Learning a group of weights to combine view features forms many popular schemes in multi-view clustering. The learning ways mainly include two classes: explicit and implicit. For the explicit weight learning ways, there are mainly Norm Regularization (NR) and Exponential Decay (ED). In the NR, the p-norm Norm Regularization (pNR) and Negative Entropy (NE) can be adopted. Moreover, Self-paced Learning is also included in the NR. For the implicit weight learning ways, there are mainly q-Root Loss (qRL), Logarithm Self-weighted Loss (LSL) and Capped Self-weighted Loss (CSL) and so on. Further, these implicit weight ways can be extended to a more generalized implicit weight learning framework. The above involved numerous methods are the special case. In this paper, we firstly show the explicit and implicit weight learning ways in which the multiple ways are set. Then, considering the connections among pNR, ED and qRL three paradigms, a Unified Paradigm (UP) which can represent the above three paradigms by setting different constraints is introduced. From the multiple weight learning manners, we can find that pNR, ED, NE, qRL, LSL and CSL six learning paradigms are used to learn view weights. However, the SPL is used to learn sample weights. Next, the weight sparsity, hyper-parameter setting and limiting behavior of these schemes are discussed. Finally, we conduct experiments on seven multi-view data sets. The experimental results support the theoretical analysis.
Pre-trained 3D vision models have substantially advanced point cloud analysis, yet adapting them to downstream tasks via full fine-tuning is computationally expensive and storage-intensive. Parameter-Efficient Fine-Tuning (PEFT) offers a promising alternative by reducing both adaptation cost and storage burden. However, existing prompting-based approaches ignore the intrinsic geometric structures of point clouds, thereby limiting their adaptation capability. This limitation stems from their inability to encode both fine-grained geometric cues and coarse-grained structural semantics, as well as failing to propagate such information effectively through the model hierarchy. To address these challenges, we propose GAPrompt++, a multi-granular geometry-aware prompting method that provides richer geometric guidance for efficient 3D task adaptation. Specifically, we introduce a Point Shift Prompter that extracts multi-granular geometric features across different scales, enabling instance-specific geometric adjustments during adaptation. Next, a Keypoint Prompter adaptively generates point-level prompts to highlight local geometric saliency and fine-grained structural details. Furthermore, a Prompt Propagation mechanism injects these multi-granular geometric cues throughout the feature extraction hierarchy, strengthening the ability to capture essential geometric characteristics. Extensive experiments show that GAPrompt++ achieves state-of-the-art performance among prompting-based PEFT methods and even surpasses full fine-tuning across diverse benchmarks, while requiring less than 2% trainable parameters. In addition, to address the saturation of existing evaluation datasets, we construct two more challenging benchmarks derived from 3D Gaussian Splatting and Multi-View Stereo reconstruction, offering diverse and realistic point cloud scenarios to promote future research.
Current Multiview Graph Clustering (MGC) methods primarily focus on fusing multiple graphs and/or spectral embeddings to seek common clustering sets, however, the tendency between nodes and the locality between clusters and views are scarcely considered. Specifically, there are consistent topological structures in multiview graphs, which imply the tendency for pairwise nodes to be connected or disconnected across views. The locality refers to specific local structures, where some clusters exhibit clear discriminative patterns in partial views, whereas these patterns are broken in other views. With this awareness, we first propose a Credible Signed Information (CSI) extraction module to capture the tendency across all views and promote raw multiview graphs into signed multiview graphs. Building on CSI, we develop a novel mixed local graph cut model CSI-MGC to effectively capture the complex locality between clusters, views, and the signed graphs using a three-layer weight learning scheme. To solve the optimization problem involved in CSI-MGC, we propose an efficient discrete optimization algorithm and provide corresponding theoretical analyses. Finally, we conduct extensive experiments on eight benchmark datasets and against 13 state-of-the-art competitors and the results demonstrate the effectiveness of our proposals.
