WACV 2026 Papers

Generating high-quality sketches from video requires a nuanced understanding of semantic content and visual structure, particularly for complex scenes across diverse sketch styles. Efficient and flexible video-to-sketch style transformation remains a significant challenge. We introduce VideoSketcher, a training-free framework for style-controllable sketch video generation that preserves frame structure while applying specified sketch aesthetics. Leveraging text-to-image diffusion models, VideoSketcher utilizes strong semantic priors without the need for extensive training. Our approach enforces temporal consistency by retaining latent information across frames and employs a Time-Linked Attention mechanism to capture structural elements from the source video and inject stylistic information from the reference image. To bridge the semantic gap between sketches and original video content, we introduce Sketch Directive Amplification for selective transfer of stylistic features. Additionally, a Stroke Graph Regularization strategy, comprising line and point loss, refines line consistency in the latent space. Extensive experiments validate VideoSketcher's superior temporal stability and fidelity across diverse sketch styles and content. Video demos can be found in the supplementary materials.

Accurately recovering 6D poses in densely packed industrial bin-picking environments remain a serious challenge, owing to occlusions, reflections, and textureless parts. We introduce a holistic depth-only 6D pose estimation approach that fuses multi-view depth maps into either a fine-grained 3D point cloud in its vanilla version, or a sparse Truncated Signed Distance Field (TSDF). At the core of our framework lies a staged heatmap mechanism that yields scene-adaptive attention priors across different resolutions, steering computation toward foreground regions, thus keeping memory requirements at high resolutions feasible. Along, we propose a density-aware sparse transformer block that dynamically attends to (self-) occlusions and the non-uniform distribution of 3D data. While sparse 3D approaches has proven effective for long-range perception, its potential in close-range robotic applications remains underexplored. Our framework operates fully sparse, enabling high-resolution volumetric representations to capture fine geometric details crucial for accurate pose estimation in clutter. Our method processes the entire scene integrally, predicting the 6D pose via a novel per-voxel voting strategy, allowing simultaneous pose predictions for an arbitrary number of target objects. We validate our method on the recently published IPD and MV-YCB multi-view datasets, demonstrating competitive performance in heavily cluttered industrial and household bin picking scenarios.

Text-to-image diffusion models have demonstrated tremendous success in synthesizing visually stunning images given textual instructions. Despite remarkable progress in creating high-fidelity visuals, text-to-image models can still struggle with precisely rendering subjects, such as text spelling. To address this challenge, this paper explores using additional conditions of an image that provides visual guidance of the particular subjects for diffusion models to generate. In addition, this reference condition empowers the model to be conditioned in ways that the vocabularies of the text tokenizer cannot adequately represent, and further extends the model's generalization to novel capabilities such as generating non-English text spellings. We develop several small-scale expert plugins that efficiently endow a Stable Diffusion model with the capability to take different references. Each plugin is trained with auxiliary networks and loss functions customized for applications such as English scene-text generation, multi-lingual scene-text generation, and logo-image generation. Our expert plugins demonstrate superior results than the existing methods on all tasks, each containing only 28.55M trainable parameters.

Modern Video Large Language Models (VLLMs) often rely on uniform frame sampling for video understanding, but this approach frequently fails to capture critical information due to frame redundancy and variations in video content. We propose MaxInfo, the first training-free method based on the maximum volume principle, which is available in Fast and Slow versions and a Chunk-based version that selects and retains the most representative frames from a video. By maximizing the geometric volume formed by selected embeddings, MaxInfo ensures that the chosen frames cover the most informative regions of the embedding space, effectively reducing redundancy while preserving diversity. This method enhances the quality of input representations and improves long video comprehension performance across benchmarks. For instance, MaxInfo achieves a 3.28% improvement on LongVideoBench and a 6.4% improvement on EgoSchema for LLaVA-Video-7B. Moreover, MaxInfo boosts LongVideoBench performance by 3.47% on LLaVA-Video-72B and 3.44% on MiniCPM4.5. The approach is simple to implement and works with existing VLLMs without the need for additional training and very lower latency, making it a practical and effective alternative to traditional uniform sampling methods. Our code are available at https://github.com/FusionBrainLab/MaxInfo.git

