Mixed Reality (MR) headsets promise a future of immersive telepresence where virtual humans blend indistinguishably into their real or virtual surroundings. Achieving this vision requires a method capable of capturing a user's motion, estimating their appearance under novel lighting, and understanding the surrounding environment - all from the constrained viewpoint of a head-mounted display (HMD). Existing approaches treat these as isolated problems: they either focus on driving avatars with baked-in lighting or rely on complex studio setups for relighting. In this paper, we present EgoRelight, a holistic framework for egocentric telepresence that simultaneously captures full-body human performance, synthesizes photorealistic and relightable appearance, and estimates high dynamic range (HDR) environment maps from a single HMD. First, to ensure accurate motion and surface reconstruction, we propose an egocentric perception module that leverages stereo down-facing cameras to extract dense depth maps, which serve as robust geometric control signals to drive a mesh-based avatar. Second, we introduce a novel neural appearance model that learns to synthesize view-dependent specular and view-independent diffuse shading separately. By employing a specialized ray-sampling strategy, our model generalizes to unseen illumination without relying on restrictive analytical BRDF priors. Third, we enable seamless avatar integration into the physical world via a test-time inverse rendering process, which recovers an HDR environment map by matching the pre-trained avatar's appearance to the live egocentric camera observations. We demonstrate the efficiency of our system through a social telepresence application, where remote users are coherently relit according to their local physical environment. Extensive experiments show that our individual components and the integrated system significantly outperform the best combinations of state-of-the-art baselines in both geometric accuracy and rendering as well as relighting fidelity. Further details can be found on our project page1.
论文检索
输入标题、作者或关键词，从 700 篇学术成果中精准定位
Generating high-quality stereo videos requires consistent depth perception and temporal coherence across frames. Despite advances in image and video synthesis using diffusion models, producing high-quality stereo videos remains a challenging task due to the difficulty of maintaining consistent temporal and spatial coherence between left and right views. We introduce DissolveStereo, a novel framework for zero-shot stereo video generation that leverages video diffusion priors without requiring paired training data. Our key innovations include a noisy restart strategy to initialize stereo-aware latent representations and an iterative refinement process that progressively harmonizes the latent space, addressing issues like temporal flickering and view inconsistencies. Importantly, we propose the use of dissolved depth maps to streamline latent space operations by reducing high-frequency depth information. Our comprehensive evaluations, including quantitative metrics and user studies, demonstrate that DissolveStereo produces high-quality stereo videos with enhanced depth consistency and temporal smoothness. In terms of epipolar consistency, our method achieves an 11.7% improvement in MEt3R score over the current state-of-the-art. Furthermore, user studies indicate strong perceptual gains over the previous arts, with an 8.0% higher perceived frame quality and 10.9% higher perceived temporal coherence. Our code is in https://github.com/shijianjian/DissolveStereo.
Developments in display technology mean that it may soon be possible to create stereoscopic 3-D imagery that cannot be distinguished from its real-world counterpart, i.e., it could pass the so-called Visual Turing Test. To measure how far current display technologies are from this ultimate goal, we built a custom ultra-realistic display — a high dynamic range (HDR) stereoscope, with accurate color and geometric calibration. Then, we used the display to measure how the reduction of display resolution and dynamic range (contrast) affects the sense of realism, relative to a real scene. These two display capabilities have improved greatly in recent years, and it is useful to understand what benefits further development might bring. The display resolution measurements showed that the realism of presented scenes starts dropping rapidly below ∼ 60 pixels per degree, and the gains diminish above that resolution. Our contrast measurement showed a linear reduction in realism as the contrast was reduced. Furthermore, we found that the observers could discriminate between real and virtual scenes in less than half of the trials (when adjusted for the guess rate), demonstrating the ultra-realistic capabilities of our display. Our results provide a clear guide to potential gains in realism, and can inform display design and HDR tone mapping.
In human vision, the inputs perceived by each eye do not contribute equally to the final percept. Instead, visual dominance influences how these inputs are fused, giving more weight to one eye view over the other. While eye dominance has been traditionally treated as a static, eye-fixed property, recent evidence suggests that dominance can vary with viewing conditions. In this work, we systematically characterize dynamic visual dominance across the visual field, with a particular focus on peripheral vision, where perceptual asymmetries are most relevant for stereoscopic rendering. Through two complementary psychophysical experiments, we first show that tolerance to eye-asymmetric blur at the fovea under binocular viewing depends on gaze direction, confirming that eye dominance is not spatially invariant. We then show that peripheral dominance is primarily governed by retinal eccentricity, with consistent naso-temporal asymmetries and dominance reversals around the blind spots. We leverage our insights in a dominance-contingent rendering application, where additional blur is selectively applied to the perceptually non-dominant eye regions under binocular viewing. Compared to static dominance approaches, our method enables stronger localized quality reductions, illustrating the practical relevance of dynamic peripheral dominance for stereoscopic foveated rendering. Thus, our goal through this work is to show how visual dominance behaves dynamically in both the fovea and the periphery, indicating how foveation techniques could benefit from it.
