PyTorchKR Papers

최신 논문

총 16,105편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 카테고리 구독

arXiv:2611.25814v1cs.CV

What Does Anti-Hallucination Training Buy? Saying Less, Cheap Pruning, and the Rest

Preference training is the standard recipe for reducing object hallucination in vision-language models, and each new method reports a lower CHAIR score. We ask what that reduction consists of. Every trained system is compared with its base model's caption truncated to the same object recall, which removes the part of the gain that comes from saying less; the remainder is compared with what a deliberately weak, frozen, eleven-feature post-hoc probe fitted on 500 annotated images can reproduce. All criteria were specified in the analysis code and frozen before the outputs they judge existed. For eight public methods on LLaVA-1.5-7B, the probe reproduces at least two thirds of what training removes beyond saying less for 7 of 8 methods on COCO, 8 of 8 on AMBER and 6 of 8 on Object HalBench, with RLAIF-V the consistent exception; a later public method released for both LLaVA-1.5 and Qwen2.5-VL (SENTINEL) meets the same bar on both. On our own DPO checkpoints the accounting holds on Qwen2.5-VL but fails on InternVL2.5 on COCO (2 of 4 training designs). At matched recall, the base model plus the probe beats truncation for 8 of 9 LLaVA-1.5 systems on COCO and 9 of 9 on AMBER, on the two newer backbones, and when realized as actual sentence deletion. The same measure prices training signals: DPO on the model's own samples labelled by the probe or by an object-level CLIP score recovers 0.70 and 0.87 of the beyond-brevity gain of ground-truth-labelled DPO on Qwen2.5-VL (the probe 0.73 on InternVL2.5), both beat length-matched random labels, and CLIP labels buy more than probe labels on COCO. Human judgements show that the sentences the probe deletes are wrong more often than those truncation would delete; a human check of the CHAIR evaluator's precision falls below our pre-set bar, which we report as a limitation.

arXiv:2611.25802v1cs.CV

World-delta: Continually Integrating Observations for Incremental 3D Scene Generation

Existing 3D generation methods synthesize visually compelling scenes from fixed inputs, yet struggle to continually expand and revise a 3D world as new observations arrive. Iterative approaches often fuse real observations and generated content into a single representation, treating generated content as established structure rather than revisable predictions. Consequently, early generation errors can persist and accumulate despite new evidence. We propose World-delta, an incremental 3D scene generation framework combining uncertainty-aware persistent world modeling with delta-guided diffusion for continual scene expansion and correction. Our framework comprises three components: 1) Visual Evidence Encoding integrates geometric and semantic cues from incoming images into unified world tokens. 2) Observation-Calibrated World Modeling maintains a persistent latent world state that distinguishes observation-supported content from provisional completion hypotheses. New evidence drives selective world deltas to revise conflicting hypotheses while preserving well-supported content. 3) Incremental World Diffusion Generation uses delta-derived masks to restrict diffusion to regions requiring updates, freezing unchanged regions. Only affected 3D chunks are decoded and integrated into the existing scene. Together, these components enable continual world expansion and revision while reducing redundant generation and preserving scene consistency. Experiments demonstrate the effectiveness of World-delta in incremental 3D scene generation, observation-driven correction, scene consistency, and update efficiency over successive scene updates.

