ConsistentRFT: Reducing Visual Hallucinations in Flow-based Reinforcement Fine-Tuning

Reinforcement Fine-Tuning (RFT) on flow-based models is crucial for preference alignment. However, they often introduce visual hallucinations like over-optimized details and semantic misalignment. This work preliminarily explores why visual hallucinations arise and how to reduce them. We first investigate RFT methods from a unified perspective, and reveal the core problems stemming from two aspects, exploration and exploitation: (1) limited exploration during stochastic differential equation (SDE) rollouts, leading to an over-emphasis on local details at the expense of global semantics, and (2) trajectory imitation process inherent in policy gradient methods, distorting the model's foundational vector field and its cross-step consistency. Building on this, we propose ConsistentRFT, a general framework to mitigate these hallucinations. Specifically, we design a Dynamic Granularity Rollout (DGR) mechanism to balance exploration between global semantics and local details by dynamically scheduling different noise sources. We then introduce a Consistent Policy Gradient Optimization (CPGO) that preserves the model's consistency by aligning the current policy with a more stable prior. Extensive experiments demonstrate that ConsistentRFT significantly mitigates visual hallucinations, achieving average reductions of 49\% for low-level and 38\% for high-level perceptual hallucinations. Furthermore, ConsistentRFT outperforms other RFT methods on out-of-domain metrics, showing an improvement of 5.1\% (v.s. the baseline's decrease of -0.4\%) over FLUX1.dev. This is \href{https://xiaofeng-tan.github.io/projects/ConsistentRFT}{Project Page}.

Fast Inference of Visual Autoregressive Model with Adjacency-Adaptive Dynamical Draft Trees

Autoregressive (AR) image models achieve diffusion-level quality but suffer from sequential inference, requiring approximately 2,000 steps for a 576x576 image. Speculative decoding with draft trees accelerates LLMs yet underperforms on visual AR models due to spatially varying token prediction difficulty. We identify a key obstacle in applying speculative decoding to visual AR models: inconsistent acceptance rates across draft trees due to varying prediction difficulties in different image regions. We propose Adjacency-Adaptive Dynamical Draft Trees (ADT-Tree), an adjacency-adaptive dynamic draft tree that dynamically adjusts draft tree depth and width by leveraging adjacent token states and prior acceptance rates. ADT-Tree initializes via horizontal adjacency, then refines depth/width via bisectional adaptation, yielding deeper trees in simple regions and wider trees in complex ones. The empirical evaluations on MS-COCO 2017 and PartiPrompts demonstrate that ADT-Tree achieves speedups of 3.13xand 3.05x, respectively. Moreover, it integrates seamlessly with relaxed sampling methods such as LANTERN, enabling further acceleration. Code is available at https://github.com/Haodong-Lei-Ray/ADT-Tree.

Multimodal Skeleton-Based Action Representation Learning via Decomposition and Composition

Multimodal human action understanding is a significant problem in computer vision, with the central challenge being the effective utilization of the complementarity among diverse modalities while maintaining model efficiency. However, most existing methods rely on simple late fusion to enhance performance, which results in substantial computational overhead. Although early fusion with a shared backbone for all modalities is efficient, it struggles to achieve excellent performance. To address the dilemma of balancing efficiency and effectiveness, we introduce a self-supervised multimodal skeleton-based action representation learning framework, named Decomposition and Composition. The Decomposition strategy meticulously decomposes the fused multimodal features into distinct unimodal features, subsequently aligning them with their respective ground truth unimodal counterparts. On the other hand, the Composition strategy integrates multiple unimodal features, leveraging them as self-supervised guidance to enhance the learning of multimodal representations. Extensive experiments on the NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD II datasets demonstrate that the proposed method strikes an excellent balance between computational cost and model performance.