The increasing reliance on large-scale datasets imposes significant storage and computational burdens on training deep learning models. Dataset distillation methods, particularly those based on sample generation, aim to condense large original datasets into compact synthetic sets while preserving essential information. Existing subset synthesis approaches typically minimize a homogeneous distance, assigning uniform contributions from all real instances to the construction of each synthetic sample. We show that such equal allocation neglects instance-level relationships between real-synthetic pairs, leading to inadequate modeling of the geometric structural discrepancies between the distilled and original datasets. In this work, we reformulate homogeneous distance minimization as a bi-level optimization problem via a matching-and-approximating paradigm. In the matching stage, we employ an optimal transport matrix to dynamically allocate contributions from real instances. Building upon this transport-based allocation, we further introduce a difficulty-aware marginal reweighting mechanism to emphasize informative instances while preserving global geometric consistency. In the subsequent approximation stage, synthetic samples are refined according to the established allocation scheme to better approximate the real data distribution. This strategy enables a more faithful characterization of intricate geometric structures and improved handling of intra-class variations, thereby enhancing distillation fidelity. Extensive experiments across diverse architectures, modalities, and learning paradigms demonstrate that the proposed framework consistently improves performance, with gains observed in standard supervised, federated, continual, and multimodal learning settings.
Multi-modal homography estimation aims to align multi-modal images captured by different sensors. Existing methods are largely confined to clean images, neglecting the noise degradation in real-world imaging scenarios. To tackle this issue, we propose a novel diffusion-based homography estimation framework namely HomoDiff, for noise-robust multi modal homography estimation. The design of our network is motivated by an interesting finding, revealing that stronger denoising does not always result in higher registration accuracy. Inspired by this finding, we formulate the registration process in a novel diffusion-based architecture. Specifically, in each inference step, a Multi-modal Image Denoising (MID) subnet is designed to remove the noise with the multi-modal guidance, followed by a Multi-modal Homography Estimation (MHE) subnet to predict the homography matrix at the current denoising intensity. Finally, an Adaptive Homography Selection (AHS) subnet is used to select the optimal homography matrix from all inference steps with different denoising intensities. Through this selection strategy, we can avoid over-denoising to maintain consistently high registration accuracy across different noise levels. To the best of our knowledge, our HomoDiff is the first attempt to address the multi-modal homography estimation problem with noise degradation, which can handle varying noise levels with a single unified framework. Extensive experimental results show that our HomoDiff achieves high registration accuracy on various multi-modal datasets with different noise levels. The source codes are available at https://github.com/Langweng/HomoDiff.git.
Spatio-Temporal Graph Neural Networks (STGNNs) have emerged as powerful tools for modeling dynamic systems, yet they face critical generalization challenges in streaming Spatio-Temporal Out-of-Distribution (STOOD) scenarios, where spatial structures and temporal dynamics evolve beyond the distributions observed in historical training periods. Existing approaches encounter a fundamental dilemma: parametric methods suffer from catastrophic forgetting and gradient conflicts, while replay-based methods struggle to distinguish cross-period recurring patterns from period-specific transient patterns. To overcome these limitations, we propose PRAL (Pattern Retrieval-Augmented Learning for Dynamic Spatio-Temporal Graphs), a novel framework that explicitly stores historical knowledge in external pattern libraries rather than implicitly encoding it in model parameters. PRAL constructs multidimensional pattern libraries that capture representative patterns across spatial, temporal, and Spatio Temporal dimensions, where pattern keys serve as efficient retrieval indices while pattern values encapsulate rich contextual features. Recurring and transient components within the stored patterns are further distinguished through the Stable Subspace Projector (SSP) and Shift-Aware Retrieval Reweighting (SR2) mechanisms. After training, PRAL retrieves relevant patterns via similarity-based matching and adaptively fuses them with current observations through a learnable balancing mechanism. Extensive experiments on multiple real-world streaming graph datasets demonstrate that PRAL achieves the best average performance among all compared methods, with an average MAE/RMSE/MAPE improvement of 15-30% over the strongest non-retrieval baselines, while trading off against STRAP on certain backbone-dataset configurations; it further achieves robust generalization in streaming STOOD settings without task-specific fine-tuning.
Three-dimensional quantitative electromagnetic imaging is essential for accurately characterizing concealed or geometrically complex targets. Existing approaches, however, predominantly rely on point-based representations or 2D reconstruction paradigms, which often require strong target priors, neglect 3D geometric dependencies, and inadequately incorporate the physics of electromagnetic scattering. In this work, we propose EDMA, a physics-informed diffusion framework that directly reconstructs complete 3D meshes together with their associated constitutive parameters from measured electromagnetic fields. Built upon a conditional scatter-to-mesh diffusion formulation, EDMA integrates current-consistency constraints, realized through a lightweight wavelet-transform-enhanced ResMLP for efficient induced-current prediction, with an aperture-adaptive encoding mechanism for robust operation under diverse and incomplete measurement configurations. By explicitly enforcing Maxwell-consistent coupling among the scattered field, contrast function, and induced current, EDMA provides a physically interpretable reconstruction pathway in which the generated object is constrained and verifiable through electromagnetic consistency rather than solely by data-driven fitting. Extensive experiments on multiple 3D benchmark datasets and realistic electromagnetic scenarios demonstrate that EDMA achieves superior reconstruction quality while maintaining efficient single-step inference compared with state-of-the-art methods. Moreover, EDMA exhibits strong generalization and robustness under limited-aperture measurement conditions, making it a promising solution for practical 3D quantitative electromagnetic imaging applications.