arXiv:2611.25801v1cs.CV

Fast and Accurate Body Estimation in Pixels, World Space and Scenes

We present FAB (Fast and Accurate Body estimation), a human body estimator that jointly improves inference efficiency and body-recovery fidelity. Rather than predicting only parametric mesh-model parameters, FAB directly predicts human skeleton joints and dense surface points. In images, it predicts dense 2D and 3D body points alongside a parametric Momentum Human Rig (MHR) head, then fits SMPL in one learned pass followed by two Gauss-Newton steps. Its ViT-L student, FAB-L, distilled from a fine-tuned SAM 3D Body, runs in real time for one person on an RTX 4070 Ti SUPER, from detection to a fitted body, at 4.6x the speed of SAM 3D Body's eager-PyTorch body-only path, while improving accuracy on most camera-space benchmarks. The same network predicts heel, toe, hand, and knee contacts, which connect the body to the scene. FAB also estimates world-space motion, fuses static-scene depth around the visible person, and aligns body and scene using predicted contacts. For human-to-humanoid real-to-sim transfer, an in-model G1 head and FAB-Retarget, a motion-level retargeter that reads dense points or the fitted mesh, map motion to a Unitree G1. FAB-Retarget produces references that follow the person more closely than specialized retargeters including SOMA, GMR, UMR, and OmniRetarget, and substantially more closely than Fast SAM 3D Body retargeted with our sequence IK. Project page: https://www.jinkuncao.com/FAB.

arXiv:2611.25787v1cs.CV

GeoPillars: A Large-Kernel Pillar Detector for Single-Sweep LiDAR Under Point Sparsity, and a Decomposition of Its Gain

GeoPillars is a pillar-based 3D detector for single-sweep LiDAR under a hard budget of eight points per 0.4 m pillar, a budget that discards 46.6% of in-range points. It adds geometric pillar features, grouped vector attention within each pillar, a learned absolute positional prior on the bird's-eye-view (BEV) grid, a large-kernel backbone, auxiliary surface-normal supervision and a full-resolution detection head. Trained on 123k synthetic CARLA scans and evaluated on cars, it reaches 81.33 BEV AP@0.7, 13.76 points above CenterPoint at the same input budget and 1.91 points above CenterPoint at its native, denser budget. These are in-distribution figures, because the validation split is per-frame rather than scene-disjoint. A rung-by-rung decomposition shows that the new components explain little of the gain. Raising the detection-head resolution alone recovers 10.19 of the 13.76 points, the joint change of encoder, backbone, neck and positional prior adds 3.37, and the normal-supervision branch adds 0.20, which is within run-to-run variation. With heading reversals folded out, orientation error is similar for all seven models, although GeoPillars reverses headings about half as often as a resolution-matched baseline. It is also the slowest of the seven models. Evaluated zero-shot on 6,743 real nuScenes keyframes, no model transfers, and each retains 3.7% to 12.2% of its CARLA accuracy. A collection defect left LiDAR intensity constant, so every model learned from geometry alone. Code and checkpoints are publicly available.

arXiv:2611.25776v1cs.CV

Suppressing the Shortcut Hidden Behind the Gains of VideoLLM Fine-Tuning

Supervised fine-tuning (SFT) is the stage in which Video Large Language Models (VideoLLMs) advance their capability to reason over video. However, the VideoLLM also learns to answer from textual regularities of its training data, a shortcut that bypasses the video. The regularities behind the shortcut are hard to define at the data level, and the shortcut forms regardless of training epoch, data size, or backbone. The shortcut is hidden behind a rise in accuracy within the training distribution, yet it is not required for the rise and even costs SFT the gain the video offers under distribution shift. We therefore propose Video-Evidence Training via Orthogonal projection (VETO), which intervenes on the fine-tuning update where the shortcut is written rather than on the data. VETO calibrates the subspace that the text alone occupies in the activations of the base model and projects the update onto its orthogonal complement. The VideoLLM is thus left to ground its answer in the full input, with the shortcut suppressed as it forms and never defined. Across seven VideoLLMs, VETO preserves the gain of SFT on the training distribution and suppresses the shortcut. On six benchmarks under shift it gains more than SFT and answers from the full input rather than the shortcut. Code is available at https://github.com/Youngwoo-git/VETO.

arXiv:2611.25771v1cs.CV

Driver Gaze Target Identification via Gaze Estimation and Structured Driving Scene Understanding

