Yuval Atzmon

Rinon Gal, Yuval Alaluf, Yuval Atzmon et al. · nvidia

Text-to-image models offer unprecedented freedom to guide creation through natural language. Yet, it is unclear how such freedom can be exercised to generate images of specific unique concepts, modify their appearance, or compose them in new roles and novel scenes. In other words, we ask: how can we use language-guided models to turn our cat into a painting, or imagine a new product based on our favorite toy? Here we present a simple approach that allows such creative freedom. Using only 3-5 images of a user-provided concept, like an object or a style, we learn to represent it through new "words" in the embedding space of a frozen text-to-image model. These "words" can be composed into natural language sentences, guiding personalized creation in an intuitive way. Notably, we find evidence that a single word embedding is sufficient for capturing unique and varied concepts. We compare our approach to a wide range of baselines, and demonstrate that it can more faithfully portray the concepts across a range of applications and tasks. Our code, data and new words will be available at: https://textual-inversion.github.io

Yoad Tewel, Yuval Atzmon, Gal Chechik et al.

Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.

Niv Cohen, Rinon Gal, Eli A. Meirom et al. · nvidia

Large Vision & Language models pretrained on web-scale data provide representations that are invaluable for numerous V&L problems. However, it is unclear how they can be used for reasoning about user-specific visual concepts in unstructured language. This problem arises in multiple domains, from personalized image retrieval to personalized interaction with smart devices. We introduce a new learning setup called Personalized Vision & Language (PerVL) with two new benchmark datasets for retrieving and segmenting user-specific "personalized" concepts "in the wild". In PerVL, one should learn personalized concepts (1) independently of the downstream task (2) allowing a pretrained model to reason about them with free language, and (3) does not require personalized negative examples. We propose an architecture for solving PerVL that operates by extending the input vocabulary of a pretrained model with new word embeddings for the new personalized concepts. The model can then reason about them by simply using them in a sentence. We demonstrate that our approach learns personalized visual concepts from a few examples and can effectively apply them in image retrieval and semantic segmentation using rich textual queries.

Rinon Gal, Moab Arar, Yuval Atzmon et al.

Text-to-image personalization aims to teach a pre-trained diffusion model to reason about novel, user provided concepts, embedding them into new scenes guided by natural language prompts. However, current personalization approaches struggle with lengthy training times, high storage requirements or loss of identity. To overcome these limitations, we propose an encoder-based domain-tuning approach. Our key insight is that by underfitting on a large set of concepts from a given domain, we can improve generalization and create a model that is more amenable to quickly adding novel concepts from the same domain. Specifically, we employ two components: First, an encoder that takes as an input a single image of a target concept from a given domain, e.g. a specific face, and learns to map it into a word-embedding representing the concept. Second, a set of regularized weight-offsets for the text-to-image model that learn how to effectively ingest additional concepts. Together, these components are used to guide the learning of unseen concepts, allowing us to personalize a model using only a single image and as few as 5 training steps - accelerating personalization from dozens of minutes to seconds, while preserving quality.

Yuval Atzmon, Maciej Bala, Yogesh Balaji et al. · nvidia

We introduce Edify Image, a family of diffusion models capable of generating photorealistic image content with pixel-perfect accuracy. Edify Image utilizes cascaded pixel-space diffusion models trained using a novel Laplacian diffusion process, in which image signals at different frequency bands are attenuated at varying rates. Edify Image supports a wide range of applications, including text-to-image synthesis, 4K upsampling, ControlNets, 360 HDR panorama generation, and finetuning for image customization.

Moab Arar, Rinon Gal, Yuval Atzmon et al.

Text-to-image (T2I) personalization allows users to guide the creative image generation process by combining their own visual concepts in natural language prompts. Recently, encoder-based techniques have emerged as a new effective approach for T2I personalization, reducing the need for multiple images and long training times. However, most existing encoders are limited to a single-class domain, which hinders their ability to handle diverse concepts. In this work, we propose a domain-agnostic method that does not require any specialized dataset or prior information about the personalized concepts. We introduce a novel contrastive-based regularization technique to maintain high fidelity to the target concept characteristics while keeping the predicted embeddings close to editable regions of the latent space, by pushing the predicted tokens toward their nearest existing CLIP tokens. Our experimental results demonstrate the effectiveness of our approach and show how the learned tokens are more semantic than tokens predicted by unregularized models. This leads to a better representation that achieves state-of-the-art performance while being more flexible than previous methods.

