When Prompts Override Vision: Prompt-Induced Hallucinations in LVLMs

Despite impressive progress in capabilities of large vision-language models (LVLMs), these systems remain vulnerable to hallucinations, i.e., outputs that are not grounded in the visual input. Prior work has attributed hallucinations in LVLMs to factors such as limitations of the vision backbone or the dominance of the language component, yet the relative importance of these factors remains unclear. To resolve this ambiguity, We propose HalluScope, a benchmark to better understand the extent to which different factors induce hallucinations. Our analysis indicates that hallucinations largely stem from excessive reliance on textual priors and background knowledge, especially information introduced through textual instructions. To mitigate hallucinations induced by textual instruction priors, we propose HalluVL-DPO, a framework for fine-tuning off-the-shelf LVLMs towards more visually grounded responses. HalluVL-DPO leverages preference optimization using a curated training dataset that we construct, guiding the model to prefer grounded responses over hallucinated ones. We demonstrate that our optimized model effectively mitigates the targeted hallucination failure mode, while preserving or improving performance on other hallucination benchmarks and visual capability evaluations. To support reproducibility and further research, we will publicly release our evaluation benchmark, preference training dataset, and code at https://pegah-kh.github.io/projects/prompts-override-vision/ .

FlowID : Enhancing Forensic Identification with Latent Flow-Matching Models

Every day, many people die under violent circumstances, whether from crimes, war, migration, or climate disasters. Medico-legal and law enforcement institutions document many portraits of the deceased for evidence, but cannot immediately carry out identification on them. While traditional image editing tools can process these photos for public release, the workflow is lengthy and produces suboptimal results. In this work, we leverage advances in image generation models, which can now produce photorealistic human portraits, to introduce FlowID, an identity-preserving facial reconstruction method. Our approach combines single-image fine-tuning, which adapts the generative model to out-of-distribution injured faces, with attention-based masking that localizes edits to damaged regions while preserving identity-critical features. Together, these components enable the removal of artifacts from violent death while retaining sufficient identity information to support identification. To evaluate our method, we introduce InjuredFaces, a novel benchmark for identity-preserving facial reconstruction under severe facial damage. Beyond serving as an evaluation tool for this work, InjuredFaces provides a standardized resource for the community to study and compare methods addressing facial reconstruction in extreme conditions. Experimental results show that FlowID outperforms state-of-the-art open-source methods while maintaining low memory requirements, making it suitable for local deployment without compromising data privacy.

Stochastic Orthogonal Regularization for deep projective priors

Many crucial tasks of image processing and computer vision are formulated as inverse problems. Thus, it is of great importance to design fast and robust algorithms to solve these problems. In this paper, we focus on generalized projected gradient descent (GPGD) algorithms where generalized projections are realized with learned neural networks and provide state-of-the-art results for imaging inverse problems. Indeed, neural networks allow for projections onto unknown low-dimensional sets that model complex data, such as images. We call these projections deep projective priors. In generic settings, when the orthogonal projection onto a lowdimensional model set is used, it has been shown, under a restricted isometry assumption, that the corresponding orthogonal PGD converges with a linear rate, yielding near-optimal convergence (within the class of GPGD methods) in the classical case of sparse recovery. However, for deep projective priors trained with classical mean squared error losses, there is little guarantee that the hypotheses for linear convergence are satisfied. In this paper, we propose a stochastic orthogonal regularization of the training loss for deep projective priors. This regularization is motivated by our theoretical results: a sufficiently good approximation of the orthogonal projection guarantees linear stable recovery with performance close to orthogonal PGD. We show experimentally, using two different deep projective priors (based on autoencoders and on denoising networks), that our stochastic orthogonal regularization yields projections that improve convergence speed and robustness of GPGD in challenging inverse problem settings, in accordance with our theoretical findings.

SINETRA: a Versatile Framework for Evaluating Single Neuron Tracking in Behaving Animals

Accurately tracking neuronal activity in behaving animals presents significant challenges due to complex motions and background noise. The lack of annotated datasets limits the evaluation and improvement of such tracking algorithms. To address this, we developed SINETRA, a versatile simulator that generates synthetic tracking data for particles on a deformable background, closely mimicking live animal recordings. This simulator produces annotated 2D and 3D videos that reflect the intricate movements seen in behaving animals like Hydra Vulgaris. We evaluated four state-of-the-art tracking algorithms highlighting the current limitations of these methods in challenging scenarios and paving the way for improved cell tracking techniques in dynamic biological systems.

