Aharon Azulay

Nir Zabari, Aharon Azulay, Alexey Gorkor et al.

The colorization of grayscale images is a complex and subjective task with significant challenges. Despite recent progress in employing large-scale datasets with deep neural networks, difficulties with controllability and visual quality persist. To tackle these issues, we present a novel image colorization framework that utilizes image diffusion techniques with granular text prompts. This integration not only produces colorization outputs that are semantically appropriate but also greatly improves the level of control users have over the colorization process. Our method provides a balance between automation and control, outperforming existing techniques in terms of visual quality and semantic coherence. We leverage a pretrained generative Diffusion Model, and show that we can finetune it for the colorization task without losing its generative power or attention to text prompts. Moreover, we present a novel CLIP-based ranking model that evaluates color vividness, enabling automatic selection of the most suitable level of vividness based on the specific scene semantics. Our approach holds potential particularly for color enhancement and historical image colorization.

Aharon Azulay, Jan Dubiński, Zhuoyun Li et al.

The visual modality of vision-language models (VLMs) is an underexplored attack surface for bypassing safety alignment. We introduce four jailbreak attacks exploiting the vision component: (1) encoding harmful instructions as visual symbol sequences with a decoding legend, (2) replacing harmful objects with benign substitutes (e.g., bomb -> banana) then prompting for harmful actions using the substitute term, (3) replacing harmful text in images (e.g., on book covers) with benign words while visual context preserves the original meaning, and (4) visual analogy puzzles whose solution requires inferring a prohibited concept. Evaluating across six frontier VLMs, our visual attacks bypass safety alignment and expose a cross-modality alignment gap: text-based safety training does not automatically generalize to harmful intent conveyed visually. For example, our visual cipher achieves 40.9% attack success on Claude-Haiku-4.5 versus 10.7% for an equivalent textual cipher. To further our insight into the attack mechanism, we present preliminary interpretability and mitigation results. These findings highlight that robust VLM alignment requires treating vision as a first-class target for safety post-training.

Rotem Shalev-Arkushin, Aharon Azulay, Tavi Halperin et al.

Diffusion-based generative models have recently shown remarkable image and video editing capabilities. However, local video editing, particularly removal of small attributes like glasses, remains a challenge. Existing methods either alter the videos excessively, generate unrealistic artifacts, or fail to perform the requested edit consistently throughout the video. In this work, we focus on consistent and identity-preserving removal of glasses in videos, using it as a case study for consistent local attribute removal in videos. Due to the lack of paired data, we adopt a weakly supervised approach and generate synthetic imperfect data, using an adjusted pretrained diffusion model. We show that despite data imperfection, by learning from our generated data and leveraging the prior of pretrained diffusion models, our model is able to perform the desired edit consistently while preserving the original video content. Furthermore, we exemplify the generalization ability of our method to other local video editing tasks by applying it successfully to facial sticker-removal. Our approach demonstrates significant improvement over existing methods, showcasing the potential of leveraging synthetic data and strong video priors for local video editing tasks.

Aharon Azulay, Tavi Halperin, Orestis Vantzos et al.

Temporally consistent dense video annotations are scarce and hard to collect. In contrast, image segmentation datasets (and pre-trained models) are ubiquitous, and easier to label for any novel task. In this paper, we introduce a method to adapt still image segmentation models to video in an unsupervised manner, by using an optical flow-based consistency measure. To ensure that the inferred segmented videos appear more stable in practice, we verify that the consistency measure is well correlated with human judgement via a user study. Training a new multi-input multi-output decoder using this measure as a loss, together with a technique for refining current image segmentation datasets and a temporal weighted-guided filter, we observe stability improvements in the generated segmented videos with minimal loss of accuracy.

Aharon Azulay, Yair Weiss

Convolutional Neural Networks (CNNs) are commonly assumed to be invariant to small image transformations: either because of the convolutional architecture or because they were trained using data augmentation. Recently, several authors have shown that this is not the case: small translations or rescalings of the input image can drastically change the network's prediction. In this paper, we quantify this phenomena and ask why neither the convolutional architecture nor data augmentation are sufficient to achieve the desired invariance. Specifically, we show that the convolutional architecture does not give invariance since architectures ignore the classical sampling theorem, and data augmentation does not give invariance because the CNNs learn to be invariant to transformations only for images that are very similar to typical images from the training set. We discuss two possible solutions to this problem: (1) antialiasing the intermediate representations and (2) increasing data augmentation and show that they provide only a partial solution at best. Taken together, our results indicate that the problem of insuring invariance to small image transformations in neural networks while preserving high accuracy remains unsolved.