Identifying a driver's gaze target provides object-level information for visual attention monitoring in intelligent vehicles. Gaze-zone classification and continuous attention localization provide limited information about the identity of the attended scene target. This paper proposes a modular framework that identifies object- or region-level gaze targets from synchronized in-cabin driver images and forward-view road images. The framework comprises four components: MFGaze, which combines a Vision Transformer pretrained with a masked autoencoder (MAE) and head-pose conditioning through feature-wise linear modulation (FiLM); panoptic scene understanding, which constructs a structured candidate set; polynomial ridge regression, which maps gaze features to 2D scene-image coordinates through individualized calibration; and entropy-driven target inference, which adaptively fuses boundary-distance geometry with class-level semantic priors. Experiments evaluate the framework using three public gaze datasets, a laboratory test using Cityscapes road scenes and self-collected gaze data, and the Look Both Ways (LBW) dataset. The complete framework achieves an overall Top-1 gaze target identification accuracy of 96.27% in the laboratory experiment. After sequence-specific calibration, it achieves Top-1 accuracies of 79.92% for task-relevant objects and 85.58% for the road class on LBW. MFGaze achieves the lowest mean angular errors among the compared methods, with 4.06$^\circ$, 4.23$^\circ$, and 10.52$^\circ$ on MPIIFaceGaze, ETH-XGaze, and Gaze360, respectively, and demonstrates strong cross-dataset generalization. Sensitivity analyses show stable performance under hyperparameter perturbations. These results support the framework as a source of object-level gaze information for driver monitoring.

arXiv:2611.25761v1cs.CV

AdaptReg: Skip-level FiLM Adapter for Task-Specific Adaptation of a Frozen Registration Foundation Model

Longitudinal PSMA-PET/CT registration is a key step in assessing treatment response in prostate cancer, and it requires more than generic alignment accuracy: the deformation must preserve lesion volume, so that measured changes in metabolic tumor volume (MTV) and total lesion glycolysis (TLG) reflect treatment response rather than registration artifacts. We propose AdaptReg, which adapts a frozen registration foundation model to such task-specific requirements through a lightweight trainable adapter branch with identity-preserving initialization. Keeping uniGradICON fully frozen, we apply FiLM-style per-voxel affine modulation ($\gamma$, $\beta$) to its decoder skip connections; the modulation is generated from anatomical semantic features of a frozen TotalSegmentator encoder and supervised by a lesion volume-preservation loss. On the official Codabench validation set (20 pairs), AdaptReg achieves DSC 0.7585, HD95 6.33 mm, MTV/TLG errors of 1.68%/2.11%, and 0.0022% non-diffeomorphic voxels, improving on the uniGradICON + instance optimization baseline by $+0.05$ DSC and reducing the MTV/TLG errors by 70.9%/67.7%. On the hidden official test set (160 pairs), AdaptReg achieves CT DSC 0.816, PET DSC 0.598, and MTV/TLG errors of 4.02%/3.46%, ranking fourth among 16 teams in the final PSMAReg evaluation. Code: https://github.com/shishuyue/AdaptReg

arXiv:2611.25747v1cs.CV

SAM 3 Sees More Than It Detects: Promptable Concept Maps Beyond Categories

Promptable segmenters answer what a thing is. Dense prediction tasks also ask class-agnostic questions, such as what stands out or hides, how light is blocked, transmitted or reflected, and where structure lies. Robots, surveillance systems and image editors build on their answers. We study promptable concept maps (PCM), which extend SAM 3's promptable concept segmentation from noun instances to maps of concepts named by short noun phrases such as `"transparent object", with "none" as an answer. PCM-Bench groups the standard test sets of seven such tasks by the question they ask. On it, we find that SAM 3 sees more than it detects. Its semantic path responds to many of these concepts, yet a low image-level presence score can block every detection of its instance path. We present UNVEIL, which trains a copy of SAM 3's semantic path for these maps, with LoRA on copies of its last blocks. A separate presence decoder learns from verified negatives whether a concept occurs. With one text phrase per query, UNVEIL maps region concepts such as salient and camouflaged objects, shadows, transparent objects and mirrors. The same weights map thin structures such as edges and skeletons as well as object categories. Run next to SAM 3, it leaves SAM 3's own outputs unchanged. On each of the seven tasks, UNVEIL reaches the highest score among the compared promptable models, including SAM 3's official fine-tuning and pipelines built on multimodal LLMs. Code is available at https://github.com/yuhuan-wu/UNVEIL.