Dvir Samuel, Yuval Atzmon, Gal Chechik et al.

4D mesh generation has recently emerged as a powerful paradigm for recovering dynamic 3D structure from videos, but existing methods remain slow, computationally expensive, and difficult to scale to longer sequences. We introduce a training-free approach that accelerates 4D mesh generation while improving temporal correspondence quality. Our key observation is that temporal correspondences emerge inside a 4D backbone long before its generated meshes become visually accurate. We exploit this with a general framework we call Spatio-Temporal Attention Chain which propagates information across space and time. Starting from vertices on an anchor mesh, the chain maps vertices to latent tokens. It then follows temporal correspondences in latent space, and recovers frame-specific vertices through latent-to-vertex attention. This design avoids expensive explicit matching while preserving anchor mesh details and thereby improving dynamic mesh geometry and temporal consistency. Compared to state-of-the-art, our method generates a 4D mesh in 9 seconds, achieving a $13\times$ speedup while producing higher-quality results. Moreover, our approach scales to videos up to $16\times$ longer without degrading mesh quality. Beyond generation, the improved correspondences enable competitive zero-shot performance on two downstream tasks: 2D object tracking and 4D tracking. We further show that our framework enables reliable camera estimation, a capability not supported by prior 4D mesh generation methods.

Yoad Tewel, Omri Kaduri, Rinon Gal et al.

Text-to-image models offer a new level of creative flexibility by allowing users to guide the image generation process through natural language. However, using these models to consistently portray the same subject across diverse prompts remains challenging. Existing approaches fine-tune the model to teach it new words that describe specific user-provided subjects or add image conditioning to the model. These methods require lengthy per-subject optimization or large-scale pre-training. Moreover, they struggle to align generated images with text prompts and face difficulties in portraying multiple subjects. Here, we present ConsiStory, a training-free approach that enables consistent subject generation by sharing the internal activations of the pretrained model. We introduce a subject-driven shared attention block and correspondence-based feature injection to promote subject consistency between images. Additionally, we develop strategies to encourage layout diversity while maintaining subject consistency. We compare ConsiStory to a range of baselines, and demonstrate state-of-the-art performance on subject consistency and text alignment, without requiring a single optimization step. Finally, ConsiStory can naturally extend to multi-subject scenarios, and even enable training-free personalization for common objects.

Yoad Tewel, Rinon Gal, Dvir Samuel et al.

Adding Object into images based on text instructions is a challenging task in semantic image editing, requiring a balance between preserving the original scene and seamlessly integrating the new object in a fitting location. Despite extensive efforts, existing models often struggle with this balance, particularly with finding a natural location for adding an object in complex scenes. We introduce Add-it, a training-free approach that extends diffusion models' attention mechanisms to incorporate information from three key sources: the scene image, the text prompt, and the generated image itself. Our weighted extended-attention mechanism maintains structural consistency and fine details while ensuring natural object placement. Without task-specific fine-tuning, Add-it achieves state-of-the-art results on both real and generated image insertion benchmarks, including our newly constructed "Additing Affordance Benchmark" for evaluating object placement plausibility, outperforming supervised methods. Human evaluations show that Add-it is preferred in over 80% of cases, and it also demonstrates improvements in various automated metrics.

Yuval Atzmon, Rinon Gal, Yoad Tewel et al.

Text-to-video diffusion models have shown remarkable progress in generating coherent video clips from textual descriptions. However, the interplay between motion, structure, and identity representations in these models remains under-explored. Here, we investigate how self-attention query (Q) features simultaneously govern motion, structure, and identity and examine the challenges arising when these representations interact. Our analysis reveals that Q affects not only layout, but that during denoising Q also has a strong effect on subject identity, making it hard to transfer motion without the side-effect of transferring identity. Understanding this dual role enabled us to control query feature injection (Q injection) and demonstrate two applications: (1) a zero-shot motion transfer method - implemented with VideoCrafter2 and WAN 2.1 - that is 10 times more efficient than existing approaches, and (2) a training-free technique for consistent multi-shot video generation, where characters maintain identity across multiple video shots while Q injection enhances motion fidelity.

