arXiv subjects

Weidong Cai

Publications and source records attributed to Weidong Cai.

At least 37 records · Page 2Linked to original sources

arXiv2026

Calibrate What You SHIP: Post-Selection Risk Control for Verifier-Guided Text-to-Image Generation

Verifier-guided text-to-image systems increasingly use test-time search to select, refine, or stop among multiple candidates, yet release thresholds are often calibrated on individual images. This creates a candidate-to-policy calibration mismatch: search changes both which prompts receive an output and which candidate is released, so candidate-level risk control need not imply control of released-output risk. We formalize this estimand shift through prompt reweighting and within-prompt selection, and introduce SHIP, Selection-aware Held-out calibration of Inference Policies. SHIP runs or replays the complete deployed policy on held-out prompts, evaluates the image it actually releases using an independent target judge, and selects the most permissive threshold whose risk upper bound satisfies a prescribed budget. For replayable policies with a prespecified threshold grid, simultaneous confidence control provides finite-sample validity. Experiments across fixed, sequential, and adaptive T2I inference procedures show that policy-level calibration recovers lower-risk operating points while exposing policy-dependent tradeoffs among risk, coverage, and compute. On GenEval2 with FLUX at N=16, a pooled-candidate threshold yields released risk 0.310, whereas SHIP reduces it to 0.162. Across 200 cached-stream splits, the fixed-grid certificate has no target crossing. Reliable inference-time scaling therefore requires calibrating the output distribution induced by the complete deployed policy.

cs.CV↗

arXiv2026

DefaultShift: Auditing Semantic Default Shift in Accelerated Text-to-Image Models

Few-step text-to-image models increasingly replace slower generators, yet acceleration can silently change distributions over unspecified attributes even when individual outputs remain plausible and aligned. We call these distributions semantic defaults and their change under replacement semantic default shift. Existing quality, preference, and diversity evaluations do not test whether a replacement preserves its reference model's semantic defaults. We introduce DefaultShift, a paired audit that labels repeated samples with closed semantic vocabularies, measures probability-mass movement, and separates interpretable ranking from confirmatory cross-fit inference. Across 14 reference and replacement pairs, adjusted color discrepancies range from 0.054 to 0.303 with recipe-specific directions. A 1,000-image human audit reproduces the ordering. We further introduce DefaultShift-Select, an offline calibration method that reduces human-measured shift by 10.3 percent to 35.1 percent across Turbo, DMD2, and FLUX without material quality loss. Under balanced evaluation, selected data recover 4.3 accuracy points and 7.5 worst-group points over uncalibrated replacement data. DefaultShift makes semantic preservation under acceleration measurable and actionable.

cs.CV↗

arXiv2026

Towards Conditional Feature Alignment for Cross-Domain Counting

Object counting models often degrade under cross-domain deployment because density composition varies across domains and is itself task-relevant. Standard feature alignment methods tend to suppress such variation by encouraging global domain invariance, which can be harmful when source and target domains contain different proportions of background, sparse foreground, and dense foreground. We propose Conditional Feature Alignment (CFA), a cross-domain counting framework that aligns representations within label-induced conditions rather than across full marginal feature distributions. Given density annotations or pseudo-density predictions, CFA constructs foreground/background or density-level conditions and aligns only features belonging to matching conditions. We formalise this idea through a conditional divergence perspective, characterising an ideal conditionally aligned state with no within-condition discrepancy while preserving condition-marginal density shift. For unsupervised domain adaptation, CFA estimates source conditions from annotations and target conditions from detached pseudo-density maps, then performs condition-wise adversarial alignment with full-image consistency regularisation. For source-domain generalisation, we instantiate the same principle with MPCount by enforcing condition-wise memory-consistency between generated source-domain views. Experiments on crowd and cell counting benchmarks show competitive or improved performance across diverse UDA and DG settings. For example, on JHU-CROWD++ FH->SN, CFA-DG reduces MAE/RMSE from MPCount's 216.3/421.4 to 90.5/169.9, showing a marked improvement on this large weather- and density-induced shift. These results suggest that condition-wise alignment is a promising design principle for domain-adaptive counting.

cs.CV↗

arXiv2026

Reflex First, Reflect Later: Latency-Aware Embodied LLM Agents for Dynamic Response