LOTA: Bit-Planes Guided AI-Generated Image Detection

The rapid advancement of GAN and Diffusion models makes it more difficult to distinguish AI-generated images from real ones. Recent studies often use image-based reconstruction errors as an important feature for determining whether an image is AI-generated. However, these approaches typically incur high computational costs and also fail to capture intrinsic noisy features present in the raw images. To solve these problems, we innovatively refine error extraction by using bit-plane-based image processing, as lower bit planes indeed represent noise patterns in images. We introduce an effective bit-planes guided noisy image generation and exploit various image normalization strategies, including scaling and thresholding. Then, to amplify the noise signal for easier AI-generated image detection, we design a maximum gradient patch selection that applies multi-directional gradients to compute the noise score and selects the region with the highest score. Finally, we propose a lightweight and effective classification head and explore two different structures: noise-based classifier and noise-guided classifier. Extensive experiments on the GenImage benchmark demonstrate the outstanding performance of our method, which achieves an average accuracy of \textbf{98.9\%} (\textbf{11.9}\%~$\uparrow$) and shows excellent cross-generator generalization capability. Particularly, our method achieves an accuracy of over 98.2\% from GAN to Diffusion and over 99.2\% from Diffusion to GAN. Moreover, it performs error extraction at the millisecond level, nearly a hundred times faster than existing methods. The code is at https://github.com/hongsong-wang/LOTA.

Foundation Model for Skeleton-Based Human Action Understanding

Human action understanding serves as a foundational pillar in the field of intelligent motion perception. Skeletons serve as a modality- and device-agnostic representation for human modeling, and skeleton-based action understanding has potential applications in humanoid robot control and interaction. \RED{However, existing works often lack the scalability and generalization required to handle diverse action understanding tasks. There is no skeleton foundation model that can be adapted to a wide range of action understanding tasks}. This paper presents a Unified Skeleton-based Dense Representation Learning (USDRL) framework, which serves as a foundational model for skeleton-based human action understanding. USDRL consists of a Transformer-based Dense Spatio-Temporal Encoder (DSTE), Multi-Grained Feature Decorrelation (MG-FD), and Multi-Perspective Consistency Training (MPCT). The DSTE module adopts two parallel streams to learn temporal dynamic and spatial structure features. The MG-FD module collaboratively performs feature decorrelation across temporal, spatial, and instance domains to reduce dimensional redundancy and enhance information extraction. The MPCT module employs both multi-view and multi-modal self-supervised consistency training. The former enhances the learning of high-level semantics and mitigates the impact of low-level discrepancies, while the latter effectively facilitates the learning of informative multimodal features. We perform extensive experiments on 25 benchmarks across across 9 skeleton-based action understanding tasks, covering coarse prediction, dense prediction, and transferred prediction. Our approach significantly outperforms the current state-of-the-art methods. We hope that this work would broaden the scope of research in skeleton-based action understanding and encourage more attention to dense prediction tasks.

DreamCS: Geometry-Aware Text-to-3D Generation with Unpaired 3D Reward Supervision

While text-to-3D generation has attracted growing interest, existing methods often struggle to produce 3D assets that align well with human preferences. Current preference alignment techniques for 3D content typically rely on hardly-collected preference-paired multi-view 2D images to train 2D reward models, when then guide 3D generation -- leading to geometric artifacts due to their inherent 2D bias. To address these limitations, we construct 3D-MeshPref, the first large-scale unpaired 3D preference dataset, featuring diverse 3D meshes annotated by a large language model and refined by human evaluators. We then develop RewardCS, the first reward model trained directly on unpaired 3D-MeshPref data using a novel Cauchy-Schwarz divergence objective, enabling effective learning of human-aligned 3D geometric preferences without requiring paired comparisons. Building on this, we propose DreamCS, a unified framework that integrates RewardCS into text-to-3D pipelines -- enhancing both implicit and explicit 3D generation with human preference feedback. Extensive experiments show DreamCS outperforms prior methods, producing 3D assets that are both geometrically faithful and human-preferred. Code and models will be released publicly.

PAMD: Plausibility-Aware Motion Diffusion Model for Long Dance Generation

Computational dance generation is crucial in many areas, such as art, human-computer interaction, virtual reality, and digital entertainment, particularly for generating coherent and expressive long dance sequences. Diffusion-based music-to-dance generation has made significant progress, yet existing methods still struggle to produce physically plausible motions. To address this, we propose Plausibility-Aware Motion Diffusion (PAMD), a framework for generating dances that are both musically aligned and physically realistic. The core of PAMD lies in the Plausible Motion Constraint (PMC), which leverages Neural Distance Fields (NDFs) to model the actual pose manifold and guide generated motions toward a physically valid pose manifold. To provide more effective guidance during generation, we incorporate Prior Motion Guidance (PMG), which uses standing poses as auxiliary conditions alongside music features. To further enhance realism for complex movements, we introduce the Motion Refinement with Foot-ground Contact (MRFC) module, which addresses foot-skating artifacts by bridging the gap between the optimization objective in linear joint position space and the data representation in nonlinear rotation space. Extensive experiments show that PAMD significantly improves musical alignment and enhances the physical plausibility of generated motions. This project page is available at: https://mucunzhuzhu.github.io/PAMD-page/.