Sapir Esther Yiflach, Yuval Atzmon, Gal Chechik

Text-to-image diffusion models can generate stunning visuals, yet they often fail at tasks children find trivial--like placing a dog to the right of a teddy bear rather than to the left. When combinations get more unusual--a giraffe above an airplane--these failures become even more pronounced. Existing methods attempt to fix these spatial reasoning failures through model fine-tuning or test-time optimization with handcrafted losses that are suboptimal. Rather than imposing our assumptions about spatial encoding, we propose learning these objectives directly from the model's internal representations. We introduce Learn-to-Steer, a novel framework that learns data-driven objectives for test-time optimization rather than handcrafting them. Our key insight is to train a lightweight classifier that decodes spatial relationships from the diffusion model's cross-attention maps, then deploy this classifier as a learned loss function during inference. Training such classifiers poses a surprising challenge: they can take shortcuts by detecting linguistic traces rather than learning true spatial patterns. We solve this with a dual-inversion strategy that enforces geometric understanding. Our method dramatically improves spatial accuracy: from 0.20 to 0.61 on FLUX.1-dev and from 0.07 to 0.54 on SD2.1 across standard benchmarks. Moreover, our approach generalizes to multiple relations and significantly improves accuracy.

Ohad Rahamim, Hilit Segev, Idan Achituve et al.

Generating 3D visual scenes is at the forefront of visual generative AI, but current 3D generation techniques struggle with generating scenes with multiple high-resolution objects. Here we introduce Lay-A-Scene, which solves the task of Open-set 3D Object Arrangement, effectively arranging unseen objects. Given a set of 3D objects, the task is to find a plausible arrangement of these objects in a scene. We address this task by leveraging pre-trained text-to-image models. We personalize the model and explain how to generate images of a scene that contains multiple predefined objects without neglecting any of them. Then, we describe how to infer the 3D poses and arrangement of objects from a 2D generated image by finding a consistent projection of objects onto the 2D scene. We evaluate the quality of Lay-A-Scene using 3D objects from Objaverse and human raters and find that it often generates coherent and feasible 3D object arrangements.

Yoad Tewel, Rinon Gal, Gal Chechik et al.

Text-to-image models (T2I) offer a new level of flexibility by allowing users to guide the creative process through natural language. However, personalizing these models to align with user-provided visual concepts remains a challenging problem. The task of T2I personalization poses multiple hard challenges, such as maintaining high visual fidelity while allowing creative control, combining multiple personalized concepts in a single image, and keeping a small model size. We present Perfusion, a T2I personalization method that addresses these challenges using dynamic rank-1 updates to the underlying T2I model. Perfusion avoids overfitting by introducing a new mechanism that "locks" new concepts' cross-attention Keys to their superordinate category. Additionally, we develop a gated rank-1 approach that enables us to control the influence of a learned concept during inference time and to combine multiple concepts. This allows runtime-efficient balancing of visual-fidelity and textual-alignment with a single 100KB trained model, which is five orders of magnitude smaller than the current state of the art. Moreover, it can span different operating points across the Pareto front without additional training. Finally, we show that Perfusion outperforms strong baselines in both qualitative and quantitative terms. Importantly, key-locking leads to novel results compared to traditional approaches, allowing to portray personalized object interactions in unprecedented ways, even in one-shot settings.

Yuval Atzmon, Eli A. Meirom, Shie Mannor et al.

Reasoning and interacting with dynamic environments is a fundamental problem in AI, but it becomes extremely challenging when actions can trigger cascades of cross-dependent events. We introduce a new supervised learning setup called {\em Cascade} where an agent is shown a video of a physically simulated dynamic scene, and is asked to intervene and trigger a cascade of events, such that the system reaches a "counterfactual" goal. For instance, the agent may be asked to "Make the blue ball hit the red one, by pushing the green ball". The agent intervention is drawn from a continuous space, and cascades of events makes the dynamics highly non-linear. We combine semantic tree search with an event-driven forward model and devise an algorithm that learns to search in semantic trees in continuous spaces. We demonstrate that our approach learns to effectively follow instructions to intervene in previously unseen complex scenes. It can also reason about alternative outcomes, when provided an observed cascade of events.

Tzuf Paz-Argaman, Yuval Atzmon, Gal Chechik et al.