Parameter-Efficient Fine-Tuning (PEFT) methods have transformed the approach to fine-tuning large models for downstream tasks by enabling the adjustment of significantly fewer parameters than those in the original model matrices. In this work, we study the "very low rank regime", where we fine-tune the lowest amount of parameters per linear layer for each considered PEFT method. We propose 1LoRA (Summation Low-Rank Adaptation), a compute, parameter and memory efficient fine-tuning method which uses the feature sum as fixed compression and a single trainable vector as decompression. Differently from state-of-the-art PEFT methods like LoRA, VeRA, and the recent MoRA, 1LoRA uses fewer parameters per layer, reducing the memory footprint and the computational cost. We extensively evaluate our method against state-of-the-art PEFT methods on multiple fine-tuning tasks, and show that our method not only outperforms them, but is also more parameter, memory and computationally efficient. Moreover, thanks to its memory efficiency, 1LoRA allows to fine-tune more evenly across layers, instead of focusing on specific ones (e.g. attention layers), improving performance further.

Image animation is a rapidly developing area in multimodal research, with a focus on generating videos from reference images. While much of the work has emphasized generic video generation guided by text, music-driven dance image animation remains underexplored. In this paper, we introduce MuseDance, an end-to-end model that animates reference images using both music and text inputs. By integrating music as a conditioning modality, MuseDance generates personalized videos that not only adhere to textual descriptions but also synchronize character movements with the rhythm and dynamics of the music. Unlike existing methods, MuseDance eliminates the need for explicit motion guidance, such as pose sequences or depth maps, reducing the complexity of video generation while enhancing accessibility and flexibility. To support further research in this field, we present a new multimodal dataset comprising of 3,122 dance videos, each paired with the corresponding background music and text descriptions. Our approach leverages diffusion-based methods to achieve robust generalization, precise control, and temporal consistency, setting a new benchmark for the task of music-driven image animation. The dataset of this work is available at https://github.com/Dongzhikang/musedance.

Classifier-Free Guidance (CFG) is a fundamental technique in training conditional diffusion models. The common practice for CFG-based training is to use a single network to learn both conditional and unconditional noise prediction, with a small dropout rate for conditioning. However, we observe that the joint learning of unconditional noise with limited bandwidth in training results in poor priors for the unconditional case. More importantly, these poor unconditional noise predictions become a serious reason for degrading the quality of conditional generation. Inspired by the fact that most CFG-based conditional models are trained by fine-tuning a base model with better unconditional generation, we first show that simply replacing the unconditional noise in CFG with that predicted by the base model can significantly improve conditional generation. Furthermore, we show that a diffusion model other than the one the fine-tuned model was trained on can be used for unconditional noise replacement. We experimentally verify our claim with a range of CFG-based conditional models for both image and video generation, including Zero-1-to-3, Versatile Diffusion, DiT, DynamiCrafter, and InstructPix2Pix.

Understanding 3D scenes goes beyond simply recognizing objects; it requires reasoning about the spatial and semantic relationships between them. Current 3D scene-language models often struggle with this relational understanding, particularly when visual embeddings alone do not adequately convey the roles and interactions of objects. In this paper, we introduce Descrip3D, a novel and powerful framework that explicitly encodes the relationships between objects using natural language. Unlike previous methods that rely only on 2D and 3D embeddings, Descrip3D enhances each object with a textual description that captures both its intrinsic attributes and contextual relationships. These relational cues are incorporated into the model through a dual-level integration: embedding fusion and prompt-level injection. This allows for unified reasoning across various tasks such as grounding, captioning, and question answering, all without the need for task-specific heads or additional supervision. When evaluated on five benchmark datasets, including ScanRefer, Multi3DRefer, ScanQA, SQA3D, and Scan2Cap, Descrip3D consistently outperforms strong baseline models, demonstrating the effectiveness of language-guided relational representation for understanding complex indoor scenes.