In most existing multi-view clustering scenarios, samples of the same object across views are typically assumed to be strictly aligned, forming a key prerequisite for cross-view consistency. In practice, however, view-wise data acquisition and processing are often independent, making the View-unaligned Problem (VuP) ubiquitous and non-negligible. Several pioneering methods have been proposed to address VuP, but their computational efficiency remains a major bottleneck in large-scale applications. Therefore, there is an urgent need for more efficient solutions to the VuP in multi-view clustering. Motivated by this bottleneck, we propose an efficient solution to the VuP in multi-view clustering, termed Scalable Graph-Based Clustering for Unaligned Multi-View Data (SGCU). Specifically, SGCU aligns samples across views via sample permutation and learns a cross-view consensus bipartite graph: it models the sample permutation matrix of each view as an entropy-regularized optimal transport (OT) plan and incorporates a Probabilistic Sparse Sinkhorn solver for efficient optimization. This solver maintains an unbiased approximation to the full solution while reducing the sample alignment complexity to near-linear in sample size. Moreover, the proposed SGCU model can be seamlessly integrated into existing VuP-oriented clustering frameworks to enhance computational efficiency. Extensive experiments on multiple real-world datasets demonstrate its effectiveness, efficiency, and robustness.
Customized generation synthesizes user-specified visual concepts into novel contexts conditioned on prompts. Scaling the number of concepts in customized generation meets a broader demand for personalization, whereas existing methods face challenges with generation quality and computational efficiency. In this paper, we introduce a novel Latent Textual Space (LTS), which is positioned after the text encoder and a linear projection layer. We identify that LTS is the pivot in text-conditional diffusion models, offering three key merits for scaling customized generation: 1) sufficient customized information, 2) effective mitigation of denoising deviation, and 3) excellent scalability. We devise a customization framework to embed each individual concept into compact features in LTS, serving as a versatile concept representation for customized generation. Based on the introduced LTS, we develop LATEXBLEND-I/G/V inference pipelines for three prominent customized generation tasks, enabling seamless plug-and-play combination of multiple concepts in customized text-to-image (T2I), grounded T2I, and inflation-based text-to-video (T2V) generation. These pipelines are built upon the core idea of Blending multiple single-concept representations in the Latent Textual Space, with dedicated designs to address the unique challenges specific to each task. Extensive experiments demonstrate that our method effectively addresses key challenges in scaling multi-concept customized generation, outperforming baselines in both generation quality and computational efficiency
Depression ranks as one of the most prevalent psychological disorders and has received growing attention. Currently, there are two primary challenges in automatic video depression detection: the scarcity of labeled depressed data and effective feature representations characterizing human depression. To address these challenges, this work fully considers the correlation between human personality traits and depression, and proposes a Personality-guided Cross-domain Multitask Learning Network (PerDepNet) for automated video Depression detection. PerDepNet consists of a shared encoder for learning spatiotemporal joint representations, and a personality-guided decoder for cross-domain multitask learning tasks. The encoder contains a cross-scale shared feature extractor for capturing interactive spatial joint features at different scales, and a mamba-based temporal feature extractor for modeling long-term dynamics of depressed videos in personality and depression domains. The personality-guided decoder integrates the task of predicting personality traits into the depression detection task, enabling the personality traits to guide the estimate of depression levels in an adaptive weighted balance module. This is the first attempt to consider the role of personality traits in the depression detection task, thereby improving the performance of depression detection. Extensive experiments on three public depression video datasets demonstrate the superiority of the proposed method over state-of-the-art methods.
