale control but also a new surface energy-saving mechanism. Here, we report the formation of W-Cu nanoalloys with a very large miscibility gap in the bulk via the use of an atomically size-selected cluster beam source as an example. The face-centered cubic (FCC) structure was demonstrated through electron diffraction, which indicated a lattice constant of 3.88{\AA} for W0.85Cu0.15 nanoalloys (~2280 atoms). In this comprehensive study that covers a large parameter space of W/Cu compositions and numbers of atoms, an asymmetric miscibility nanophase diagram in which W-rich compositions favor mixing and the critical size is approximately 6000 atoms, which far exceeds the approximately tens of atoms predicted via classical theory, was obtained for the first time. Density functional theory (DFT) calculations revealed a mutual strain-induced mechanism that simultaneously lowers the surface energies while reducing the size to the atomic scale. This approach paves the way for the development of new high-performance nonequilibrium phase alloys.

cond-mat.mtrl-sci↗

arXiv2025

C-Flat++: Towards a More Efficient and Powerful Framework for Continual Learning

Balancing sensitivity to new tasks and stability for retaining past knowledge is crucial in continual learning (CL). Recently, sharpness-aware minimization has proven effective in transfer learning and has also been adopted in continual learning (CL) to improve memory retention and learning efficiency. However, relying on zeroth-order sharpness alone may favor sharper minima over flatter ones in certain settings, leading to less robust and potentially suboptimal solutions. In this paper, we propose \textbf{C}ontinual \textbf{Flat}ness (\textbf{C-Flat}), a method that promotes flatter loss landscapes tailored for CL. C-Flat offers plug-and-play compatibility, enabling easy integration with minimal modifications to the code pipeline. Besides, we present a general framework that integrates C-Flat into all major CL paradigms and conduct comprehensive comparisons with loss-minima optimizers and flat-minima-based CL methods. Our results show that C-Flat consistently improves performance across a wide range of settings. In addition, we introduce C-Flat++, an efficient yet effective framework that leverages selective flatness-driven promotion, significantly reducing the update cost required by C-Flat. Extensive experiments across multiple CL methods, datasets, and scenarios demonstrate the effectiveness and efficiency of our proposed approaches. Code is available at https://github.com/WanNaa/C-Flat.

cs.LG↗

arXiv2025

Towards Understanding The Calibration Benefits of Sharpness-Aware Minimization

Deep neural networks have been increasingly used in safety-critical applications such as medical diagnosis and autonomous driving. However, many studies suggest that they are prone to being poorly calibrated and have a propensity for overconfidence, which may have disastrous consequences. In this paper, unlike standard training such as stochastic gradient descent, we show that the recently proposed sharpness-aware minimization (SAM) counteracts this tendency towards overconfidence. The theoretical analysis suggests that SAM allows us to learn models that are already well-calibrated by implicitly maximizing the entropy of the predictive distribution. Inspired by this finding, we further propose a variant of SAM, coined as CSAM, to ameliorate model calibration. Extensive experiments on various datasets, including ImageNet-1K, demonstrate the benefits of SAM in reducing calibration error. Meanwhile, CSAM performs even better than SAM and consistently achieves lower calibration error than other approaches

cs.LG↗

arXiv2025

Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers

Diffusion transformers (DiT) have demonstrated exceptional performance in video generation. However, their large number of parameters and high computational complexity limit their deployment on edge devices. Quantization can reduce storage requirements and accelerate inference by lowering the bit-width of model parameters. Yet, existing quantization methods for image generation models do not generalize well to video generation tasks. We identify two primary challenges: the loss of information during quantization and the misalignment between optimization objectives and the unique requirements of video generation. To address these challenges, we present Q-VDiT, a quantization framework specifically designed for video DiT models. From the quantization perspective, we propose the Token-aware Quantization Estimator (TQE), which compensates for quantization errors in both the token and feature dimensions. From the optimization perspective, we introduce Temporal Maintenance Distillation (TMD), which preserves the spatiotemporal correlations between frames and enables the optimization of each frame with respect to the overall video context. Our W3A6 Q-VDiT achieves a scene consistency of 23.40, setting a new benchmark and outperforming current state-of-the-art quantization methods by 1.9$\times$. Code will be available at https://github.com/cantbebetter2/Q-VDiT.