Large language models (LLMs) have substantially improved the planning capabilities of embodied agents, enabling their deployment in dynamic and safety-critical environments. However, these settings expose a critical limitation: inference latency. Delayed LLM responses can weaken real-time responsiveness and misalign agent reasoning with rapidly changing environmental states. This paper systematically studies the impact of inference latency on LLM-based embodied agents in dynamic environments. We introduce an FPS-based Time Conversion Mechanism (TCM) that maps inference time to elapsed simulation time, allowing computational delays to directly affect environmental evolution and agent outcomes. We instantiate this protocol in HAZARD and introduce Response Latency (RL) and Latency-to-Action Ratio (LAR) to evaluate agent responsiveness. Building on this framework, we propose the Rapid-Reflex Async-Reflect Agent (RRARA), which integrates rapid reflexive actions with asynchronous LLM reflection to mitigate latency-induced errors. We further introduce an LLM-based PrePlanner that generates cached object-centric subgoals, reducing repeated LLM calls while retaining the model's high-level reasoning capability. Experiments show that accounting for inference latency substantially changes embodied-agent performance and that RRARA achieves a stronger balance between decision quality and responsiveness.

cs.AI↗

arXiv2026

When Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models

Safety guards are widely used to filter harmful content and are typically trained via supervised fine-tuning on labeled prompt-response pairs. We audit two widely used safety-guard training datasets, WildGuardMix and GR-Train, and find that among responses to harmful prompts, refusal expressions co-occur almost exclusively with unharmful labels. This imbalance motivates what we term the refusal-cue shortcut: inserting a refusal cue into a harmful response could flip the guard's verdict from harmful to unharmful. The shortcut affects not only guards trained on these datasets but also officially released models such as LlamaGuard3 and Qwen3Guard whose training data is undisclosed. It persists across response positions and is generally stronger in smaller variants within a family. To mitigate it, we adapt sparse complementary masking as a lightweight post-hoc intervention that identifies and suppresses a small set of shortcut-associated attention heads and MLP neurons without retraining. On two primary benchmarks, the intervention achieves an approximately 79% relative reduction in response-initial detection failures induced by refusal cues, while preserving standard detection performance. Although optimized using cues at a single response position, the suppression effect transfers to unseen positions and datasets, suggesting that shortcut manifestations across positions are partly mediated by shared internal components. Further analysis provides evidence that shortcut reliance and legitimate refusal recognition are partially functionally separable, as suppressing the shortcut broadly preserves the guard's ability to recognize genuine refusals.

cs.AI↗

arXiv2026

MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection

Small-object detection in Unmanned Aerial Vehicle (UAV) imagery requires preserving weak local evidence while using broader context to separate tiny foreground targets from cluttered backgrounds. Existing multi-scale fusion methods improve feature aggregation, but they often add computation or blur fine details during repeated cross-scale fusion. The central challenge is to balance low-SNR target preservation, clutter suppression, and efficient cross-scale context exchange. To address this challenge, we propose the Multi-scale Global-detail Feature Integration Strategy (MGDFIS), a neck-level feature-fusion strategy that couples global context exchange, local-detail recovery, and pixel-level foreground-background recalibration. MGDFIS integrates three coordinated modules: FusionLock-TSS Attention for stabilizing spectral-spatial responses, Global-detail Integration for combining long-range mixing with local detail capture, and Dynamic Pixel Attention for reweighting compact foreground regions. On the controlled VisDrone setting, YOLO26m + MGDFIS improves AP50:95 from 25.7 to 30.2 and AP50 from 37.2 to 44.2 over the YOLO26m baseline, with 96.1 GFLOPs. Additional dataset-specific evaluations report 38.9 AP50 and 21.9 AP50:95 on UAVDT and 97.4 AP50 on CARPK. The code is available at: https://github.com/Bai-Xuecheng/MGDFIS.

cs.CV↗

arXiv2026

Learning Color Grading, No Photo Sharing: Federated Aesthetic Preference Learning for Personalized Image Enhancement

Personalized image enhancement should reflect individual aesthetic taste, yet learning such preferences commonly depends on private photos and ratings that are unsuitable for centralized collection. The task must infer preference from sparse, heterogeneous feedback and translate it into natural-looking color transformations on resource-constrained user devices. We introduce FedPAIE, a federated personalized aesthetic image enhancement framework for user-adaptive color grading without centralizing raw photos or ratings. FedPAIE trains a lightweight dual-cue aesthetic scorer, calibrates it into a personalized scorer on a small local support set, and freezes it to guide regularized adaptation of a lightweight CLUT enhancer from unpaired local photographs. Fidelity constraints and an excess-gap penalty regularize scorer-guided adaptation to limit proxy-score over-optimization while preserving content and natural appearance. Training remains lightweight throughout the pipeline: scorer learning updates at most 0.787M parameters, enhancer adaptation updates 0.265M, and inference retains only a 0.293M-parameter personalized enhancer. Experiments on MIT-Adobe FiveK and Flickr-AES demonstrate effective open-world personalization and a favorable balance between user preference and image fidelity. FedPAIE thus connects decentralized preference learning with efficient personalized image transformation without requiring paired user retouches.