Source-free Domain Adaptation (SFDA) realizes cross-domain transfer by adapting a pretrained source model to an unlabeled target domain without accessing data in the source domain. Existing SFDA methods mainly develop subdomain-distribution-oriented framework to extract the fine-grained communal semantics from the target domain. However, they may neglect the facts that (a) smaller knowledge subsets exist within subcategories and overlaps may occur among some of these subsets; and (b) an object may appear at different granularities in the source and target domains so the semantics between them cannot be fully aligned. To this end, this study proposes a Target-oriented Multi-granular Semantic-aware Network (TMSN) for SFDA, including (a) Semantic-aware Micro-community Learning (SML) that introduces micro-communities as small-scale knowledge subsets, employs a learnable semantic-aware module and a semantic-promotability-based segmentation mechanism to evaluate the assignment of target samples to micro-communities, and establishes a multi-scale mapping from the target set into the micro-communities based on the learned assignment patterns; and (b) Target-oriented Semantic Promotability Enhancement (TSPE) that leverages semantic promotability to allocate features from the target domain to micro-communities in the same superordinate unit, and push away those in the different superordinate unit. In this way, the overlapping and asymmetrical relationships across micro-communities are captured, while the multi-granular semantics is learned. Building on the micro-community, we formally show the feasibility of SML and TSPE in batch training, and extensive experiments conducted in close-set and partial-set SFDA scenarios and in object detection demonstrate the superiority of TMSN. Source codes are available at https://github.com/zzf495/TMSN.
When the camera is facing the light source, light scattering and reflections within the lens system may produce both scattering and reflective flares in the captured image. While scattering flares are typically radial and have been largely addressed by existing flare removal methods via modeling in the image spatial domain, reflective flares (which appear as bright spots or ghostings of the light source) are more challenging to remove as they may have similar intensity values and shapes to the light source in the camera-finished images. We observe that in an image burst, reflective flares tend to present larger spatial-temporal disturbances than the light source, as their disturbances tend to be amplified internally by the lens system during the camera shake. This phenomenon can be used to help separate reflective flares from light sources for detection. Inspired by this, we propose a novel neural approach for removing reflective flares based on image bursts. We first formulate an image formation model to model the “moving” reflective flares in a burst, which produces realistic data for training deep models. We then propose a novel neural network with a novel reflective flare localization (REFO) module for locating the reflective flares, a novel guided burst-feature alignment (GBFA) module for handling the spatial and color misalignments caused by camera shakes, and an adaptive fusion (ADFU) module for removing reflective flares. Extensive experiments show that our approach outperforms state-of-the-art methods and generalizes well on image bursts captured in the wild.
Learning structured lane topology from visual observations is challenging due to structural ambiguity, occlusions, and incomplete lane markings in complex urban environments. Most existing centerline graph learning methods rely on deterministic regression, which implicitly assumes a single optimal topology for each scene and therefore struggles under uncertain visual conditions. In this work, we propose LaneDiffusion++, a unified feature-level generative framework that reformulates centerline graph learning as the conditional restoration of prior-enhanced Bird's Eye View (BEV) features, rather than directly generating vectorized graphs in a discrete output space. Specifically, LaneDiffusion++ first injects structural lane priors into intermediate BEV features through a Lane Prior Injection Module (LPIM), and then employs a Lane Prior Diffusion Module (LPDM) to perform conditional generative modeling in feature space, progressively restoring topology-aware representations while capturing multiple plausible lane configurations. To further exploit the stochasticity of diffusion inference, we introduce a Lane Hypothesis Aggregation Module (LHAM), which explicitly consolidates multiple diffusion-generated feature hypotheses into a stable topology-aware representation before decoding. Extensive experiments on large-scale benchmarks demonstrate consistent improvements in both geometric precision and topological consistency over strong deterministic and generative baselines. These results suggest that feature-level generative modeling, together with uncertainty-aware hypothesis aggregation, provides an effective paradigm for structured topology reasoning in autonomous perception.
Multi-agent reinforcement learning (MARL) is a powerful paradigm for large-scale collaborative scenarios, yet it is often hampered by partial observability and non-stationarity. While communication can alleviate these issues, designing efficient protocols remains a significant challenge, especially in decentralized settings. Many existing methods suffer from high communication volume and training complexity. To overcome these limitations, we propose CSN, which enables efficient Communication with Skill Neurons in decentralized MARL by exchanging the essential components of learned knowledge at neuron level. Specifically, we first identify skill neurons, a subset of neurons that encode the most critical knowledge acquired during local training. By communicating only a sparse subset of model parameters and doing so intermittently, CSN fundamentally reduces communication volume by avoiding redundant information and frequent communication. Extensive experiments on the SMAC, SMACv2, MPE and Predator-Prey benchmarks validate that CSN significantly outperforms state-of-the-art methods in both performance and communication efficiency, verifying its effectiveness in decentralized MARL.
