Uriel Singer

Lili Yu, Bowen Shi, Ramakanth Pasunuru et al. · berkeley, meta-ai

We present CM3Leon (pronounced "Chameleon"), a retrieval-augmented, token-based, decoder-only multi-modal language model capable of generating and infilling both text and images. CM3Leon uses the CM3 multi-modal architecture but additionally shows the extreme benefits of scaling up and tuning on more diverse instruction-style data. It is the first multi-modal model trained with a recipe adapted from text-only language models, including a large-scale retrieval-augmented pre-training stage and a second multi-task supervised fine-tuning (SFT) stage. It is also a general-purpose model that can do both text-to-image and image-to-text generation, allowing us to introduce self-contained contrastive decoding methods that produce high-quality outputs. Extensive experiments demonstrate that this recipe is highly effective for multi-modal models. CM3Leon achieves state-of-the-art performance in text-to-image generation with 5x less training compute than comparable methods (zero-shot MS-COCO FID of 4.88). After SFT, CM3Leon can also demonstrate unprecedented levels of controllability in tasks ranging from language-guided image editing to image-controlled generation and segmentation.

Shelly Sheynin, Oron Ashual, Adam Polyak et al. · meta-ai

Recent text-to-image models have achieved impressive results. However, since they require large-scale datasets of text-image pairs, it is impractical to train them on new domains where data is scarce or not labeled. In this work, we propose using large-scale retrieval methods, in particular, efficient k-Nearest-Neighbors (kNN), which offers novel capabilities: (1) training a substantially small and efficient text-to-image diffusion model without any text, (2) generating out-of-distribution images by simply swapping the retrieval database at inference time, and (3) performing text-driven local semantic manipulations while preserving object identity. To demonstrate the robustness of our method, we apply our kNN approach on two state-of-the-art diffusion backbones, and show results on several different datasets. As evaluated by human studies and automatic metrics, our method achieves state-of-the-art results compared to existing approaches that train text-to-image generation models using images only (without paired text data)

Uriel Singer, Adam Polyak, Thomas Hayes et al.

We propose Make-A-Video -- an approach for directly translating the tremendous recent progress in Text-to-Image (T2I) generation to Text-to-Video (T2V). Our intuition is simple: learn what the world looks like and how it is described from paired text-image data, and learn how the world moves from unsupervised video footage. Make-A-Video has three advantages: (1) it accelerates training of the T2V model (it does not need to learn visual and multimodal representations from scratch), (2) it does not require paired text-video data, and (3) the generated videos inherit the vastness (diversity in aesthetic, fantastical depictions, etc.) of today's image generation models. We design a simple yet effective way to build on T2I models with novel and effective spatial-temporal modules. First, we decompose the full temporal U-Net and attention tensors and approximate them in space and time. Second, we design a spatial temporal pipeline to generate high resolution and frame rate videos with a video decoder, interpolation model and two super resolution models that can enable various applications besides T2V. In all aspects, spatial and temporal resolution, faithfulness to text, and quality, Make-A-Video sets the new state-of-the-art in text-to-video generation, as determined by both qualitative and quantitative measures.

Felix Kreuk, Gabriel Synnaeve, Adam Polyak et al.

We tackle the problem of generating audio samples conditioned on descriptive text captions. In this work, we propose AaudioGen, an auto-regressive generative model that generates audio samples conditioned on text inputs. AudioGen operates on a learnt discrete audio representation. The task of text-to-audio generation poses multiple challenges. Due to the way audio travels through a medium, differentiating ``objects'' can be a difficult task (e.g., separating multiple people simultaneously speaking). This is further complicated by real-world recording conditions (e.g., background noise, reverberation, etc.). Scarce text annotations impose another constraint, limiting the ability to scale models. Finally, modeling high-fidelity audio requires encoding audio at high sampling rate, leading to extremely long sequences. To alleviate the aforementioned challenges we propose an augmentation technique that mixes different audio samples, driving the model to internally learn to separate multiple sources. We curated 10 datasets containing different types of audio and text annotations to handle the scarcity of text-audio data points. For faster inference, we explore the use of multi-stream modeling, allowing the use of shorter sequences while maintaining a similar bitrate and perceptual quality. We apply classifier-free guidance to improve adherence to text. Comparing to the evaluated baselines, AudioGen outperforms over both objective and subjective metrics. Finally, we explore the ability of the proposed method to generate audio continuation conditionally and unconditionally. Samples: https://felixkreuk.github.io/audiogen

Uriel Singer, Shelly Sheynin, Adam Polyak et al.