cs.CV↗

arXiv2026

MusicWeaver: Programmable Long-Form Music Generation with Provably Local Editing

Music generation systems produce increasingly realistic audio, yet they expose no interface between a creator's structural intent and the rendered sound. Form can only be steered through prompts, while local revisions require regenerating entire pieces. We recast music creation as program-guided generation, using an explicit, human-interpretable program layer between intent and audio. We present MusicWeaver, which decomposes generation into planning and rendering. The planning stage predicts a structured plan, a multi-level song program encoding musical form, motif recurrence, and bar-level attributes. The rendering stage synthesizes audio conditioned on this plan. We formalize editing as an algebra of typed plan operations, including section replacement, insertion, deletion, motif retagging, and attribute changes, and prove that they preserve plan validity. To realize edits, we propose Projected Diffusion Inpainting, which guarantees that audio outside the edited span is preserved exactly across repeated revisions. For rendering, we design a Global-Local Diffusion Transformer with Motif Memory Retrieval that executes minute-scale plans and produces section returns that are consistent yet varied. Our framework also supports plan induction, recovering an editable program from an existing recording, and a natural-language editor that compiles free-form instructions into validated operation sequences. We introduce plan-faithfulness and edit-fidelity measures and validate them against human judgments. Experiments on generation conditioned on text, video, and their combination show state-of-the-art structural coherence and editability, while edit quality remains stable over sequential revisions where prior methods degrade.

cs.SD↗

arXiv2026

SciFigAlign: Scoring Scientific Figures by Fine-tuned Alignment of Visuals with Manuscript Evidence

Scientific figure assessment in peer review differs fundamentally from general image quality evaluation: a figure must be visually legible, faithfully support the manuscript's claims, and communicate evidence with a clear visual hierarchy. However, if we apply traditional image assessment methods to scientific figure quality assessment, limitations emerge: classic IQA models capture perceptual quality or aesthetics but cannot judge whether a figure serves the paper's scientific argument; CLIP-based methods assess generic image-text correspondence, yet lack understanding of manuscript context; and zero-shot LLM/VLM judges, when repurposed for figure scoring, often yield overly concentrated scores with limited fusion of visual and textual evidence. We introduce an annotated dataset of 3,857 scientific figures from peer-reviewed conference papers, each rated along four peer-review-oriented dimensions: Clarity, Relevance, Informativeness, and Structure. We propose SciFigAlign, a fine-tuned multimodal scorer that grounds figure quality assessment in manuscript evidence. Given a figure crop, caption, citing paragraphs, and light paper context, SciFigAlign fine-tunes CLIP and SciBERT end-to-end with per-modality cross-attention and CubeMLP fusion, jointly optimizing SmoothL1 regression with a within-paper ranking hinge loss. Under paper-level splits, SciFigAlign achieves a macro MAE of 0.3524 and a within-paper pairwise accuracy of 81.64% on the test set with n = 396, a 59% relative error reduction over the best LLM-as-judge baseline with MAE 0.864. Ablations confirm that manuscript-grounded inputs, citing-context denoising, and ranking supervision are all critical, showing that scientific figure assessment requires learned alignment between visual content and manuscript evidence rather than prompting alone, even with state-of-the-art VLMs.

cs.CV↗

arXiv2026

LUMOS: Latent Universal Medical Priors for Segmentation

General vision foundation models (VFMs) have been primarily developed on natural images, and their utility for medical image segmentation is therefore often considered to depend on costly adaptation or domain-specific fine-tuning. In this paper, we revisit this assumption from a different perspective: rather than requiring VFM segmentors to relearn visual regularities, we investigate whether the low-level visual priors necessary for anatomical delineation already lie dormant within general VFMs. We observe that frozen VFMs, despite lacking medical supervision, encode transferable visual regularities. These properties are not exclusive to natural images but are also fundamental to medical image understanding. Motivated by this observation, we propose Latent Universal Medical PriOrs for Segmentation (LUMOS), a novel framework that amplifies general VFM priors to conventional medical segmentors. LUMOS consists of two key components: (1) Pathfinder that distills visual cues from a frozen vision foundation model, and (2) Inspiror that sparks the conventional medical networks with spatial guidance from distilled visual regularities. In this way, the segmentor is relieved from learning complex visual regularities entirely from limited medical annotations and can instead focus on task-specific anatomical delineation. Across diverse medical datasets and token-based VFMs, LUMOS shows that general VFMs can serve as spatial prior generators when their frozen token spaces preserve patch-level pattern relevance. DINO provides stable matched-backbone gains, while SigLIP exposes VFM-specific sensitivity caused by its different token granularity and representation objective.