We present SHARP-GS, a high-performance framework that unlocks real-time 8K rendering for immersive virtual experiences by overcoming the scaling bottlenecks of 3D Gaussian Splatting (3DGS). While 3DGS excels at real-time view synthesis, its performance degrades non-linearly at ultra-high resolutions due to excessive binning overhead, uncoalesced memory access, and redundant per-pixel arithmetic. SHARP-GS addresses these inefficiencies through three key contributions: (i) Resolution-Aware Adaptive Binning, which maintains a constant overlap factor via dynamic tile sizing and employs fine-grained sub-tile culling; (ii) a Morton-Ordered Memory Layout to ensure spatially coherent memory access; and (iii) Forward Differencing, which replaces expensive probability density function evaluations with efficient incremental updates. On high-end consumer GPUs, our framework achieves an average 2.55× speedup at 8K resolution, sustaining over 250 FPS with negligible visual quality loss. This performance delivers the necessary pixel throughput to enable high-fidelity, 120 Hz stereoscopic VR experiences. Code and data for this paper are at https://github.com/dingjr7/SHARP-GS.
Despite recent progress, text-to-image models still struggle to generate semantically diverse and compositionally accurate multi-person interaction scenes, often collapsing to repetitive layouts, stereotypical poses, and poorly grounded interactions. In this work, we bridge this gap by introducing a dual pose–image representation that brings person-centric structural priors into pretrained diffusion transformers. Our model jointly predicts a 2D pose visualization image and its corresponding RGB image, enabling structure and appearance to co-evolve during learning. At its core, a cross-modal alignment scheme binds text, pose, and image representations, ensuring consistent grounding across modalities. Furthermore, we design an iterative scene construction scheme, progressively generating complex multi-human interactions while effectively decomposing the overall generation complexity. Extensive experiments demonstrate that our method substantially improves prompt alignment and scene diversity in multi-person image generation.
The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language---making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.
Multimodal meme understanding is increasingly used to analyze socially sensitive content, yet existing models often exhibit biased behavior when interpreting economic dependence and social roles under ambiguity. Many memes express economic relationships through sparse text or symbolic visual cues, providing insufficient evidence for gendered attribution. In such underspecified settings, models tend to rely on pretraining correlations, leading to hallucinated and stereotypical economic role assignments. In this work, we study gendered economic dependence in image-text memes through the lens of contextual sufficiency and identify epistemic overcommitment—inferring roles without adequate evidence—as a primary source of bias. We propose CGER-Net, a context-grounded multimodal framework that estimates whether the input provides sufficient evidence for gendered economic reasoning and applies evidence-gated inference to enable confident attribution when cues are explicit while favoring principled abstention otherwise. We evaluate CGER-Net on EconMeme-GE, a curated dataset of image-text memes annotated as Men, Women, Neutral, or Ambiguous. Across strong contemporary multimodal baselines, CGER-Net reduces Gender Overcommitment Rate by up to 44% on ambiguous instances while maintaining comparable accuracy on unambiguous cases. Human evaluation further shows that 79% of generated rationales are judged as epistemically aligned with the available evidence. These results highlight the importance of modeling when not to infer for reliable and responsible multimodal analysis.
The Spherical Sliced Wasserstein (SSW) distance has emerged as a fundamental tool for comparing probability measures on the hypersphere, with broad applications spanning geology, medical imaging, computer vision, and representation learning. However, its reliance on uniformly sampling a large number of projection directions from the unit hypersphere can lead to suboptimal performance, as many directions are weakly informative and fail to capture salient differences between distributions. We address this limitation through Fusion Stereographic Spherical Sliced Wasserstein (FS3W), a data-adaptive divergence that incorporates distribution-dependent information into projection selection. FS3W uses a neural slicing fusion mechanism to refine prior directions, steering them toward projections aligned with discriminative geometric structures in the data. This adaptive approach produces more expressive and accurate characterizations of distributional discrepancies on the hypersphere. We evaluate FS3W across five diverse tasks, including gradient flows, Earth density estimation, and self-supervised representation learning. Empirical results consistently demonstrate that FS3W outperforms existing spherical optimal transport–based methods.