arXiv:2611.25731v1cs.CV

Towards High-Fidelity Learned Video Compression

Professional video workflows operate at qualities that learned video codecs do not reach. We measure how much of a professional quality range a codec covers, from the lowest quality of the standard test conditions, VVC at QP 37, to the highest 4:2:2 grade of the production codec ProRes. To compare inter-frame coding alone, every codec receives original frames as references at the intra positions. On 10-bit YCbCr 4:2:2 test sequences, four recent learned codecs cover 24 to 38% of this range. Their encoders are feed-forward networks with a fixed latent capacity, the number of latent values per sample, of at most 0.25, and removing their quantization raises the quality by at most 1 dB. We propose a learned inter-frame codec whose encoder searches, as the encoders of conventional codecs do. It searches block motion, which a learned network refines, and it searches among learned block transforms whose latent capacity ranges up to 1, selecting the partition, the skip decisions, and the quantized coefficients by a rate-distortion cost. Our codec reaches the upper end of the range on all six sequences. With all residual blocks fixed to 16x16, whose latent capacity is 0.25, the same codec covers 55% of the range, with or without quantization.

arXiv:2611.25730v1cs.CV

CGM: Chaos-Guided Graph Attention and Expert Routing for Autism Stereotypy Detection

Skeleton graph attention for autism video analysis weights joints by content alone, giving an erratic wrist the same standing as a still one. Stereotypy, however, is defined by how unpredictably movement unfolds rather than how much of it occurs: in our cohort, every amplitude-based kinematic descriptor falls in stereotyped clips, while windowed Sample Entropy of angular velocity alone rises (Hedges' g = +1.05). We compute this irregularity as a per-joint, per-frame tensor that structurally drives computation. The proposed chaos-guided graph attention and mixture-of-experts model (CGM) couples two mechanisms. First, a learned pairwise term added to the graph-attention logit concentrates attention on irregular joints and re-ranks the remaining edge weight toward jointly irregular pairs; its unconstrained coefficient neither vanishes nor inverts, and zeroing the tensor realigns attention in 94% of clips. Second, a gate conditioned on the same signal routes each frame to two of four experts, whose weights track frame-level irregularity. Under subject-grouped five-fold cross-validation on 116 clips, CGM reaches 0.930 balanced accuracy, exceeding the graph-attention baseline by +0.187; the component ordering replicates on the public MMASD dataset, and on held-out Panoptic Studio sessions, CGM keeps 0.86 specificity without retraining, against at most 0.35 for chaos-free baselines. In ablation, routing without irregularity yields no reliable gain, whereas chaos modulation alone contributes +0.132. A severity index from the same tensor reaches 0.87 split-half reliability, separates the cohorts with Hedges' g = +1.08, and stays consistent within children across therapy recordings, supporting video-only monitoring of stereotypy during intervention.

arXiv:2611.25728v1cs.CV

What's in the box? Revealing Object Layouts Inside Opaque Boxes via Contactless Vibration Excitation and Sensing

Can we reveal what is inside an opaque box by observing its surface vibrations? Objects inside a box alter its mechanical response, providing cues about their hidden locations. We introduce a method that uses these cues to predict a box's 2D occupancy layout entirely remotely. A nearby speaker excites vibrations, which we measure on one visible wall using a grid of laser points and a camera measuring each point's speckle-pattern shifts. Our model, WhisperBox, learns to encode these vibrations into a latent representation and decode it into an occupancy map of the box's contents. The method requires no physical contact to excite or sense vibrations. We demonstrate layout recovery across various boxes and single- and multiple-object scenes, with generalization to unseen object instances, positions, orientations, and weights, as well as changes in laser placement and excitation source. Our analysis reveals that informative frequency bands need not coincide with the box's resonances and that the learned latent space captures object placement and supports arithmetic operations. These results demonstrate the potential of remote vibration sensing to infer spatial layouts hidden from view.