cs.CV↗

arXiv2026

Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance

Autoregressive (AR) models based on next-scale prediction have emerged as a powerful tool for image generation, but they face a critical weakness: information inconsistencies between patches across timesteps introduced by progressive resolution scaling. These inconsistencies scatter guidance signals, causing them to drift away from salient regions within the image and leaving behind ambiguous, unfaithful features during sampling. We tackle this challenge with Information-Grounding Guidance (IGG), a novel framework that anchors guidance to semantically important tokens via an attention-based dynamic weighting formulation, consequently ensuring that guidance and semantic contents remain tightly aligned. Across both class-conditioned and text-to-image generation tasks, IGG delivers sharper, more coherent, and semantically grounded images, demonstrating its efficacy for correcting AR-based methods. Our code is available at https://github.com/dnngky/infoground-guidance.

cs.CV↗

arXiv2026

Cell as Point: One-Stage Framework for Efficient Cell Tracking

Conventional multi-stage cell tracking approaches rely heavily on detection or segmentation in each frame as a prerequisite, requiring substantial resources for high-quality segmentation masks and increasing the overall prediction time. To address these limitations, we propose CAP, a novel end-to-end one-stage framework that reimagines cell tracking by treating Cell as Point. Unlike traditional methods, CAP eliminates the need for explicit detection or segmentation, instead jointly tracking cells for sequences in one stage by leveraging the inherent correlations among their trajectories. This simplification reduces both labeling requirements and pipeline complexity. However, directly processing the entire sequence in one stage poses challenges related to data imbalance in capturing cell division events and long sequence inference. To solve these challenges, CAP introduces two key innovations: (1) adaptive event-guided (AEG) sampling, which prioritizes cell division events to mitigate the occurrence imbalance of cell events, and (2) the rolling-as-window (RAW) inference strategy, which ensures continuous and stable tracking of newly emerging cells over extended sequences. By removing the dependency on segmentation-based preprocessing while addressing the challenges of imbalanced occurrence of cell events and long-sequence tracking, CAP demonstrates promising cell tracking performance and is 8 to 32 times more efficient than existing methods. The code and model checkpoints are available at https://github.com/YXSong000/CAP.

eess.IV↗

arXiv2026

Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations

Diffusion models generate conditional samples by progressively denoising Gaussian noise, yet the denoising trajectory can stall at visually plausible but low-quality outcomes with conditional misalignment or structural artifacts. We interpret this behavior as local optima in a surrogate quality landscape: Once early denoising commits to a suboptimal global structure, later steps mainly sharpen details and seldom correct the underlying mistake. While existing inference-time approaches explore alternative diffusion states via re-noising with fixed strength or direction, they exhibit limited capacity to escape steep quality plateaus. We propose Controlled Random Zigzag Sampling (Ctrl-Z Sampling),a scalable sampling strategy that detects plateaus in quality landscape via a surrogate score, and allocates exploration only when a plateau is detected. Upon detection, Ctrl-Z Sampling rolls back to noisier states, samples a set of alternative continuations, and updates the trajectory when a candidate improves the score, otherwise escalating the exploration depth to escape the current plateau. The proposed method is model-agnostic and broadly compatible with existing diffusion frameworks. Experiments show that Ctrl-Z Sampling consistently improves generation quality over other inference-time scaling samplers across different NFE budgets, offering a scalable compute-quality trade-off. Code available at: https://github.com/ShunqiM/Ctrl-Z-Sampling.

cs.CV↗

arXiv2026

A First Exploration of Neuromorphic OT-CFM for Multi-Speaker VSR

Visual Speech Recognition (VSR) tasks in complex multi-speaker scenarios are severely hindered by rapid head motions, occlusions, and subtle lip articulations. Traditional RGB-based methods struggle here due to low rates and motion blur of frames. To overcome these, we propose LipsFlow, a neuromorphic-inspired VSR framework that converts RGB videos into high-temporal-resolution event streams. For multi-speaker, we employ ByteTrack tracking and TalkNet active speaker detection to temporally segment scenes into single-speaker clips, enabling focused per-speaker analysis. By explicitly capturing microsecond-level articulatory dynamics via learnable event-based representations, LipsFlow achieves inherent robustness against visual degradation. To efficiently model these dense event-based features and adapt to speaker-specific articulatory patterns, we introduce Optimal Transport Conditional Flow Matching (OT-CFM). It enforces deterministic, straight-line trajectory generation in a semantic latent space, slashing inference latency to just two Ordinary Differential Equation (ODE) steps. Furthermore, we design a Dual-Level Semantic Supervision mechanism combining token-level BERT weight tying and sentence-level priors to resolve homophene ambiguities. Validated on competitive benchmarks, LipsFlow achieves a state-of-the-art WER of 22.3\% at 240 ms latency, establishing a highly robust and efficient paradigm for event-based VSR.