PHOTONS: Pose-Free Human-Centric Photo-Realistic Real-Time Novel View Synthesis from Sparse Views
PDF ↗We present PHOTONS (Pose-Free Human-Centric Photo-Realistic Real-Time Novel View Synthesis from Sparse Views), a real-time framework for novel view synthesis without requiring camera calibration. Our method reconstructs consistent 3D Gaussian point clouds and synthesizes 2K photo-realistic novel views from arbitrary numbers (>=2) of freely placed cameras. PHOTONS faithfully renders dynamic human bodies amid complex backgrounds, including interactive object manipulation and fine-grained details (e.g., hair strands), while maintaining 25 FPS throughput on commodity GPU like NVIDIA RTX 4090. By combining pose-free spatial point cloud reconstruction with Gaussian parameter estimation, our method demonstrates strong resilience to occlusions and camera perturbations. Additionally, we develop a 3D stereo system that drastically reduces setup complexity compared to existing solutions. Experiments on public and custom datasets show that PHOTONS outperforms state-of-the-art methods in both efficiency and visual quality.
Large language models (LLMs) have achieved remarkable success in many domains, but concerns about data quality and privacy are growing. Federated Learning (FL) offers a privacy-preserving solution by training a model on local clients without sharing data. However, the impact of biased private data on LLMs fine-tuned through FL remains understudied. This work investigates how client-side biased data affects the global model during federated fine-tuning of LLMs. We simulate realistic scenarios where some clients possess datasets containing social biases (stereotypes, discriminatory language) while others have clean data through extensive experiments with popular FL algorithms (FedAvg, FedAdam and FedProx) and popular LLMs (LLaMA, Mistral, Phi-3 and Gemma) across datasets with varying bias proportions (33%, 66%, 100%). Our findings reveal that 1) FedAdam consistently shows the lowest bias propagation, reducing CrowS-Pairs scores by up to 15% compared to FedAvg; 2) Even small amounts of biased data (33%) can significantly influence global model bias; 3) Mixed biased and neutral data distributions lead to 5%-7% higher bias scores than segregated distributions. Additionally, we propose Bias-Aware Model Aggregation (BAMA), a novel debiasing method for federated fine-tuning that consistently reduces bias across various models and algorithms.
Image classification systems often inherit biases from uneven group representation in training data. For example, in face datasets for hair color classification, blond hair may be disproportionately associated with females, reinforcing stereotypes. A recent approach leverages the Stable Diffusion model to generate balanced training data, but these models often struggle to preserve the original data distribution. In this work, we explore multiple diffusion-finetuning techniques, e.g., LoRA and DreamBooth, to generate images that more accurately represent each training group by learning directly from their samples. Additionally, in order to prevent a single DreamBooth model from being overwhelmed by excessive intra-group variations, we explore a technique of clustering images within each group and train a DreamBooth model per cluster. These models are then used to generate group-balanced data for pretraining, followed by fine-tuning on real data. Experiments on multiple benchmarks demonstrate that the studied finetuning approaches outperform vanilla Stable Diffusion on average and achieve results comparable to SOTA debiasing techniques like Group-DRO, while surpassing them as the dataset bias severity increases.
As LLMs (large language models) are increasingly used to generate synthetic personas, particularly in data-limited domains such as health, privacy, and HCI, it becomes necessary to understand how these narratives represent identity, especially that of minority communities. In this paper, we audit synthetic personas generated by 3 LLMs (GPT4o, Gemini 1.5 Pro, Deepseek v2.5) through the lens of representational harm, focusing specifically on racial identity. Using a mixed-methods approach combining close reading, lexical analysis, and a parameterized creativity framework, we compare 1,512 LLM-generated persona to human-authored responses. Our findings reveal that LLMs disproportionately foreground racial markers, overproduce culturally coded language, and construct personas that are syntactically elaborate yet narratively reductive. These patterns result in a range of sociotechnical harms, including stereotyping, exoticism, erasure, and benevolent bias, that are often obfuscated by superficially positive narrations. We formalize this phenomenon as algorithmic othering, where minoritized identities are rendered hypervisible but less authentic.