We present MAV3D (Make-A-Video3D), a method for generating three-dimensional dynamic scenes from text descriptions. Our approach uses a 4D dynamic Neural Radiance Field (NeRF), which is optimized for scene appearance, density, and motion consistency by querying a Text-to-Video (T2V) diffusion-based model. The dynamic video output generated from the provided text can be viewed from any camera location and angle, and can be composited into any 3D environment. MAV3D does not require any 3D or 4D data and the T2V model is trained only on Text-Image pairs and unlabeled videos. We demonstrate the effectiveness of our approach using comprehensive quantitative and qualitative experiments and show an improvement over previously established internal baselines. To the best of our knowledge, our method is the first to generate 3D dynamic scenes given a text description.

Shelly Sheynin, Adam Polyak, Uriel Singer et al.

Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present Emu Edit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop Emu Edit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance Emu Edit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for Emu Edit's outstanding performance. Furthermore, we show that Emu Edit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks.

Dan Kalifa, Uriel Singer, Kira Radinsky

Proteins play a vital role in biological processes and are indispensable for living organisms. Accurate representation of proteins is crucial, especially in drug development. Recently, there has been a notable increase in interest in utilizing machine learning and deep learning techniques for unsupervised learning of protein representations. However, these approaches often focus solely on the amino acid sequence of proteins and lack factual knowledge about proteins and their interactions, thus limiting their performance. In this study, we present GOProteinGNN, a novel architecture that enhances protein language models by integrating protein knowledge graph information during the creation of amino acid level representations. Our approach allows for the integration of information at both the individual amino acid level and the entire protein level, enabling a comprehensive and effective learning process through graph-based learning. By doing so, we can capture complex relationships and dependencies between proteins and their functional annotations, resulting in more robust and contextually enriched protein representations. Unlike previous methods, GOProteinGNN uniquely learns the entire protein knowledge graph during training, which allows it to capture broader relational nuances and dependencies beyond mere triplets as done in previous work. We perform a comprehensive evaluation on several downstream tasks demonstrating that GOProteinGNN consistently outperforms previous methods, showcasing its effectiveness and establishing it as a state-of-the-art solution for protein representation learning.

Haggai Roitman, Uriel Singer, Yotam Eshel et al.

We address the product question generation task. For a given product description, our goal is to generate questions that reflect potential user information needs that are either missing or not well covered in the description. Moreover, we wish to cover diverse user information needs that may span a multitude of product types. To this end, we first show how the T5 pre-trained Transformer encoder-decoder model can be fine-tuned for the task. Yet, while the T5 generated questions have a reasonable quality compared to the state-of-the-art method for the task (KPCNet), many of such questions are still too general, resulting in a sub-optimal global question diversity. As an alternative, we propose a novel learning-to-diversify (LTD) fine-tuning approach that allows to enrich the language learned by the underlying Transformer model. Our empirical evaluation shows that, using our approach significantly improves the global diversity of the underlying Transformer model, while preserves, as much as possible, its generation relevance.

Uriel Singer, Haggai Roitman, Ido Guy et al.

Temporal graph neural networks (temporal GNNs) have been widely researched, reaching state-of-the-art results on multiple prediction tasks. A common approach employed by most previous works is to apply a layer that aggregates information from the historical neighbors of a node. Taking a different research direction, in this work, we propose tBDFS -- a novel temporal GNN architecture. tBDFS applies a layer that efficiently aggregates information from temporal paths to a given (target) node in the graph. For each given node, the aggregation is applied in two stages: (1) A single representation is learned for each temporal path ending in that node, and (2) all path representations are aggregated into a final node representation. Overall, our goal is not to add new information to a node, but rather observe the same exact information in a new perspective. This allows our model to directly observe patterns that are path-oriented rather than neighborhood-oriented. This can be thought as a Depth-First Search (DFS) traversal over the temporal graph, compared to the popular Breath-First Search (BFS) traversal that is applied in previous works. We evaluate tBDFS over multiple link prediction tasks and show its favorable performance compared to state-of-the-art baselines. To the best of our knowledge, we are the first to apply a temporal-DFS neural network.