arXiv:2611.25724v1cs.CV

Verifying Before Enhancing: Target Evidence Chain Verification for Infrared Small Target Detection

Infrared small target detection is essential for early warning, surveillance, remote sensing, and space-based target perception. However, infrared small targets are typically tiny, weakly textured, and boundary-blurred, making them difficult to distinguish from complex backgrounds. Moreover, target-like background structures can produce salient responses across multiple scales, making response strength alone insufficient for reliable target discrimination. To address this challenge, we propose VBE-UNet, a Target Evidence Chain Verification Network that assesses candidate reliability beyond response magnitude through complementary evidence cues. Specifically, counterfactual background explanation models local target evidence, while joint evidence verification integrates cross-scale support and perturbation-based contribution verification to guide feature recomposition. The resulting reliability estimates guide selective feature refinement, promoting target-relevant responses while reducing unsupported background activations. Extensive experiments on IRSTD-1K, NUDT-SIRST, and NUAA-SIRST demonstrate that VBE-UNet achieves superior detection accuracy and lower false-alarm rates compared with state-of-the-art methods. Ablation studies and evidence reliability analysis further validate the effectiveness of the proposed verification framework.

arXiv:2611.25722v1cs.CV

DACA: Dual-Anchor Collaborative Agent for Zero-Shot 3D Visual Grounding

3D visual grounding (3D-VG) is a task that bridges the natural language and 3D space, serving as a foundational capability for embodied intelligence and human-computer interaction. Although recent 3D-VG agents overcome the reliance on pre-processed 3D point clouds, they essentially employ a unidirectional, pipeline-style tool-calling architecture. This architecture separates semantic understanding from geometric reconstruction, leading to error accumulation and an inability to autonomously recover from failures. To address this flaw, this paper proposes the Dual-Anchor Collaborative Agent (DACA), which reinterprets 3D-VG as a semantic-geometric dual-anchor collaborative reasoning paradigm. DACA comprises two complementary specialized tools: the semantic anchor tool holds the visual semantic features of the target and achieves efficient target tracking via optical flow and VLM local verification; the geometric anchor tool holds the 3D spatial geometric features of the target and performs global spatial reasoning via voxelization and multi-keypoint projection. The two components exchange information bidirectionally via a standardized tool call protocol, autonomously deciding on the next perception-action step to form a closed-loop optimization. Experiments on the ScanRefer and Nr3D benchmarks demonstrate that zero-shot DACA achieves an overall Acc@0.25 of 73.6\% on ScanRefer and 70.4\% overall accuracy on Nr3D, far surpassing the latest zero-shot 3D-VG methods. The code and models are available at https://github.com/liujia99/DACA

arXiv:2611.25719v1cs.CV

SteerTrace: Tracing How Target Attacks Steer World Action Models

World action models (WAMs) predict future videos and actions with one backbone, and their imagined future is often seen as a safety signal. We introduce SteerTrace, a targeted universal attack on WAMs that also traces its effect inside the model. One $\ell_\infty$-bounded (4/255) perturbation per model, shared across frames, tasks, and camera views, pushes the executed action toward an attacker-chosen target. On eight WAMs from four families (FastWAM, Faster-WAM, OpenWAM and LingBot-VA) on LIBERO and RoboTwin, it drives task success close to zero and moves the attacked action toward the target. Same-budget noise and BadWAM produce little action shift in the target direction. We use the signed targeted effect to measure how much of the effect each intervention removes or reproduces. Under the standard attack, restoring future tokens removes 61-94% of the targeted effect in the five models whose action experts attend to them. When the attack is constrained to keep the predicted future close to the clean prediction, the targeted effect can become more dependent on current-frame tokens. In three of these five models, restoring current-frame tokens removes 87-99% of the effect, compared with 9-15% for future tokens. In six models, restoring just 0.36% of the FFN neurons in layers 10-29 of the video expert removes 72-96% of the targeted effect. Residual directions 2-16 hold only 11-18% of the shift energy, yet removing them cancels most of the targeted effect.