Restyling Unsupervised Concept Based Interpretable Networks with Generative Models

Developing inherently interpretable models for prediction has gained prominence in recent years. A subclass of these models, wherein the interpretable network relies on learning high-level concepts, are valued because of closeness of concept representations to human communication. However, the visualization and understanding of the learnt unsupervised dictionary of concepts encounters major limitations, specially for large-scale images. We propose here a novel method that relies on mapping the concept features to the latent space of a pretrained generative model. The use of a generative model enables high quality visualization, and naturally lays out an intuitive and interactive procedure for better interpretation of the learnt concepts. Furthermore, leveraging pretrained generative models has the additional advantage of making the training of the system more efficient. We quantitatively ascertain the efficacy of our method in terms of accuracy of the interpretable prediction network, fidelity of reconstruction, as well as faithfulness and consistency of learnt concepts. The experiments are conducted on multiple image recognition benchmarks for large-scale images. Project page available at https://jayneelparekh.github.io/VisCoIN_project_page/

A Concept-Based Explainability Framework for Large Multimodal Models

Large multimodal models (LMMs) combine unimodal encoders and large language models (LLMs) to perform multimodal tasks. Despite recent advancements towards the interpretability of these models, understanding internal representations of LMMs remains largely a mystery. In this paper, we present a novel framework for the interpretation of LMMs. We propose a dictionary learning based approach, applied to the representation of tokens. The elements of the learned dictionary correspond to our proposed concepts. We show that these concepts are well semantically grounded in both vision and text. Thus we refer to these as "multi-modal concepts". We qualitatively and quantitatively evaluate the results of the learnt concepts. We show that the extracted multimodal concepts are useful to interpret representations of test samples. Finally, we evaluate the disentanglement between different concepts and the quality of grounding concepts visually and textually. We will publicly release our code.

Diffusion-based image inpainting with internal learning

Diffusion models are now the undisputed state-of-the-art for image generation and image restoration. However, they require large amounts of computational power for training and inference. In this paper, we propose lightweight diffusion models for image inpainting that can be trained on a single image, or a few images. We show that our approach competes with large state-of-the-art models in specific cases. We also show that training a model on a single image is particularly relevant for image acquisition modality that differ from the RGB images of standard learning databases. We show results in three different contexts: texture images, line drawing images, and materials BRDF, for which we achieve state-of-the-art results in terms of realism, with a computational load that is greatly reduced compared to concurrent methods.

A Compact and Semantic Latent Space for Disentangled and Controllable Image Editing

Recent advances in the field of generative models and in particular generative adversarial networks (GANs) have lead to substantial progress for controlled image editing, especially compared with the pre-deep learning era. Despite their powerful ability to apply realistic modifications to an image, these methods often lack properties like disentanglement (the capacity to edit attributes independently). In this paper, we propose an auto-encoder which re-organizes the latent space of StyleGAN, so that each attribute which we wish to edit corresponds to an axis of the new latent space, and furthermore that the latent axes are decorrelated, encouraging disentanglement. We work in a compressed version of the latent space, using Principal Component Analysis, meaning that the parameter complexity of our autoencoder is reduced, leading to short training times ($\sim$ 45 mins). Qualitative and quantitative results demonstrate the editing capabilities of our approach, with greater disentanglement than competing methods, while maintaining fidelity to the original image with respect to identity. Our autoencoder architecture simple and straightforward, facilitating implementation.

Infusion: Internal Diffusion for Video Inpainting

Video inpainting is the task of filling a desired region in a video in a visually convincing manner. It is a very challenging task due to the high dimensionality of the signal and the temporal consistency required for obtaining convincing results. Recently, diffusion models have shown impressive results in modeling complex data distributions, including images and videos. Diffusion models remain nonetheless very expensive to train and perform inference with, which strongly restrict their application to video. We show that in the case of video inpainting, thanks to the highly auto-similar nature of videos, the training of a diffusion model can be restricted to the video to inpaint and still produce very satisfying results. This leads us to adopt an internal learning approch, which also allows for a greatly reduced network size. We call our approach "Infusion": an internal learning algorithm for video inpainting through diffusion. Due to our frugal network, we are able to propose the first video inpainting approach based purely on diffusion. Other methods require supporting elements such as optical flow estimation, which limits their performance in the case of dynamic textures for example. We introduce a new method for efficient training and inference of diffusion models in the context of internal learning. We split the diffusion process into different learning intervals which greatly simplifies the learning steps. We show qualititative and quantitative results, demonstrating that our method reaches state-of-the-art performance, in particular in the case of dynamic backgrounds and textures.

Patch-Based Stochastic Attention for Image Editing

Attention mechanisms have become of crucial importance in deep learning in recent years. These non-local operations, which are similar to traditional patch-based methods in image processing, complement local convolutions. However, computing the full attention matrix is an expensive step with a heavy memory and computational load. These limitations curb network architectures and performances, in particular for the case of high resolution images. We propose an efficient attention layer based on the stochastic algorithm PatchMatch, which is used for determining approximate nearest neighbors. We refer to our proposed layer as a "Patch-based Stochastic Attention Layer" (PSAL). Furthermore, we propose different approaches, based on patch aggregation, to ensure the differentiability of PSAL, thus allowing end-to-end training of any network containing our layer. PSAL has a small memory footprint and can therefore scale to high resolution images. It maintains this footprint without sacrificing spatial precision and globality of the nearest neighbours, which means that it can be easily inserted in any level of a deep architecture, even in shallower levels. We demonstrate the usefulness of PSAL on several image editing tasks, such as image inpainting and image colorization.