Heli Ben-Hamu, Omri Puny, Itai Gat et al. · meta-ai

Taming the generation outcome of state of the art Diffusion and Flow-Matching (FM) models without having to re-train a task-specific model unlocks a powerful tool for solving inverse problems, conditional generation, and controlled generation in general. In this work we introduce D-Flow, a simple framework for controlling the generation process by differentiating through the flow, optimizing for the source (noise) point. We motivate this framework by our key observation stating that for Diffusion/FM models trained with Gaussian probability paths, differentiating through the generation process projects gradient on the data manifold, implicitly injecting the prior into the optimization process. We validate our framework on linear and non-linear controlled generation problems including: image and audio inverse problems and conditional molecule generation reaching state of the art performance across all.

Hila Chefer, Uriel Singer, Amit Zohar et al.

Despite tremendous recent progress, generative video models still struggle to capture real-world motion, dynamics, and physics. We show that this limitation arises from the conventional pixel reconstruction objective, which biases models toward appearance fidelity at the expense of motion coherence. To address this, we introduce VideoJAM, a novel framework that instills an effective motion prior to video generators, by encouraging the model to learn a joint appearance-motion representation. VideoJAM is composed of two complementary units. During training, we extend the objective to predict both the generated pixels and their corresponding motion from a single learned representation. During inference, we introduce Inner-Guidance, a mechanism that steers the generation toward coherent motion by leveraging the model's own evolving motion prediction as a dynamic guidance signal. Notably, our framework can be applied to any video model with minimal adaptations, requiring no modifications to the training data or scaling of the model. VideoJAM achieves state-of-the-art performance in motion coherence, surpassing highly competitive proprietary models while also enhancing the perceived visual quality of the generations. These findings emphasize that appearance and motion can be complementary and, when effectively integrated, enhance both the visual quality and the coherence of video generation. Project website: https://hila-chefer.github.io/videojam-paper.github.io/

Dan Kalifa, Uriel Singer, Ido Guy et al.

Consumer demand forecasting is of high importance for many e-commerce applications, including supply chain optimization, advertisement placement, and delivery speed optimization. However, reliable time series sales forecasting for e-commerce is difficult, especially during periods with many anomalies, as can often happen during pandemics, abnormal weather, or sports events. Although many time series algorithms have been applied to the task, prediction during anomalies still remains a challenge. In this work, we hypothesize that leveraging external knowledge found in world events can help overcome the challenge of prediction under anomalies. We mine a large repository of 40 years of world events and their textual representations. Further, we present a novel methodology based on transformers to construct an embedding of a day based on the relations of the day's events. Those embeddings are then used to forecast future consumer behavior. We empirically evaluate the methods over a large e-commerce products sales dataset, extracted from eBay, one of the world's largest online marketplaces. We show over numerous categories that our method outperforms state-of-the-art baselines during anomalies.

Neta Shaul, Uriel Singer, Ricky T. Q. Chen et al.

This paper introduces Bespoke Non-Stationary (BNS) Solvers, a solver distillation approach to improve sample efficiency of Diffusion and Flow models. BNS solvers are based on a family of non-stationary solvers that provably subsumes existing numerical ODE solvers and consequently demonstrate considerable improvement in sample approximation (PSNR) over these baselines. Compared to model distillation, BNS solvers benefit from a tiny parameter space ($<$200 parameters), fast optimization (two orders of magnitude faster), maintain diversity of samples, and in contrast to previous solver distillation approaches nearly close the gap from standard distillation methods such as Progressive Distillation in the low-medium NFE regime. For example, BNS solver achieves 45 PSNR / 1.76 FID using 16 NFE in class-conditional ImageNet-64. We experimented with BNS solvers for conditional image generation, text-to-image generation, and text-2-audio generation showing significant improvement in sample approximation (PSNR) in all.

Neta Shaul, Uriel Singer, Itai Gat et al.