arXiv:2611.25713v1cs.CV

Interpreting DeepStack-Style Vision-Language Models: Do Input or Deeper-Layer Visual Tokens Drive Answers?

DeepStack-style vision-language models feed visual tokens to the language model at its input, and again at deeper layers. We call these the main entry and the later entries. What do the later entries add when the model already has the image? We give the entries versions that differ in one object's color and see which one the answer follows. On photographs, the original DeepStack follows the later entries, while Qwen3-VL-8B/32B follow the main entry. Does weakening the main entry give the later features more influence? On synthetic images, we keep this conflict and reduce color saturation at the main entry only. The unchanged later features gain a stronger effect on the relative scores of the two colors, in all four models, including Granite Vision. This holds even under mild desaturation, when accuracy without the conflict stays at 100%. To test whether later features can also correct errors, we degrade both entries and restore only the later ones. The same blur costs both Qwen models 50 percentage points of accuracy on stripe orientation. The 32B model recovers 48 points and the 8B model only 8. Accuracy with one clean entry predicted which entry each model can lose, and five of six tested predictions held on new GQA attribute questions. With the other entry zeroed, Qwen3-VL-8B loses only 0.3 points on its main entry alone, and DeepStack at most 3.45 points on either entry alone. In split inference on these questions, DeepStack can then send only its main-entry features at higher precision. This takes 49% fewer bytes than all entries at a shared precision, at similar accuracy.

arXiv:2611.25712v1cs.CV

SkillLens: Action-Guided Dense Visual Dynamics for Proficiency Estimation

Human proficiency estimation is challenging because skill is a high-level and often subjective property that is not captured by action identity alone. Existing approaches leverage multimodal observations such as ego-exo video, gaze, pose, and inertial signals, but the gap between sensor-level evidence and abstract proficiency labels remains difficult to bridge. We present SkillLens, a proficiency estimation framework that uses atomic action descriptions as temporal guidance, rather than direct semantic supervision, to localize proficiency-relevant windows for dense visual-dynamics extraction. Each window is summarized into an action-dynamics feature and scored by a high-low prototype margin against higher- and lower-proficiency executions from the same activity-action category. Fusing these ego/exo PDM cues with a global ego-exo-gaze baseline through CORAL ordinal calibration, our method achieves 0.542 accuracy on Ego-Exo4D. Our project page is available at https://skill-lens.net/

arXiv:2611.25711v1cs.CV

Calibrate in Context: Efficient Data-Free Quantization from Small Synthetic Pools

We enable accurate low-bit recovery from small synthetic pools by improving both image synthesis and student preparation. A grid with lower local response error can incur greater error in the student's block context. Functional Grid Calibration (FGC) selects activation scales and integer offsets against full-precision responses constructed from the student's current block input. With weights fixed during calibration, FGC evaluates candidates only at their consuming operators, bringing block context to calibration at nearly local-scoring cost. For DeiT and ViT, progressive inversion updates the same images through increasingly deep teacher prefixes. With 1,000 images, the complete pipeline exceeds MimiQ's reported accuracy in all 18 W4A4/W5A5 settings across nine backbones. With 256 images, DeiT-S and ViT-S reach 68.20% and 66.90% ImageNet top-1 accuracy at W4A4 in 26-27 minutes on one RTX 5090. The pipeline outperforms strengthened 1K-image MimiQ by 13.03 and 24.36 percentage points at lower preparation cost and the same recovery-update count. The same low-budget checkpoints retain their advantage on ImageNetV2.

