DA-VAE tackles the challenge of scaling latent diffusion models to higher resolutions without linearly increasing token counts. The core idea is a structured latent representation: keep the original pretrained VAE latent channels as a 'base' and append additional 'detail' channels that encode high-resolution information, enforced by a simple alignment loss. This allows a pretrained diffusion model to be fine-tuned rather than retrained from scratch, promising significant compute savings.

FeatDistill tackles robust detection of AI-generated images under real-world degradations via a multi-expert ensemble of CLIP and SigLIP backbones. The framework combines extensive data expansion with a two-stage training paradigm featuring feature-level self-distillation. It aims to balance strong generalization across unseen generators with practical inference efficiency.

This paper tackles coronary artery segmentation from CTA images, a challenging task due to slender tubular morphology and severe class imbalance. The authors propose MDSVM-UNet, a two-stage framework that combines multidirectional snake convolution (MDSConv)—extending deformable convolution to three anatomical planes—with residual visual Mamba (RVM) for linear-complexity long-range dependency modeling. The approach aims to capture both local geometric priors of vessels and global inter-slice context while maintaining computational efficiency suitable for clinical deployment.

This paper tackles a fundamental question in multimodal large language models (MLLMs): should the vision encoder be fine-tuned or frozen during instruction tuning? The authors identify visual preference conflicts—where diverse linguistic instructions pull encoder parameters in conflicting directions—as the root cause of instability in existing visual fine-tuning (VFT) methods. They propose CoVFT, a context-aware framework that extracts multimodal context vectors and routes visual tokens through mixture-of-experts layers to decompose these conflicts, achieving consistent gains across 12 benchmarks.

Federated learning for medical imaging typically requires task-specific pipelines and assumes homogeneous modalities across institutions, limiting real-world deployment where hospitals use diverse scanners (MRI, CT, PET) and need to support multiple downstream tasks. OmniFM proposes a frequency-domain insight: low-frequency spectral components exhibit cross-modality consistency and encode modality-invariant anatomical structures, enabling a single reusable optimization pipeline. The framework combines Global Spectral Knowledge Retrieval, Embedding-wise Cross-Attention Fusion, and Prefix-Suffix Spectral Prompting, regularized by Spectral-Proximal Alignment to stabilize aggregation under severe modality heterogeneity.

This paper proposes SparseVoxelDet, the first fully sparse object detector for event cameras that processes asynchronous event data using 3D sparse convolutions throughout the entire pipeline—from voxelization through backbone, feature pyramid, and detection head—without ever instantiating a dense feature tensor. On the FRED drone detection benchmark, the model achieves 83.38% mAP@50 (within 4.3 points of the dense YOLOv11 baseline) while processing only ~14,900 active voxels per frame (0.23% occupancy at 640×640) instead of all 409,600 pixel positions, yielding 858× GPU memory compression and storage costs that scale with scene activity rather than sensor resolution.

PEARL tackles training-free open-vocabulary semantic segmentation (OVSS), where the goal is to segment images into classes defined by arbitrary text prompts without fine-tuning the vision-language backbone. The core idea is an align-then-propagate pipeline: (1) Procrustes alignment rotates attention keys toward the query subspace inside the last self-attention block to fix spatially inconsistent patch geometry, and (2) a text-aware Laplacian propagation refines logits on a compact grid using a confidence-weighted graph that couples image gradients with text-based semantic similarity. This matters because it delivers state-of-the-art training-free accuracy with a frozen CLIP encoder, adding only modest computational overhead.

DUO-VSR tackles the prohibitive sampling cost of diffusion-based video super-resolution by enabling efficient one-step generation. The paper identifies critical limitations when applying Distribution Matching Distillation (DMD) to VSR—specifically training instability, degraded supervision from frozen score models, and insufficient guidance capped by teacher quality—and proposes a dual-stream strategy that unifies DMD with adversarial supervision via Real–Fake Score Feature GAN (RFS-GAN). This three-stage pipeline achieves approximately $50\times$ speedup over multi-step counterparts while delivering superior perceptual quality, making high-fidelity video upscaling practical for real-world deployment.

ALADIN tackles person Re-identification by distilling fine-grained attribute knowledge from a frozen CLIP teacher into a lightweight student network. The core innovation uses a Multimodal LLM (Qwen-VL) to generate structured attribute descriptions, which are converted via CLIP into spatial attention maps for supervising local feature alignment. A Scene-Aware Prompt Generator (SAPG) creates image-specific soft prompts via $\mathbf{p}=\mathrm{MLP}(\mathbf{f}_{g})$ to adapt text embeddings to surveillance scenes. At inference, only the student runs, promising deployable efficiency.

NoOVD tackles a critical issue in open-vocabulary object detection (OVD): during training, novel-category objects are forcibly aligned with background embeddings, causing them to be filtered out by the RPN and misclassified by the RoI head. The authors propose a framework built on frozen CLIP that identifies latent novel objects during training via generic text prompts (e.g., 'This is an object, specifically an animal') and integrates them through self-distillation. At test time, a Re-weighted RPN (R-RPN) boosts proposal scores using CLIP-based knowledge to improve novel-category recall. The method aims to eliminate the training-inference gap without requiring additional labeled data or pseudo-labeling noise.