For RGB-D salient object detection (SOD), a fundamental challenge lies in establishing effective cross-modality interactions between the input graphic domain (RGB and depth modalities) and the output saliency domain. While existing deep learning methods primarily focus on modeling image-level consistency through carefully designed feature extraction and fusion modules, this paper reveals a crucial discovery: the discrepancy in the phase component between the graphic and saliency domains can be reasonably approximated by a Laplacian distribution. Inspired by this observation, we establish a RGB-D-Saliency phase consistency model for RGB-D SOD, which explicitly formulates the cross-modality relationship from a phase perspective. Building upon this model, we develop a novel Phase Updating and Transform Network (PUTNet). It performs two key operations including (a) updating the saliency phase component under RGB-D-Saliency phase consistency constraints, and (b) generating the final saliency map through global phase-to-image transform. Extensive experiments across nine different RGB-D SOD datasets demonstrate that PUTNet outperforms state-of-the-art methods in both quantitative and qualitative evaluations.
The existing works focus on extracting distinguishable features of a scene to enhance the performance of place recognition. However, the fact is ignored that the place recognition task is a process of probability learning, which still lacks comprehensive investigations. That is, the retrieved point cloud results may have multiple candidates. Therefore, in this paper, we formulate place recognition as an uncertainty probability problem. First, inspired by the uncertainty probability of quantum theory, we model point cloud-based place recognition in quantum states and propose Point Cloud Hilbert Space, which consists of six components including meta, point, point clouds, abstraction, combination, and positioning. This modeling endows Point Cloud Hilbert Space with the capability of uncertainty delivery from points to combinations, and then the uncertainty is eliminated by positioning measurement to generate the ultimate retrieved result. Second, to concretize this model for the place recognition task, we design a quantum-inspired framework, dubbed as QuantumPR, which includes point cloud embedding, point cloud mixture, and point cloud measurement. These three modules separately fulfill the quantum state embedding of point clouds, the mixture of quantum states, and the measurement of quantum states. The QuantumPR provides a feasible solution, which is also interpretable inheriting from the Point Cloud Hilbert Space. Third, to implement the above framework, we devise an instance network, where the amplitude-phase representation is adopted for initializing quantum states, and five QuantumPointConv blocks are established for exploring relations among points. Moreover, geometric-mean-pooling and half-output dimensions are adopted to lighten the network and get the combination. This network can be flexibly applied to other diverse methods. Finally, extensive experiments demonstrate the efficiency and generalizability of the proposed method for point cloud-based place recognition. By establishing the first bridge between quantum-theoretic uncertainty modeling and large-scale place recognition, we offer a mathematically interpretation and provide a reusable formalism for future uncertainty-aware place recognition.
Despite significant advancements in multimodal learning (MML), it has been unexpectedly shown to underperform compared to unimodal approaches in practice, largely due to the modality imbalance problem, ultimately affecting the overall performance of the model. Naturally, most existing methods aim to rebalance optimization speeds across different modalities to avoid performance degeneration caused by modality imbalance. However, in addition to task-oriented modality fusion, we experimentally find that multimodal learning requires explicit modality alignment to stimulate weak modal capabilities so that they can be fully exploited, which is ignored by existing works. Therefore, in this paper, we explore the impact of modality fusion and alignment on multimodal learning from a unified perspective, and develops a dynamic strategy that jointly optimizes both, with particular emphasis on addressing modality imbalance. Concretely, we initially design a soft alignment strategy to impose the positive intervention from the prediction level by integrating modality fusion and alignment into a unified framework. We further extend this strategy to the representation level and hybrid level, enabling compatibility with a wider range of architectures. Subsequently, we design a heuristic strategy to dynamically integrate fusion and alignment. Furthermore, we develop a learning-based strategy using a bi-level optimization framework and theoretically prove the convergence of the learning algorithm to ensure its reliability. These two dynamic integration strategies are incorporated into a unified framework applicable to both supervised and semi-supervised scenarios, further enhancing performance. We conduct a series of experiments to demonstrate the effectiveness of our method on diverse datasets. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art multimodal learning approaches, achieving accuracy improvements of 1.30%, 2.69%, and 0.60% on representative bimodal benchmarks, namely KSounds, CREMA-D, andSarcasm, respectively, as well as gains of 1.35% and 0.65% on trimodal datasets, namely NVGesture and IEMOCAP.