Diffusion and flow matching models have significantly advanced media generation, yet their design space is well-explored, somewhat limiting further improvements. Concurrently, autoregressive (AR) models, particularly those generating continuous tokens, have emerged as a promising direction for unifying text and media generation. This paper introduces Transition Matching (TM), a novel discrete-time, continuous-state generative paradigm that unifies and advances both diffusion/flow models and continuous AR generation. TM decomposes complex generation tasks into simpler Markov transitions, allowing for expressive non-deterministic probability transition kernels and arbitrary non-continuous supervision processes, thereby unlocking new flexible design avenues. We explore these choices through three TM variants: (i) Difference Transition Matching (DTM), which generalizes flow matching to discrete-time by directly learning transition probabilities, yielding state-of-the-art image quality and text adherence as well as improved sampling efficiency. (ii) Autoregressive Transition Matching (ARTM) and (iii) Full History Transition Matching (FHTM) are partially and fully causal models, respectively, that generalize continuous AR methods. They achieve continuous causal AR generation quality comparable to non-causal approaches and potentially enable seamless integration with existing AR text generation techniques. Notably, FHTM is the first fully causal model to match or surpass the performance of flow-based methods on text-to-image task in continuous domains. We demonstrate these contributions through a rigorous large-scale comparison of TM variants and relevant baselines, maintaining a fixed architecture, training data, and hyperparameters.

Uriel Singer, Yaron Lipman

Transition Matching (TM) is an emerging paradigm for generative modeling that generalizes diffusion and flow-matching models as well as continuous-state autoregressive models. TM, similar to previous paradigms, gradually transforms noise samples to data samples, however it uses a second ``internal'' generative model to implement the transition steps, making the transitions more expressive compared to diffusion and flow models. To make this paradigm tractable, TM employs a large backbone network and a smaller "head" module to efficiently execute the generative transition step. In this work, we present a large-scale, systematic investigation into the design, training and sampling of the head in TM frameworks, focusing on its time-continuous bidirectional variant. Through comprehensive ablations and experimentation involving training 56 different 1.7B text-to-image models (resulting in 549 unique evaluations) we evaluate the affect of the head module architecture and modeling during training as-well as a useful family of stochastic TM samplers. We analyze the impact on generation quality, training, and inference efficiency. We find that TM with an MLP head, trained with a particular time weighting and sampled with high frequency sampler provides best ranking across all metrics reaching state-of-the-art among all tested baselines, while Transformer head with sequence scaling and low frequency sampling is a runner up excelling at image aesthetics. Lastly, we believe the experiments presented highlight the design aspects that are likely to provide most quality and efficiency gains, while at the same time indicate what design choices are not likely to provide further gains.

Peter Holderrieth, Uriel Singer, Tommi Jaakkola et al.

The performance of flow matching and diffusion models can be greatly improved at inference time using reward alignment algorithms, yet efficiency remains a major limitation. While several algorithms were proposed, we demonstrate that a common bottleneck is the sampling method these algorithms rely on: many algorithms require to sample Markov transitions via SDE sampling, which is significantly less efficient and often less performant than ODE sampling. To remove this bottleneck, we introduce GLASS Flows, a new sampling paradigm that simulates a "flow matching model within a flow matching model" to sample Markov transitions. As we show in this work, this "inner" flow matching model can be retrieved from a pre-trained model without any re-training, combining the efficiency of ODEs with the stochastic evolution of SDEs. On large-scale text-to-image models, we show that GLASS Flows eliminate the trade-off between stochastic evolution and efficiency. Combined with Feynman-Kac Steering, GLASS Flows improve state-of-the-art performance in text-to-image generation, making it a simple, drop-in solution for inference-time scaling of flow and diffusion models.

Uriel Singer, Amit Zohar, Yuval Kirstain et al.

We introduce Emu Video Edit (EVE), a model that establishes a new state-of-the art in video editing without relying on any supervised video editing data. To develop EVE we separately train an image editing adapter and a video generation adapter, and attach both to the same text-to-image model. Then, to align the adapters towards video editing we introduce a new unsupervised distillation procedure, Factorized Diffusion Distillation. This procedure distills knowledge from one or more teachers simultaneously, without any supervised data. We utilize this procedure to teach EVE to edit videos by jointly distilling knowledge to (i) precisely edit each individual frame from the image editing adapter, and (ii) ensure temporal consistency among the edited frames using the video generation adapter. Finally, to demonstrate the potential of our approach in unlocking other capabilities, we align additional combinations of adapters

Yuval Kirstain, Adam Polyak, Uriel Singer et al.