This paper investigates whether video transformers can detect respiratory distress from video recordings of post-exercise recovery. The authors frame the problem as a temporal ordering task—predicting which of two clips shows greater shortness of breath—and propose augmenting ViViT with Lie Relative Encodings (LieRE) and Motion-Guided Masking (MGM). An F1 score of 0.81 is achieved, though on only 7 test videos from 3 participants.

Remote sensing text-to-image generation suffers from a lack of domain-specific diffusion transformers and prohibitive costs for high-resolution training. Existing training-free resolution promotion methods apply static RoPE scaling that uniformly compresses the spatial spectrum, which is particularly harmful for RS imagery due to its characteristically denser high-frequency energy. This paper proposes SHARP, a spectrum-aware dynamic adaptation strategy that uses a rational decay scheduler $\kappa_{rs}(t)$ to apply strong positional extrapolation early in denoising (for layout formation) while progressively relaxing it later (for detail recovery). The approach enables robust multi-scale generation up to 2.5$\times$ extrapolation factors with negligible overhead, addressing a critical gap in large-scale RS synthesis.

Accurate respiratory motion modeling is critical for radiotherapy precision, yet patient-specific breathing patterns are difficult to predict outside observed ranges. This paper proposes PRISM-RM, a trajectory-aware implicit neural representation (INR) that models lung motion as a continuous diffeomorphic flow driven by external surrogate signals. By integrating neohookean hyperelastic constraints with temporal total-variation regularization, the method eliminates the need for fixed reference breathing states and aims to improve extrapolation to unseen respiratory phases.

Pointing-based methods improve Large Vision-Language Models (LVLMs) by grounding objects before answering, yet the underlying mechanism remains unclear. This work investigates why pointing helps by comparing Direct Counting against Point-then-Count (PtC) in zero-shot counting tasks using synthetic data with controlled spatial layouts. The authors find that intermediate coordinate supervision encourages skill learning rather than narrow task memorization, yielding stronger out-of-distribution generalization while providing verifiable visual explanations.

Diffusion models generate high-quality images but require hundreds of denoising steps, making deployment on edge devices impractical. This paper proposes Coarse-to-Fine Diffusion Models that start with low-resolution denoising early in the process (when outputs are noisy anyway) before switching to high-resolution, plus a fast time-step search method that finds good sampling schedules in under 10 minutes instead of days.

Generative zero-shot learning (ZSL) synthesizes visual features for unseen classes conditioned on semantic prototypes, but existing methods often produce task-agnostic features that overlap for semantically similar yet visually distinct categories. This paper proposes RLVC, an outcome-reward reinforcement learning framework that treats the feature generator as a policy model and optimizes it using classifier confidence as the reward signal. The method further incorporates class-wise visual prototypes via a distillation loss to align synthesized features with real data distributions, achieving reported state-of-the-art results on CUB, SUN, and AWA2 benchmarks.

This paper tackles real-world image restoration (Real-IR) by adapting the 12B-parameter FLUX.1-dev flow matching model to low-level vision tasks. The core innovation is ResFlow-Tuner, which combines Unified Multi-Modal Fusion (UMMF) of image and text cues with a novel test-time scaling (TTS) paradigm that greedily optimizes ODE sampling trajectories using a multi-reward ensemble during inference. This establishes a new compute-quality trade-off for generative image restoration, showing that carefully perturbing intermediate flow states can yield substantial perceptual gains without retraining the base model.

The paper tackles the computational bottleneck of radiative transfer models (RTMs) for hyperspectral image (HSI) generation by proposing a VAE-based emulation framework that learns latent representations conditioned on biophysical parameters. It introduces both pixel-to-pixel (P2P) and fully convolutional (FC-VAE) variants, trained via either direct one-step mapping or a two-step pretraining strategy that decouples representation learning from parameter-to-latent interpolation. The work is significant for remote sensing applications as it provides empirical evidence that optimal emulator architecture depends critically on whether the target data is simulated (where P2P excels) or real-world imagery (where FC-VAE-pre dominates), and demonstrates that emulated data preserves downstream utility for parameter retrieval tasks.

Optical flow estimation traditionally requires expensive ground-truth annotations or relies on unreliable brightness constancy assumptions that fail under occlusion and illumination changes. This paper introduces GenOpticalFlow, a framework that synthesizes perfectly aligned training pairs by using monocular depth estimates to generate pseudo-optical flow, then conditioning a latent diffusion model to render corresponding next frames. The core innovation is converting unsupervised optical flow learning into a supervised training paradigm using synthetic data with geometrically consistent motion fields, potentially eliminating the need for manual annotation at scale.

StreamingClaw addresses real-time streaming video understanding for embodied intelligence applications such as autonomous driving and robotics. The framework unifies continuous perception, hierarchical multimodal memory, and proactive interaction through a main–sub-agent architecture where StreamingReasoning orchestrates StreamingMemory and StreamingProactivity sub-agents. By integrating incremental KV-cache reuse with dynamic pruning, memory evolution from atomic actions to events, and trigger-based proactive responses, it aims to close the perception–decision–action loop for physical world deployment.