Optical flow forms a fundamental information for various motion related vision problems: e.g., SLAM, visual odometry, and object motion estimation. Event cameras are ideal vision sensors for on-line, dynamic tasks that require optical flow estimation, as they have high temporal resolution, high dynamic range and low latency. However, efficiently decoding optical flow from events for high frequency operation while maintaining accuracy is still an open problem. Batch-based optical flow algorithms (CNN or contrast maximisation) accumulate event data over a short period of time and achieve state-of-the-art performance in terms of accuracy, but at the cost of algorithm latency and lower update rates (on par with traditional cameras). In contrast, event-by-event algorithms only compute flow vectors in small, local regions, achieving a lower latency, but losing accuracy when global information is ignored. In this paper, we introduce a spatio-temporal registration framework to increase accuracy of current state-of-the-art event-by-event flow estimation, while also introducing a twofold algorithm acceleration approach and a real-time implementation strategy to mitigate the impact of computation scaling with event rate. We evaluate our event by-event optical flow algorithm on MVSEC, achieving state-of-the-art results for event-by-event algorithms, and performance comparable to batch-based methods. Our method is also computationally efficient, enabling processing of the higher resolution DSEC dataset, and is the only event-by-event algorithm tested to run completely in real time. Furthermore, we demonstrate its effectiveness and efficiency through qualitative evaluations on the ECD and the high-resolution M3ED datasets. Finally, we introduce a moving object dataset, which is outside the autonomous driving domain, to evaluate the general applicability of the proposed optical flow algorithm. The code is available open-source [CODE AVAILABLE ON ACCEPTANCE].
Zero-shot Learning (ZSL) enables classifiers to recognize classes unseen during training, commonly via generative two stage methods: (1) learn visual semantic correlations from seen classes; (2) synthesize unseen class features from semantics to train classifiers. In this paper, we identify spurious visual semantic correlations in existing generative ZSL worsened by scarce seen class samples and introduce two metrics to quantify spuriousness for seen and unseen classes. Furthermore, we point out a more critical bottleneck: existing unadaptive fully noised generators produce features disconnected from real test samples, which also leads to the spurious correlation. To enhance the visual-semantic correlations on both seen and unseen classes, we propose ZeroDiff++, a diffusion-based generative framework. In training, ZeroDiff++ uses (i) diffusion augmentation to produce diverse noised samples, (ii) supervised contrastive (SC) representations for instance level semantics, and (iii) multi-view discriminators with Wasserstein mutual learning to assess generated features. At generation time, we introduce (iv) Diffusion-based Test time Adaptation (DiffTTA) to adapt the generator using pseudo label reconstruction, and (v) Diffusion-based Test time Generation (DiffGen) to trace the diffusion denoising path and produce partially synthesized features that connect real and generated data, and mitigates data scarcity further. Extensive experiments on three ZSL benchmarks demonstrate that ZeroDiff++ not only achieves significant improvements over existing ZSL methods but also maintains robust performance even with scarce training data.
Fashion image editing demands high-dimensional, fine-grained control to follow personalized, unpredictable natural-language instructions. Yet current methods are limited by a fundamental trade-off: fashion-specific approaches offer structural accuracy but lack semantic flexibility, while general text-driven editors are semantically flexible but structurally inaccurate. To bridge this gap, we propose Pose-Star++, a training-free, plug-and-play framework that introduces two core innovations: an LVLM-based Understanding Module that shifts from word- to sentence-level semantic-visual comprehension, eliminating cumbersome instruction pre-parsing and enabling robust understanding of complex natural language; a Bidirectional Calibration Module that co-optimizes semantic and structural constraints through forward pose-guided and backward attention-guided refinement, achieving precise, whole-body-reachable region calibration even under challenging in-the-wild poses. We further contribute the first real-world-oriented fashion-editing benchmark with diverse data, instructions, and tasks, exposing long-overlooked practical challenges. Extensive experiments demonstrate that Pose-Star++ significantly outperforms existing methods in semantic alignment, pose robustness, and in-the-wild generalization across complex scenarios, advancing toward practical, user-guided fashion creation.