cs.CV↗

arXiv2025

A Unified Solution to Video Fusion: From Multi-Frame Learning to Benchmarking

The real world is dynamic, yet most image fusion methods process static frames independently, ignoring temporal correlations in videos and leading to flickering and temporal inconsistency. To address this, we propose Unified Video Fusion (UniVF), a novel and unified framework for video fusion that leverages multi-frame learning and optical flow-based feature warping for informative, temporally coherent video fusion. To support its development, we also introduce Video Fusion Benchmark (VF-Bench), the first comprehensive benchmark covering four video fusion tasks: multi-exposure, multi-focus, infrared-visible, and medical fusion. VF-Bench provides high-quality, well-aligned video pairs obtained through synthetic data generation and rigorous curation from existing datasets, with a unified evaluation protocol that jointly assesses the spatial quality and temporal consistency of video fusion. Extensive experiments show that UniVF achieves state-of-the-art results across all tasks on VF-Bench. Project page: https://vfbench.github.io.

cs.CV↗

arXiv2025

RefComp: A Reference-guided Unified Framework for Unpaired Point Cloud Completion

The unpaired point cloud completion task aims to complete a partial point cloud by using models trained with no ground truth. Existing unpaired point cloud completion methods are class-aware, i.e., a separate model is needed for each object class. Since they have limited generalization capabilities, these methods perform poorly in real-world scenarios when confronted with a wide range of point clouds of generic 3D objects. In this paper, we propose a novel unpaired point cloud completion framework, namely the Reference-guided Completion (RefComp) framework, which attains strong performance in both the class-aware and class-agnostic training settings. The RefComp framework transforms the unpaired completion problem into a shape translation problem, which is solved in the latent feature space of the partial point clouds. To this end, we introduce the use of partial-complete point cloud pairs, which are retrieved by using the partial point cloud to be completed as a template. These point cloud pairs are used as reference data to guide the completion process. Our RefComp framework uses a reference branch and a target branch with shared parameters for shape fusion and shape translation via a Latent Shape Fusion Module (LSFM) to enhance the structural features along the completion pipeline. Extensive experiments demonstrate that the RefComp framework achieves not only state-of-the-art performance in the class-aware training setting but also competitive results in the class-agnostic training setting on both virtual scans and real-world datasets.

cs.CV↗

arXiv2025

Task-driven Image Fusion with Learnable Fusion Loss

Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TDFusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks.

cs.CV↗

arXiv2025

Retinex-MEF: Retinex-based Glare Effects Aware Unsupervised Multi-Exposure Image Fusion

Multi-exposure image fusion (MEF) synthesizes multiple, differently exposed images of the same scene into a single, well-exposed composite. Retinex theory, which separates image illumination from scene reflectance, provides a natural framework to ensure consistent scene representation and effective information fusion across varied exposure levels. However, the conventional pixel-wise multiplication of illumination and reflectance inadequately models the glare effect induced by overexposure. To address this limitation, we introduce an unsupervised and controllable method termed Retinex-MEF. Specifically, our method decomposes multi-exposure images into separate illumination components with a shared reflectance component, and effectively models the glare induced by overexposure. The shared reflectance is learned via a bidirectional loss, which enables our approach to effectively mitigate the glare effect. Furthermore, we introduce a controllable exposure fusion criterion, enabling global exposure adjustments while preserving contrast, thus overcoming the constraints of a fixed exposure level. Extensive experiments on diverse datasets, including underexposure-overexposure fusion, exposure controlled fusion, and homogeneous extreme exposure fusion, demonstrate the effective decomposition and flexible fusion capability of our model. The code is available at https://github.com/HaowenBai/Retinex-MEF

cs.CV↗

arXiv2025

ReFusion: Learning Image Fusion from Reconstruction with Learnable Loss via Meta-Learning