ReAlign: Bilingual Text-to-Motion Generation via Step-Aware Reward-Guided Alignment

Bilingual text-to-motion generation, which synthesizes 3D human motions from bilingual text inputs, holds immense potential for cross-linguistic applications in gaming, film, and robotics. However, this task faces critical challenges: the absence of bilingual motion-language datasets and the misalignment between text and motion distributions in diffusion models, leading to semantically inconsistent or low-quality motions. To address these challenges, we propose BiHumanML3D, a novel bilingual human motion dataset, which establishes a crucial benchmark for bilingual text-to-motion generation models. Furthermore, we propose a Bilingual Motion Diffusion model (BiMD), which leverages cross-lingual aligned representations to capture semantics, thereby achieving a unified bilingual model. Building upon this, we propose Reward-guided sampling Alignment (ReAlign) method, comprising a step-aware reward model to assess alignment quality during sampling and a reward-guided strategy that directs the diffusion process toward an optimally aligned distribution. This reward model integrates step-aware tokens and combines a text-aligned module for semantic consistency and a motion-aligned module for realism, refining noisy motions at each timestep to balance probability density and alignment. Experiments demonstrate that our approach significantly improves text-motion alignment and motion quality compared to existing state-of-the-art methods. Project page: https://wengwanjiang.github.io/ReAlign-page/.

Training-Free Zero-Shot Temporal Action Detection with Vision-Language Models

Existing zero-shot temporal action detection (ZSTAD) methods predominantly use fully supervised or unsupervised strategies to recognize unseen activities. However, these training-based methods are prone to domain shifts and require high computational costs, which hinder their practical applicability in real-world scenarios. In this paper, unlike previous works, we propose a training-Free Zero-shot temporal Action Detection (FreeZAD) method, leveraging existing vision-language (ViL) models to directly classify and localize unseen activities within untrimmed videos without any additional fine-tuning or adaptation. We mitigate the need for explicit temporal modeling and reliance on pseudo-label quality by designing the LOGarithmic decay weighted Outer-Inner-Contrastive Score (LogOIC) and frequency-based Actionness Calibration. Furthermore, we introduce a test-time adaptation (TTA) strategy using Prototype-Centric Sampling (PCS) to expand FreeZAD, enabling ViL models to adapt more effectively for ZSTAD. Extensive experiments on the THUMOS14 and ActivityNet-1.3 datasets demonstrate that our training-free method outperforms state-of-the-art unsupervised methods while requiring only 1/13 of the runtime. When equipped with TTA, the enhanced method further narrows the gap with fully supervised methods.

STHFL: Spatio-Temporal Heterogeneous Federated Learning

Federated learning is a new framework that protects data privacy and allows multiple devices to cooperate in training machine learning models. Previous studies have proposed multiple approaches to eliminate the challenges posed by non-iid data and inter-domain heterogeneity issues. However, they ignore the \textbf{spatio-temporal} heterogeneity formed by different data distributions of increasing task data in the intra-domain. Moreover, the global data is generally a long-tailed distribution rather than assuming the global data is balanced in practical applications. To tackle the \textbf{spatio-temporal} dilemma, we propose a novel setting named \textbf{Spatio-Temporal Heterogeneity} Federated Learning (STHFL). Specially, the Global-Local Dynamic Prototype (GLDP) framework is designed for STHFL. In GLDP, the model in each client contains personalized layers which can dynamically adapt to different data distributions. For long-tailed data distribution, global prototypes are served as complementary knowledge for the training on classes with few samples in clients without leaking privacy. As tasks increase in clients, the knowledge of local prototypes generated in previous tasks guides for training in the current task to solve catastrophic forgetting. Meanwhile, the global-local prototypes are updated through the moving average method after training local prototypes in clients. Finally, we evaluate the effectiveness of GLDP, which achieves remarkable results compared to state-of-the-art methods in STHFL scenarios.