Contrast-enhanced Computed Tomography (CT) plays a vital role in modern medical diagnostics, particularly in cardiovascular assessment. However, the use of intravenous contrast agents can pose potential health risks for vulnerable patient populations. To address this limitation, we present 2S-CEDiff, a clinically-inspired image translation framework that synthesizes high-fidelity contrast-enhanced 3D CT volumes from non-contrast inputs. Our method adopts a two-stage architecture. The first stage employs a 2.5D diffusion model to generate anatomically accurate slice-wise predictions, guided by cross-attention-based positional conditioning and structural priors derived from segmentation masks produced by a pre-trained TotalSegmentator model. The second stage involves applying a 3D UNet model trained with a multi-objective loss that jointly optimizes pixel-level fidelity and volumetric coherence.We evaluate 2S-CEDiff on an in-house paired 3D cardiac CT dataset, where it achieves state-of-the-art performance in PSNR (29.87), SSIM (0.89), and inter-slice coherence. Moreover, the synthesized contrast-enhanced images significantly enhance downstream anatomical segmentation accuracy, improving the overall Dice score to 0.9447 and yielding a substantial +0.0798 increase in myocardium segmentation performance with TotalSegmentator -- underscoring their clinical utility and translational potential.

Text-to-image models have made significant strides, producing impressive results in generating images from textual descriptions. However, creating a scalable pipeline for deploying these models in production remains a challenge. Achieving the right balance between automation and human feedback is critical to maintain both scale and quality. While automation can handle large volumes, human oversight is still an essential component to ensure that the generated images meet the desired standards and are aligned with the creative vision. This paper presents a new pipeline that offers a fully automated, scalable solution for generating marketing images of commercial products using text-to-image models. The proposed system maintains the quality and fidelity of images, while also introducing sufficient creative variation to adhere to marketing guidelines. By streamlining this process, we ensure a seamless blend of efficiency and human oversight, achieving a 30.77% increase in marketing object fidelity using DINOV2 and a 52.00% increase in human preference over the generated outcome.

Linear attention is sparse, recurrent, and GPU-parallel; these are essential features for processing sparse data from event-based cameras. We argue that locality is missing to efficiently model event-to-event relationships for continuous spatiotemporal perception. We propose CoL^2\mkern-3muA by introducing locality into linear attention without using a computationally demanding convolution operation. The key idea for the convolution-free formulation is restricting the positional embedding local convolutional kernel into the special class which can be decomposed into two global positional embeddings which can be absorbed into query and key; this replaces convolution with a local sum. To the best of our knowledge, CoL^2\mkern-3muAis the first to equip sparsity, recurrence, GPU parallelism and locality, simultaneously. We demonstrate CoL^2\mkern-3muA's effectiveness on dense, high-temporal-resolution (> 1000 fps ) prediction task from events, demonstrating real-time capability while maintaining competitive results over the conventional method.

Car-crash safety assessments require experts to analyze and document numerous vehicle components from various angles, resulting in a large number of post-crash images. Currently, this process relies on manual image classification and integration into structured reports -- a time-consuming and error-prone workflow that limits scalability and consistency. In this paper, we present Crash2DocAI, a tool designed to automate the classification and integration of post-crash car part images into technical reports. Our system leverages ConvNeXt, a state-of-the-art image classification model, which achieves a top-1 accuracy of 94.4% on a newly compiled dataset of 5,772 publicly available post-crash images spanning 32 car part categories. To enable real-time deployment on CPU-only devices, we apply structured pruning and quantization, reducing the model size from 334.3MB to 77.6MB and inference time from 342ms to 94ms per image--while preserving classification performance. To enhance the robustness of our tool, we introduce an Out-of-Model-Scope (OMS) monitor based on Mahalanobis distance, which filters images outside the target domain. This binary detector achieves a precision of 71% and a recall of 95%, with only a 1% overhead on inference time. We further demonstrate the practical utility of Crash2DocAI in real-world scenarios through a user study involving 26 automotive safety experts. The results reflect a 90% speed-up and significantly more consistent completion times. Finally, we release the National Highway Traffic Safety Administration-Post-Crash Car Parts (NHTSA-PCCP) dataset to the research community, together with the application and evaluation materials. https://gitlab.com/divisvaclav/crash2docai