We study the problem of recognizing visual entities from the textual descriptions of their classes. Specifically, given birds' images with free-text descriptions of their species, we learn to classify images of previously-unseen species based on specie descriptions. This setup has been studied in the vision community under the name zero-shot learning from text, focusing on learning to transfer knowledge about visual aspects of birds from seen classes to previously-unseen ones. Here, we suggest focusing on the textual description and distilling from the description the most relevant information to effectively match visual features to the parts of the text that discuss them. Specifically, (1) we propose to leverage the similarity between species, reflected in the similarity between text descriptions of the species. (2) we derive visual summaries of the texts, i.e., extractive summaries that focus on the visual features that tend to be reflected in images. We propose a simple attention-based model augmented with the similarity and visual summaries components. Our empirical results consistently and significantly outperform the state-of-the-art on the largest benchmarks for text-based zero-shot learning, illustrating the critical importance of texts for zero-shot image-recognition.

Yuval Atzmon, Felix Kreuk, Uri Shalit et al.

People easily recognize new visual categories that are new combinations of known components. This compositional generalization capacity is critical for learning in real-world domains like vision and language because the long tail of new combinations dominates the distribution. Unfortunately, learning systems struggle with compositional generalization because they often build on features that are correlated with class labels even if they are not "essential" for the class. This leads to consistent misclassification of samples from a new distribution, like new combinations of known components. Here we describe an approach for compositional generalization that builds on causal ideas. First, we describe compositional zero-shot learning from a causal perspective, and propose to view zero-shot inference as finding "which intervention caused the image?". Second, we present a causal-inspired embedding model that learns disentangled representations of elementary components of visual objects from correlated (confounded) training data. We evaluate this approach on two datasets for predicting new combinations of attribute-object pairs: A well-controlled synthesized images dataset and a real-world dataset which consists of fine-grained types of shoes. We show improvements compared to strong baselines.

Dvir Samuel, Yuval Atzmon, Gal Chechik

Real-world data is predominantly unbalanced and long-tailed, but deep models struggle to recognize rare classes in the presence of frequent classes. Often, classes can be accompanied by side information like textual descriptions, but it is not fully clear how to use them for learning with unbalanced long-tail data. Such descriptions have been mostly used in (Generalized) Zero-shot learning (ZSL), suggesting that ZSL with class descriptions may also be useful for long-tail distributions. We describe DRAGON, a late-fusion architecture for long-tail learning with class descriptors. It learns to (1) correct the bias towards head classes on a sample-by-sample basis; and (2) fuse information from class-descriptions to improve the tail-class accuracy. We also introduce new benchmarks CUB-LT, SUN-LT, AWA-LT for long-tail learning with class-descriptions, building on existing learning-with-attributes datasets and a version of Imagenet-LT with class descriptors. DRAGON outperforms state-of-the-art models on the new benchmark. It is also a new SoTA on existing benchmarks for GFSL with class descriptors (GFSL-d) and standard (vision-only) long-tailed learning ImageNet-LT, CIFAR-10, 100, and Places365.

Gilad Vered, Gal Oren, Yuval Atzmon et al.

When describing images with natural language, the descriptions can be made more informative if tuned using downstream tasks. This is often achieved by training two networks: a "speaker network" that generates sentences given an image, and a "listener network" that uses them to perform a task. Unfortunately, training multiple networks jointly to communicate to achieve a joint task, faces two major challenges. First, the descriptions generated by a speaker network are discrete and stochastic, making optimization very hard and inefficient. Second, joint training usually causes the vocabulary used during communication to drift and diverge from natural language. We describe an approach that addresses both challenges. We first develop a new effective optimization based on partial-sampling from a multinomial distribution combined with straight-through gradient updates, which we name PSST for Partial-Sampling Straight-Through. Second, we show that the generated descriptions can be kept close to natural by constraining them to be similar to human descriptions. Together, this approach creates descriptions that are both more discriminative and more natural than previous approaches. Evaluations on the standard COCO benchmark show that PSST Multinomial dramatically improve the recall@10 from 60% to 86% maintaining comparable language naturalness, and human evaluations show that it also increases naturalness while keeping the discriminative power of generated captions.