Operationalizing Pluralistic Values in Large Language Model Alignment Reveals Trade-offs in Safety, Inclusivity, and Model Behavior
PDF ↗Although large language models (LLMs) are increasingly trained using human feedback for safety and alignment with human values, alignment decisions often overlook human social diversity. This study examines how incorporating pluralistic values affects LLM behavior by systematically evaluating demographic variation and design parameters in the alignment pipeline. We collect alignment data from US and German participants (N = 1,095 participants, 27,375 ratings) who rated LLM responses across five dimensions: Toxicity, Emotional Awareness (EA), Sensitivity, Stereotypical Bias, and Helpfulness. We fine-tuned multiple Large Language Models and Large Reasoning Models using preferences from different social groups while varying rating scales, disagreement handling methods, and optimization techniques. The results revealed systematic demographic effects: male participants rated responses 18% less toxic than female participants; conservative and Black participants rated responses 27.9% and 44% higher on EA than liberal and White participants, respectively. Models fine-tuned on group-specific preferences exhibited distinct behaviors. Technical design choices showed strong effects: the preservation of rater disagreement achieved roughly 53% greater toxicity reduction than majority voting, and 5-point scales yielded about 22% more reduction than binary formats; and Direct Preference Optimization (DPO) consistently outperformed Group Relative Policy Optimization (GRPO) in multi-value optimization. These findings represent a preliminary step in answering a critical question: How should alignment balance expert-driven and user-driven signals to ensure both safety and fair representation?
ShadeEdit: A Utility-Preserving and Defense-Evasive Knowledge Manipulation Attack in Federated LLMs
PDF ↗Recent studies reveal that adversaries can manipulate the internal knowledge of large language models (LLMs) on selected topics through model editing, causing attacker-specified harmful or biased outputs when queried about the edited content. Once such tampered LLMs are distributed, they can mislead users on the targeted topics, thereby potentially propagating misinformation or reinforcing stereotypes. However, existing knowledge manipulation attacks rely on the ability to redistribute compromised models, which is infeasible in constrained settings like Federated Instruction Tuning (FedIT), where a central server controls LLM's training and distribution. In this work, we introduce ShadeEdit, the first attack framework that leverages strengthened model editing to enable knowledge manipulation in FedIT scenarios. ShadeEdit introduces two key components to address two challenges posed by the training process of FedIT: (1) a paraphrase-based editing dataset selection strategy to mitigate the dilution from benign updates on malicious ones by constructing a high-quality editing dataset, and (2) an adaptive manipulation mechanism to evade aggregation-based defenses via an adaptive clipping strategy. ShadeEdit achieves an average 99.5% attack success rate over eight robust aggregation algorithms while preserving instruction-following accuracy, demonstrating its strong attack effectiveness and model-utility preservation.
Existing stereotype auditing methods for Large Language Models (LLMs) typically rely on isolated rating schemes or task-specific probes, lacking theoretical grounding and failing to reveal internal organization beyond surface-level output patterns. In this paper, we introduce SCoUT (Stereotype Content-oriented Utility structure via Thurstonian modeling), a closed-loop framework that structurally models, explicitly probes, and functionally steers stereotype dimensions (warmth and competence) in LLMs. SCoUT first reconstructs a global stereotype utility structure aligned with Stereotype Content Model theory via Thurstonian comparative judgments. Across multiple open-source LLMs, this modeling achieves high pairwise-preference prediction accuracy (≥ 0.90 on larger-scale models) and exhibits strong cross-model consistency. Probing internal attention mechanisms localizes this structure to specific heads (Spearman’s ρ up to 0.83 for warmth and 0.90 for competence) and surfaces a salient asymmetry between warmth and competence. Further, targeted inference-time activation modifications on these dimension-sensitive heads consistently steer model outputs along the intended axes. By bridging behavioral measurement with internal representation and controllable steering, SCoUT offers an end-to-end framework that uncovers and interprets the latent structure of stereotypes, advancing stereotype auditing from surface detection to structural analysis.
Libra-MIL: Multimodal Prototypes Stereoscopic Infused with Task-specific Language Priors for Few-shot Whole Slide Image Classification
PDF ↗While Large Language Models (LLMs) are emerging as a promising direction in computational pathology, the substantial computational cost of giga-pixel Whole Slide Images (WSIs) necessitates the use of Multi-Instance Learning (MIL) to enable effective modeling. A key challenge is that pathological tasks typically provide only bag-level labels, while instance-level descriptions generated by LLMs often suffer from bias due to a lack of fine-grained medical knowledge. To address this, we propose that constructing task-specific pathological entity prototypes is crucial for learning generalizable features and enhancing model interpretability. Furthermore, existing vision-language MIL methods often employ unidirectional guidance, limiting cross-modal synergy. In this paper, we introduce a novel approach, Multimodal Prototype-based Multi-Instance Learning, that promotes bidirectional interaction through a balanced information compression scheme. Specifically, we leverage a frozen LLM to generate task-specific pathological entity descriptions, which are learned as text prototypes. Concurrently, the vision branch learns instance-level prototypes to mitigate the model's reliance on redundant data. For the fusion stage, we employ the Stereoscopic Optimal Transport (SOT) algorithm, which is based on a similarity metric, thereby facilitating broader semantic alignment in a higher-dimensional space. We conduct few-shot classification and explainability experiments on three distinct cancer datasets, and the results demonstrate the superior generalization capabilities of our proposed method.