arXiv:2611.25708v1cs.CV

Slide2PPT: Native PowerPoint Reconstruction with Localization-Guided Refinement

Image-to-slide reconstruction must recover both visual appearance and native object structure for subsequent editing. Yet visually faithful charts and tables may lack editable series and cells, and direct code generation couples content prediction to rendering implementation. We present Slide2PPT, a model-agnostic, training-free framework that equips both open- and closed-source vision-language models with native PowerPoint reconstruction capabilities. Through a unified typed interface, the model predicts object records while a fixed renderer materializes their native properties, enabling integration with different backbones without task-specific training or architectural changes. Localization-guided refinement then connects visual discrepancies to object records for correction. Regional updates preserve unselected records, while whole-slide updates address cross-region relationships. Record refinement and targeted object editing support structural and attribute-level changes, with rendered feedback governing iterative acceptance. To evaluate this framework across diverse reconstruction demands, we introduce Slide2PPT-Bench, covering controlled elements, real slides, and scientific figures. Evaluation separately measures file production, rendered agreement, and native-object availability. Experiments across open- and closed-source backbones demonstrate substantially more reliable PPTX generation than direct code generation, improved layout and text recovery on scientific figures, and native chart and table interfaces beyond geometric primitives.

arXiv:2611.25696v1cs.CV

LIFT: Lifting Language-Controlled Image Fusion to Free-Viewpoint 3D Scenes

Multimodal image fusion traditionally generates a single fused image from paired observations at a specific viewpoint, limiting comprehensive scene perception and adaptability to diverse perceptual objectives. To overcome these limitations, we present LIFT, a novel framework for our proposed language-controlled 3D image fusion (L3D-IF) paradigm that extends multimodal image fusion from 2D images to free-viewpoint 3D scenes, enabling language-guided fusion tailored to diverse perceptual objectives. First, LIFT formulates Semantic Modality Binding (SMB), which interprets language instructions to identify semantic targets and their associated modality preferences and grounds these targets in corresponding regions of a shared Gaussian scene, establishing persistent semantic-to-spatial correspondences for consistent fusion control across viewpoints. Second, Language-Conditioned Gaussian Appearance (LCGA) is developed to achieve scene-wide multimodal appearance modeling by combining adaptive multimodal blending with language- and viewpoint-conditioned appearance refinement, enabling spatially adaptive realization of diverse fusion appearances while preserving source fidelity at pure-modality endpoints. Finally, we construct a Multi-View Language-Supervision (MVLS) dataset that associates language instructions and calibrated viewpoints with modulated fused images, providing effective supervision for learning instruction-dependent fusion from finite multi-view observations. Extensive experiments demonstrate that LIFT achieves coherent multimodal fusion across continuous viewpoints, accurate semantic target localization, and consistent regional fusion control under diverse language instructions.

arXiv:2611.25689v1cs.CV

D3-Cache: Accelerating Diffusion Models with Dynamic Mode Decomposition

Diffusion Transformers generate high-quality images and videos, but every sampling step requires a full network evaluation. Training-free caching reduces this cost by reusing or forecasting network outputs across steps, but common approaches use per-model calibrated skip criteria or maintain per-layer prediction histories. We present D3-Cache, which applies Dynamic Mode Decomposition (DMD) to Diffusion sampling: it fits DMD to a short history of exact velocities and forecasts only the velocity that the scheduler consumes. A horizon-aware gate decides when to skip: whenever the network is evaluated, its output is compared with the forecast for that step, and the larger the error, the sooner the network is evaluated again, up to a cap on consecutive skipped steps. D3-Cache requires no training or per-model offline calibration curve and adds negligible memory. It shares one history length and risk tolerance across the four evaluated models. On four state-of-the-art image and video models, D3-Cache achieves speedups of 2.6-3.8x with high fidelity to the base model output.