The ability to collect a large dataset of human preferences from text-to-image users is usually limited to companies, making such datasets inaccessible to the public. To address this issue, we create a web app that enables text-to-image users to generate images and specify their preferences. Using this web app we build Pick-a-Pic, a large, open dataset of text-to-image prompts and real users' preferences over generated images. We leverage this dataset to train a CLIP-based scoring function, PickScore, which exhibits superhuman performance on the task of predicting human preferences. Then, we test PickScore's ability to perform model evaluation and observe that it correlates better with human rankings than other automatic evaluation metrics. Therefore, we recommend using PickScore for evaluating future text-to-image generation models, and using Pick-a-Pic prompts as a more relevant dataset than MS-COCO. Finally, we demonstrate how PickScore can enhance existing text-to-image models via ranking.

Uriel Singer, Haggai Roitman, Yotam Eshel et al.

In e-commerce, the watchlist enables users to track items over time and has emerged as a primary feature, playing an important role in users' shopping journey. Watchlist items typically have multiple attributes whose values may change over time (e.g., price, quantity). Since many users accumulate dozens of items on their watchlist, and since shopping intents change over time, recommending the top watchlist items in a given context can be valuable. In this work, we study the watchlist functionality in e-commerce and introduce a novel watchlist recommendation task. Our goal is to prioritize which watchlist items the user should pay attention to next by predicting the next items the user will click. We cast this task as a specialized sequential recommendation task and discuss its characteristics. Our proposed recommendation model, Trans2D, is built on top of the Transformer architecture, where we further suggest a novel extended attention mechanism (Attention2D) that allows to learn complex item-item, attribute-attribute and item-attribute patterns from sequential-data with multiple item attributes. Using a large-scale watchlist dataset from eBay, we evaluate our proposed model, where we demonstrate its superiority compared to multiple state-of-the-art baselines, many of which are adapted for this task.

Jonathan Kavitzky, Jonathan Zarecki, Idan Brusilovsky et al.

Recent advances in deep learning have transformed many fields by introducing generic embedding spaces, capable of achieving great predictive performance with minimal labeling effort. The geology field has not yet met such success. In this work, we introduce an extension for self-supervised learning techniques tailored for exploiting the fractal-effect in remote-sensing images. The fractal-effect assumes that the same structures (for example rivers, peaks and saddles) will appear in all scales. We demonstrate our method's effectiveness on elevation data, we also use the effect in inference. We perform an extensive analysis on several classification tasks and emphasize its effectiveness in detecting the same class on different scales. To the best of our knowledge, it is the first attempt to build a generic representation for topographic images.

Uriel Singer, Kira Radinsky

Graph neural networks (GNNs), has been widely used for supervised learning tasks in graphs reaching state-of-the-art results. However, little work was dedicated to creating unbiased GNNs, i.e., where the classification is uncorrelated with sensitive attributes, such as race or gender. Some ignore the sensitive attributes or optimize for the criteria of statistical parity for fairness. However, it has been shown that neither approaches ensure fairness, but rather cripple the utility of the prediction task. In this work, we present a GNN framework that allows optimizing representations for the notion of Equalized Odds fairness criteria. The architecture is composed of three components: (1) a GNN classifier predicting the utility class, (2) a sampler learning the distribution of the sensitive attributes of the nodes given their labels. It generates samples fed into a (3) discriminator that discriminates between true and sampled sensitive attributes using a novel "permutation loss" function. Using these components, we train a model to neglect information regarding the sensitive attribute only with respect to its label. To the best of our knowledge, we are the first to optimize GNNs for the equalized odds criteria. We evaluate our classifier over several graph datasets and sensitive attributes and show our algorithm reaches state-of-the-art results.

Uriel Singer, Ido Guy, Kira Radinsky

In this work, we present a method for node embedding in temporal graphs. We propose an algorithm that learns the evolution of a temporal graph's nodes and edges over time and incorporates this dynamics in a temporal node embedding framework for different graph prediction tasks. We present a joint loss function that creates a temporal embedding of a node by learning to combine its historical temporal embeddings, such that it optimizes per given task (e.g., link prediction). The algorithm is initialized using static node embeddings, which are then aligned over the representations of a node at different time points, and eventually adapted for the given task in a joint optimization. We evaluate the effectiveness of our approach over a variety of temporal graphs for the two fundamental tasks of temporal link prediction and multi-label node classification, comparing to competitive baselines and algorithmic alternatives. Our algorithm shows performance improvements across many of the datasets and baselines and is found particularly effective for graphs that are less cohesive, with a lower clustering coefficient.