Background subtraction (BGS) is a fundamental task in computer vision with applications in video surveillance, object tracking, and recognition. Despite recent advancements, many deep learning-based BGS algorithms rely on large models to extract high-level representations, demanding significant computational resources and leading to inefficiencies in processing video streams. To address these limitations, we introduce the Sequential Feature Feedback Network (SeqFeedNet), a novel supervised algorithm for BGS in unseen videos that operates without additional pre-processing models. SeqFeedNet innovatively incorporates time-scale diverse sequential features and employs a feedback mechanism for each iteration. Moreover, we propose the Sequential Fit Training (SeqFiT) technique, enhancing model convergence during training. Evaluated on the CDNet 2014 dataset, SeqFeedNet not only achieves ~5 times increase in inference speed but also outperforms F-Measure scores of the leading supervised algorithms, making it highly suitable for real-world applications. Our experiment demonstrates that SeqFeedNet surpasses state-of-the-art network without pre-trained segmentation model by 3.83% F-Measure on the CDnet 2014 dataset. Leading the way to establish a new benchmark for efficient and effective BGS in unseen videos. The code is released at https://github.com/tw-yshuang/SeqFeedNet.

Large Language Models (LLMs) extended to multi-modal inputs have led to Multi-modal LLMs (MLLMs) that perform strongly on vision-language tasks. Recent MLLMs adopt multi-resolution inputs to capture both global context and local details, but this substantially increases visual tokens and computational cost. Existing pruning methods reduce redundancy but are designed for single-resolution settings, overlooking the characteristics of multi-resolution tokens. We observe two key properties: tokens from different resolutions follow distinct distributions of information content, and tokens across resolutions exhibit mutual complementarity, such that pruning one type can often be compensated by the other. Based on this observation, we propose Multi-Resolution Token Pruning method (MR-Pruner), a training-free, graph-based pruning framework for multi-resolution MLLMs. MR-Pruner incorporates three components--Intra-resolution, Cross-resolution Token Scoring, and Informativeness-aware Token Pruning--that adaptively allocate pruning ratios and facilitate information propagation across resolutions. Experiments on eight benchmarks show that MR-Pruner achieves superior efficiency-performance trade-offs. For example, when only 10% of the visual tokens are retained, it leads to an average performance degradation of 3.6%. For reproducibility, the source code is available at https://github.com/gooriiie/MR-Pruner.

Medical image segmentation is crucial for computer-aided diagnosis, surgical planning, and clinical research, requiring precise delineation of anatomical structures and pathological regions across various imaging modalities. Traditional deep learning approaches, primarily based on convolutional neural networks (CNNs) and transformers, have shown high performance but are limited by their reliance on visual features alone, which restricts their generalization and integration of clinical knowledge. The increasing availability of multimodal medical data, including paired image-text records from electronic health systems, offers a promising solution to these limitations. Vision-language segmentation (VLS) leverages natural language inputs, such as radiology reports or anatomical queries, to guide the segmentation process. This multimodal approach bridges the gap between low-level visual cues and high-level clinical concepts, reduces the need for task-specific supervision, and facilitates more intuitive human-AI interaction in medical workflows. Despite recent advancements, VLS in the medical domain faces significant challenges, including the subtlety of pathological features, inter-reader variability, and the need for fine-grained spatial accuracy. Privacy constraints and the scarcity of well-aligned, high-quality image-text pairs further complicate model training and evaluation, limiting the applicability of general-purpose VLS models in clinical settings. This paper introduces an innovative uncertainty-guided multimodal vision-language segmentation model designed to address these challenges. Our model integrates visual and textual data through advanced cross-modal learning techniques, enhancing segmentation accuracy and robustness. By incorporating uncertainty guidance, the model improves spatial precision and better captures domain-specific visual-linguistic relationships, making it more suitable for clinical applications.