cs.MM↗

arXiv2026

AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation

Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.

cs.CV↗

arXiv2026

SAFE-DiT: Semantics-Aware Fast-path Execution for High-Resolution Diffusion Transformers

High-resolution Diffusion Transformer (DiT) inference contains substantial spatial redundancy, but many spatially adaptive implementations encode regional computation as attention masks, which can inadvertently move scaled dot-product attention (SDPA) away from FlashAttention fast paths. We identify this avoidable systems bottleneck as Mask-Induced Dispatch Tax (MIDT) and show that it grows with latent sequence length. We introduce SAFE-DiT, a training-free Semantics-Aware Fast-path Execution framework that separates exact mask elision from approximation-based spatial scheduling. SAFE-DiT removes only provenance-certified image self-attention masks that induce a row-wise constant shift in attention logits, preserves semantics-bearing masks such as text-padding masks, and realizes spatial adaptation through prompt-conditioned token partitioning, selective state updates with global context, and periodic context refresh. We call this acceleration-only configuration SAFE-Core and report sensitivity-weighted classifier-free guidance separately as SAFE-DiT+SW. On the evaluated PyTorch SDPA stack, redundant masks make long-sequence attention $4.1\times$ to $5.8\times$ slower than the mask-free path. On Lumina-Next, SAFE-DiT achieves $2.69\times$ end-to-end acceleration at $1024^2$ resolution and $5.09\times$ at $2560^2$, reduces peak memory at $2560^2$ from 94.1 to 27.9 GB, and enables $3072^2$ generation when dense inference runs out of memory. Paired metrics, component ablations, and a blinded human study support visual non-inferiority of SAFE-Core to the dense fast-path baseline, while SAFE-DiT+SW provides a separate prompt-alignment operating point without reintroducing spatial self-attention masks. Code is available at https://github.com/xuanhuayin/SAFE-DiT.

cs.CV↗

arXiv2026

EvLIR: Learning Illumination Residuals from Ordered Events for Low-Light Image Enhancement

Low-light image enhancement is severely ill-posed when the input frame contains missing structure, saturated noise, and weak local contrast. Event cameras provide asynchronous brightness-change observations with high temporal resolution, but prior works often treat voxel channels as an unordered or static feature stack before fusion, rather than explicitly modeling their within-window temporal evolution, weakening the temporal evidence that makes events useful. We propose EvLIR, a temporal-residual enhancement framework that learns illumination residuals from ordered events for low-light image enhancement. Given a low-light frame and its aligned event voxel, EvLIR preserves the ordered temporal bins of the event stream and introduces a Temporal Event Residual Module (TERM) to encode short-window event dynamics with a lightweight ConvGRU. The resulting temporal state is converted into a bounded illumination correction, which provides spatially adaptive photometric guidance for Retinex-style illumination estimation and subsequent reliability-aware image-event restoration. On SDE and SDSD indoor/outdoor benchmarks, EvLIR achieves the best result on eleven of twelve dataset-metric pairs, with average scores of 25.63~dB PSNR, 28.30~dB PSNR*, and 0.827 SSIM across the four benchmarks.

cs.CV↗

arXiv2026

CrossWeaver: Cross-modal Weaving for Arbitrary-Modality Semantic Segmentation

Multimodal semantic segmentation has shown great potential in leveraging complementary information across diverse sensing modalities. However, existing approaches often rely on carefully designed fusion strategies that either use modality-specific adaptations or rely on loosely coupled interactions, thereby limiting flexibility and resulting in less effective cross-modal coordination. Moreover, these methods often struggle to balance efficient information exchange with preserving the unique characteristics of each modality across different modality combinations. To address these challenges, we propose CrossWeaver, a simple yet effective multimodal fusion framework for arbitrary-modality semantic segmentation. Its core is a Modality Interaction Block (MIB), which enables selective and reliability-aware cross-modal interaction within the encoder, while a lightweight Seam-Aligned Fusion (SAF) module further aggregates the enhanced features. Extensive experiments on multiple multimodal semantic segmentation benchmarks demonstrate that our framework achieves state-of-the-art performance with minimal additional parameters and strong generalization to unseen modality combinations.

cs.CV↗