Roman Visotsky, Yuval Atzmon, Gal Chechik

Learning with few samples is a major challenge for parameter-rich models like deep networks. In contrast, people learn complex new concepts even from very few examples, suggesting that the sample complexity of learning can often be reduced. Many approaches to few-shot learning build on transferring a representation from well-sampled classes, or using meta learning to favor architectures that can learn with few samples. Unfortunately, such approaches often struggle when learning in an online way or with non-stationary data streams. Here we describe a new approach to learn with fewer samples, by using additional information that is provided per sample. Specifically, we show how the sample complexity can be reduced by providing semantic information about the relevance of features per sample, like information about the presence of objects in a scene or confidence of detecting attributes in an image. We provide an improved generalization error bound for this case. We cast the problem of using per-sample feature relevance by using a new ellipsoid-margin loss, and develop an online algorithm that minimizes this loss effectively. Empirical evaluation on two machine vision benchmarks for scene classification and fine-grain bird classification demonstrate the benefits of this approach for few-shot learning.

Yuval Atzmon, Gal Chechik

Generalized zero-shot learning (GZSL) is the problem of learning a classifier where some classes have samples and others are learned from side information, like semantic attributes or text description, in a zero-shot learning fashion (ZSL). Training a single model that operates in these two regimes simultaneously is challenging. Here we describe a probabilistic approach that breaks the model into three modular components, and then combines them in a consistent way. Specifically, our model consists of three classifiers: A "gating" model that makes soft decisions if a sample is from a "seen" class, and two experts: a ZSL expert, and an expert model for seen classes. We address two main difficulties in this approach: How to provide an accurate estimate of the gating probability without any training samples for unseen classes; and how to use expert predictions when it observes samples outside of its domain. The key insight to our approach is to pass information between the three models to improve each one's accuracy, while maintaining the modular structure. We test our approach, adaptive confidence smoothing (COSMO), on four standard GZSL benchmark datasets and find that it largely outperforms state-of-the-art GZSL models. COSMO is also the first model that closes the gap and surpasses the performance of generative models for GZSL, even-though it is a light-weight model that is much easier to train and tune. Notably, COSMO offers a new view for developing zero-shot models. Thanks to COSMO's modular structure, instead of trying to perform well both on seen and on unseen classes, models can focus on accurate classification of unseen classes, and later consider seen class models.

Yuval Atzmon, Gal Chechik

In zero-shot learning (ZSL), a classifier is trained to recognize visual classes without any image samples. Instead, it is given semantic information about the class, like a textual description or a set of attributes. Learning from attributes could benefit from explicitly modeling structure of the attribute space. Unfortunately, learning of general structure from empirical samples is hard with typical dataset sizes. Here we describe LAGO, a probabilistic model designed to capture natural soft and-or relations across groups of attributes. We show how this model can be learned end-to-end with a deep attribute-detection model. The soft group structure can be learned from data jointly as part of the model, and can also readily incorporate prior knowledge about groups if available. The soft and-or structure succeeds to capture meaningful and predictive structures, improving the accuracy of zero-shot learning on two of three benchmarks. Finally, LAGO reveals a unified formulation over two ZSL approaches: DAP (Lampert et al., 2009) and ESZSL (Romera-Paredes & Torr, 2015). Interestingly, taking only one singleton group for each attribute, introduces a new soft-relaxation of DAP, that outperforms DAP by ~40.

Yuval Atzmon, Jonathan Berant, Vahid Kezami et al.

Recurrent neural networks have recently been used for learning to describe images using natural language. However, it has been observed that these models generalize poorly to scenes that were not observed during training, possibly depending too strongly on the statistics of the text in the training data. Here we propose to describe images using short structured representations, aiming to capture the crux of a description. These structured representations allow us to tease-out and evaluate separately two types of generalization: standard generalization to new images with similar scenes, and generalization to new combinations of known entities. We compare two learning approaches on the MS-COCO dataset: a state-of-the-art recurrent network based on an LSTM (Show, Attend and Tell), and a simple structured prediction model on top of a deep network. We find that the structured model generalizes to new compositions substantially better than the LSTM, ~7 times the accuracy of predicting structured representations. By providing a concrete method to quantify generalization for unseen combinations, we argue that structured representations and compositional splits are a useful benchmark for image captioning, and advocate compositional models that capture linguistic and visual structure.