In medical healthcare, obtaining detailed annotations is challenging, highlighting the need for robust Vision-Language Models (VLMs). Pretrained VLMs enable fine-tuning on small datasets or zero-shot inference, achieving performance comparable to task-specific models. On the one hand, contrastive learning (CL) is a key paradigm for training VLMs but inherently requires large batch sizes for effective learning, making it computationally demanding and often limited to well-resourced institutions. On the other hand, with limited data in healthcare, it is important to prioritize knowledge extraction from both data and models during training to improve performance. Therefore, we focus on leveraging the momentum method combined with distillation to simultaneously address computational efficiency and knowledge exploitation. Our contributions can be summarized as follows: (1) leveraging momentum self-distillation to enhance multimodal learning, and (2) integrating momentum mechanisms with gradient accumulation to enlarge the effective batch size without increasing resource consumption. Our method surpasses state-of-the-art (SOTA) approaches in few-shot learning with over 90% AUC ROC and improves retrieval tasks by 2-3% across multiple datasets. Importantly, our method achieves high training efficiency with a single GPU while maintaining reasonable training time. Our approach aims to advance efficient multimodal learning by reducing resource requirements while improving performance over SOTA methods.

We present GC-KBVQA, a zero-shot framework for knowledge-based visual question answering (KB-VQA) that requires no additional training. GC-KBVQA leverages pre-trained models together with carefully designed, context-aware descriptive information. The framework integrates three modules--(i) question-guided visual grounding, (ii) semantics-based caption filtering, and (iii) inter-stage feedback--that work together to generate concise, relevant prompts while reducing hallucinations and noisy auxiliary text. Despite its lightweight design, GC-KBVQA outperforms strong zero-shot baselines by up to +10.97% on OK-VQA, A-OKVQA, and VQAv2, and approaches the performance of few-shot systems without labeled data. The framework is model-agnostic, maintaining effectiveness across LLMs from TinyLLaMA-1B to Llama3-8B with minimal degradation. Ablation studies confirm that grounding, dual-caption generation, and both intra-stage filtering and inter-stage feedback each contribute to accuracy improvements. By combining efficiency, robustness, and modularity, GC-KBVQA provides a practical and scalable direction for zero-shot KB-VQA.

We present FlowMorph, a simple and training-free framework for geometry-preserving and semantics-aware image interpolation. The key idea is to separate two factors inside the flow model's latent space: an offset that captures shape and geometry, and a one-step vector that carries semantic meaning. By keeping the flow model frozen and only optimizing these two variables, FlowMorph exposes a stable and interpretable neighborhood around each image. This leads to two complementary modes. Flow-Optimizer directly fits a source image toward a target image and naturally supports multi-objective combinations, producing stable reconstructions. Flow-Interpolation mixes the offset linearly and the semantic vector spherically, generating smooth and coherent transitions between images. Across a wide range of tasks including object morphing, pose changes, and scene transitions, FlowMorph outperforms prior interpolation-based methods. Quantitative experiments show that our method achieves lower perceptual error, better image fidelity, and smoother transitions. Landmark-based analysis further confirms that FlowMorph preserves geometry more effectively. We also ablate the effect of the backward step size, showing that longer steps increase semantic expressiveness and allow interpolations that move beyond trivial shape blending, enabling richer morphs across object positions and photo layouts. FlowMorph provides an interpretable and controllable tool for high-quality image morphing without the need for additional training.

We introduce CLARGA, a general-purpose multimodal fusion architecture for multimodal representation learning that works with any number and type of modalities without changing the underlying framework. Given a supervised dataset, CLARGA can be applied to virtually any machine learning task to fuse different multimodal representations for processing by downstream layers. On a sample-by-sample basis, CLARGA learns how modalities should inform one another by building an attention weighted graph over their features and passing messages along this graph with a multi-head Graph Attention Network. Not only does this make CLARGA highly adaptive, as it constructs unique graphs for different samples, it makes for efficient fusion with sub-quadratic complexity as the number of modalities grows. Through a learnable mask, it can also adapt to missing modality inputs. The model is trained with a hybrid objective that combines a supervised task loss with contrastive InfoNCE loss, improving cross-modal consistency and robustness to noisy inputs. We demonstrate CLARGA's effectiveness in diverse multimodal representation learning tasks across 7 datasets spanning finance, human-computer interaction, general multimedia classification, and affective computing. It consistently outperforms baselines, state-of-the-art models, and ablations. Additional experiments also demonstrate its robustness to missing inputs and ability to excel on niche tasks. Overall, CLARGA can be easily plugged into machine learning models for effective and efficient learning of representations effective and efficient learning of representations across a wide variety of tasks.