Splat-SAP: Feed-Forward Gaussian Splatting for Human-Centered Scene with Scale-Aware Point Map Reconstruction
PDF ↗We present Splat-SAP, a feed-forward approach to render novel views of human-centered scenes from binocular cameras with large sparsity. Gaussian Splatting has shown its promising potential in rendering tasks, but it typically necessitates per-scene optimization with dense input views. Although some recent approaches achieve feed-forward Gaussian Splatting rendering through geometry priors obtained by multi-view stereo, such approaches still require largely overlapped input views to establish the geometry prior. To bridge this gap, we leverage pixel-wise point map reconstruction to represent geometry which is robust to large sparsity for its independent view modeling. In general, we propose a two-stage learning strategy. In stage 1, we transform the point map into real space via an iterative affinity learning process, which facilitates camera control in the following. In stage 2, we project point maps of two input views onto the target view plane and refine such geometry via stereo matching. Furthermore, we anchor Gaussian primitives on this refined plane in order to render high-quality images. As a metric representation, the scale-aware point map in stage 1 is trained in a self-supervised manner without 3D supervision and stage 2 is supervised with photo-metric loss. We collect multi-view human-centered data and demonstrate that our method improves both the stability of point map reconstruction and the visual quality of free-viewpoint rendering.
Cheating Stereo Matching in Full-Scale: Physical Adversarial Attack Against Binocular Depth Estimation in Autonomous Driving
PDF ↗Though deep neural models adopted to realize the perception of autonomous driving have proven vulnerable to adversarial examples, known attacks often leverage 2D patches and target mostly monocular perception. Therefore, the effectiveness of Physical Adversarial Examples (PAEs) on stereo-based binocular depth estimation remains largely unexplored. To this end, we propose the first texture-enabled physical adversarial attack against stereo matching models in the context of autonomous driving. Our method employs a 3D PAE with global camouflage texture rather than a local 2D patch-based one, ensuring both visual consistency and attack effectiveness across different viewpoints of stereo cameras. To cope with the disparity effect of these cameras, we also propose a new 3D stereo matching rendering module that allows the PAE to be aligned with real-world positions and headings in binocular vision. We further propose a novel merging attack that seamlessly blends the target into the environment through fine-grained PAE optimization. It has significantly enhanced stealth and lethality upon existing hiding attacks that fail to get seamlessly merged into the background. Extensive evaluations show that our PAEs can successfully fool the stereo models into producing erroneous depth information.
MolSight: Optical Chemical Structure Recognition with SMILES Pretraining, Multi-Granularity Learning and Reinforcement Learning
PDF ↗Optical Chemical Structure Recognition (OCSR) plays a pivotal role in modern chemical informatics, enabling the automated conversion of chemical structure images from scientific literature, patents, and educational materials into machine-readable molecular representations. This capability is essential for large-scale chemical data mining, drug discovery pipelines, and Large Language Model (LLM) applications in related domains. However, existing OCSR systems face significant challenges in accurately recognizing stereochemical information due to the subtle visual cues that distinguish stereoisomers, such as wedge and dash bonds, ring conformations, and spatial arrangements. To address these challenges, we propose MolSight, a comprehensive learning framework for OCSR that employs a three-stage training paradigm. In the first stage, we conduct pre-training on large-scale but noisy datasets to endow the model with fundamental perception capabilities for chemical structure images. In the second stage, we perform multi-granularity fine-tuning using datasets with richer supervisory signals, systematically exploring how auxiliary tasks—specifically chemical bond classification and atom localization—contribute to molecular formula recognition. Finally, we employ reinforcement learning for post-training optimization and introduce a novel stereochemical structure dataset. Remarkably, we find that even with MolSight's relatively compact parameter size, the Group Relative Policy Optimization (GRPO) algorithm can further enhance the model's performance on stereomolecular. Through extensive experiments across diverse datasets, our results demonstrate that MolSight achieves state-of-the-art performance in (stereo)chemical optical structure recognition.