Image fusion aims to combine information from multiple source images into a single one with more comprehensive informational content. Deep learning-based image fusion algorithms face significant challenges, including the lack of a definitive ground truth and the corresponding distance measurement. Additionally, current manually defined loss functions limit the model's flexibility and generalizability for various fusion tasks. To address these limitations, we propose ReFusion, a unified meta-learning based image fusion framework that dynamically optimizes the fusion loss for various tasks through source image reconstruction. Compared to existing methods, ReFusion employs a parameterized loss function, that allows the training framework to be dynamically adapted according to the specific fusion scenario and task. ReFusion consists of three key components: a fusion module, a source reconstruction module, and a loss proposal module. We employ a meta-learning strategy to train the loss proposal module using the reconstruction loss. This strategy forces the fused image to be more conducive to reconstruct source images, allowing the loss proposal module to generate a adaptive fusion loss that preserves the optimal information from the source images. The update of the fusion module relies on the learnable fusion loss proposed by the loss proposal module. The three modules update alternately, enhancing each other to optimize the fusion loss for different tasks and consistently achieve satisfactory results. Extensive experiments demonstrate that ReFusion is capable of adapting to various tasks, including infrared-visible, medical, multi-focus, and multi-exposure image fusion.

cs.CV↗

arXiv2025

Deep Unfolding Multi-modal Image Fusion Network via Attribution Analysis

Multi-modal image fusion synthesizes information from multiple sources into a single image, facilitating downstream tasks such as semantic segmentation. Current approaches primarily focus on acquiring informative fusion images at the visual display stratum through intricate mappings. Although some approaches attempt to jointly optimize image fusion and downstream tasks, these efforts often lack direct guidance or interaction, serving only to assist with a predefined fusion loss. To address this, we propose an ``Unfolding Attribution Analysis Fusion network'' (UAAFusion), using attribution analysis to tailor fused images more effectively for semantic segmentation, enhancing the interaction between the fusion and segmentation. Specifically, we utilize attribution analysis techniques to explore the contributions of semantic regions in the source images to task discrimination. At the same time, our fusion algorithm incorporates more beneficial features from the source images, thereby allowing the segmentation to guide the fusion process. Our method constructs a model-driven unfolding network that uses optimization objectives derived from attribution analysis, with an attribution fusion loss calculated from the current state of the segmentation network. We also develop a new pathway function for attribution analysis, specifically tailored to the fusion tasks in our unfolding network. An attribution attention mechanism is integrated at each network stage, allowing the fusion network to prioritize areas and pixels crucial for high-level recognition tasks. Additionally, to mitigate the information loss in traditional unfolding networks, a memory augmentation module is incorporated into our network to improve the information flow across various network layers. Extensive experiments demonstrate our method's superiority in image fusion and applicability to semantic segmentation.

cs.CV↗

arXiv2025

Simultaneous Automatic Picking and Manual Picking Refinement for First-Break

First-break picking is a pivotal procedure in processing microseismic data for geophysics and resource exploration. Recent advancements in deep learning have catalyzed the evolution of automated methods for identifying first-break. Nevertheless, the complexity of seismic data acquisition and the requirement for detailed, expert-driven labeling often result in outliers and potential mislabeling within manually labeled datasets. These issues can negatively affect the training of neural networks, necessitating algorithms that handle outliers or mislabeled data effectively. We introduce the Simultaneous Picking and Refinement (SPR) algorithm, designed to handle datasets plagued by outlier samples or even noisy labels. Unlike conventional approaches that regard manual picks as ground truth, our method treats the true first-break as a latent variable within a probabilistic model that includes a first-break labeling prior. SPR aims to uncover this variable, enabling dynamic adjustments and improved accuracy across the dataset. This strategy mitigates the impact of outliers or inaccuracies in manual labels. Intra-site picking experiments and cross-site generalization experiments on publicly available data confirm our method's performance in identifying first-break and its generalization across different sites. Additionally, our investigations into noisy signals and labels underscore SPR's resilience to both types of noise and its capability to refine misaligned manual annotations. Moreover, the flexibility of SPR, not being limited to any single network architecture, enhances its adaptability across various deep learning-based picking methods. Focusing on learning from data that may contain outliers or partial inaccuracies, SPR provides a robust solution to some of the principal obstacles in automatic first-break picking.

cs.CV↗