We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.

Recent advancements in diffusion models have set new benchmarks in image and video generation, enabling realistic visual synthesis across single- and multi-frame contexts. However, these models still struggle with efficiently and explicitly generating 3D-consistent content. To address this, we propose World-consistent Video Diffusion (WVD), a novel framework that incorporates explicit 3D supervision using XYZ images, which encode global 3D coordinates for each image pixel. More specifically, we train a diffusion transformer to learn the joint distribution of RGB and XYZ frames. This approach supports multi-task adaptability via a flexible inpainting strategy. For example, WVD can estimate XYZ frames from ground-truth RGB or generate novel RGB frames using XYZ projections along a specified camera trajectory. In doing so, WVD unifies tasks like singleimage-to-3D generation, multi-view stereo, and camera-controlled video generation. Our approach demonstrates competitive performance across multiple benchmarks, providing a scalable solution for 3D-consistent video and image generation with a single pretrained model. Our project website is at https://zqh0253.github.io/wvd.

Building general-purpose models that can effectively perceive the world through multimodal signals has been a long-standing goal. Current approaches involve integrating separately pre-trained components, such as connecting vision encoders to LLMs and continuing multimodal training. While such approaches exhibit remarkable sample efficiency, it remains an open question whether such late-fusion architectures are inherently superior. In this work, we revisit the architectural design of native multimodal models (NMMs)-those trained from the ground up on all modalities-and conduct an extensive scaling laws study, spanning 457 trained models with different architectures and training mixtures. Our investigation reveals no inherent advantage to late-fusion architectures over early-fusion ones, which do not rely on image encoders or tokenizers. On the contrary, early-fusion exhibits stronger performance at lower parameter counts, is more efficient to train, and is easier to deploy. Motivated by the strong performance of the early-fusion architectures, we show that incorporating Mixture of Experts (MoEs) allows models to learn modality-specific weights, significantly benefiting performance.

Diffusion models have become the dominant approach for visual generation. They are trained by denoising a Markovian process which gradually adds noise to the input. We argue that the Markovian property limits the model’s ability to fully utilize the generation trajectory, leading to inefficiencies during training and inference. In this paper, we propose DART, a transformer-based model that unifies autoregressive (AR) and diffusion within a non-Markovian framework. DART iteratively denoises image patches spatially and spectrally using an AR model that has the same architecture as standard language models. DART does not rely on image quantization, which enables more effective image modeling while maintaining flexibility. Furthermore, DART seamlessly trains with both text and image data in a unified model. Our approach demonstrates competitive performance on class-conditioned and text-to-image generation tasks, offering a scalable, efficient alternative to traditional diffusion models. Through this unified framework, DART sets a new benchmark for scalable, high-quality image synthesis.

Vision foundation models pre-trained on massive data encode rich representations of real-world concepts, which can be adapted to downstream tasks by fine-tuning. However, fine-tuning foundation models on one task often leads to the issue of *concept forgetting* on other tasks. Recent methods of robust fine-tuning aim to mitigate forgetting of prior knowledge without affecting the fine-tuning performance. Knowledge is often preserved by matching the original and fine-tuned model weights or feature pairs. However, such point-wise matching can be too strong, without explicit awareness of the feature neighborhood structures that encode rich knowledge as well. We propose a novel regularization method **Proxy-FDA** that explicitly preserves the structural knowledge in feature space. Proxy-FDA performs Feature Distribution Alignment (using nearest neighbor graphs) between the pre-trained and fine-tuned feature spaces, and the alignment is further improved by informative proxies that are generated dynamically to increase data diversity. Experiments show that Proxy-FDA significantly reduces concept forgetting during fine-tuning, and we find a strong correlation between forgetting and a distributional distance metric (in comparison to L2 distance). We further demonstrate Proxy-FDA's benefits in various fine-tuning settings (end-to-end, few-shot and continual tuning) and across different tasks like image classification, captioning and VQA.

Scaling the capacity of language models has consistently proven to be a reliable approach for improving performance and unlocking new capabilities. Capacity can be primarily defined by two dimensions: the number of model parameters and the compute per example. While scaling typically involves increasing both, the precise interplay between these factors and their combined contribution to overall capacity remains not fully understood. We explore this relationship in the context of sparse Mixture-of-Expert models (MoEs), which allow scaling the number of parameters without proportionally increasing the FLOPs per example. We investigate how varying the sparsity level, i.e., the ratio of non-active to total parameters, affects model performance in terms of both pretraining and downstream performance. We find that under different constraints (e.g. parameter size and total training compute), there is an optimal level of sparsity that improves both training efficiency and model performance.These results provide a better understanding of the impact of sparsity in scaling laws for MoEs and complement existing works in this area, offering insights for designing more efficient architectures.

We study the theoretical foundations of composition in diffusion models, with a particular focus on out-of-distribution extrapolation and length-generalization. Prior work has shown that composing distributions via linear score combination can achieve promising results, including length-generalization in some cases (Du et al., 2023; Liu et al., 2022). However, our theoretical understanding of how and why such compositions work remains incomplete. In fact, it is not even entirely clear what it means for composition to "work". This paper starts to address these fundamental gaps. We begin by precisely defining one possible desired result of composition, which we call *projective composition*. Then, we investigate: (1) when linear score combinations provably achieve projective composition, (2) whether reverse-diffusion sampling can generate the desired composition, and (3) the conditions under which composition fails. We connect our theoretical analysis to prior empirical observations where composition has either worked or failed, for reasons that were unclear at the time. Finally, we propose a simple heuristic to help predict the success or failure of new compositions.

Flow matching models have emerged as a powerful method for generative modeling on domains like images or videos, and even on irregular or unstructured data like 3D point clouds or even protein structures. These models are commonly trained in two stages: first, a data compressor is trained, and in a subsequent training stage a flow matching generative model is trained in the latent space of the data compressor. This two-stage paradigm sets obstacles for unifying models across data domains, as hand-crafted compressors architectures are used for different data modalities. To this end, we introduce INRFlow, a domain-agnostic approach to learn flow matching transformers directly in ambient space. Drawing inspiration from INRs, we introduce a conditionally independent point-wise training objective that enables INRFlow to make predictions continuously in coordinate space. Our empirical results demonstrate that INRFlow effectively handles different data modalities such as images, 3D point clouds and protein structure data, achieving strong performance in different domains and outperforming comparable approaches. INRFlow is a promising step towards domain-agnostic flow matching generative models that can be trivially adopted in different data domains.

Normalizing Flows (NFs) are likelihood-based models for continuous inputs. They have demonstrated promising results on both density estimation and generative modeling tasks, but have received relatively little attention in recent years. In this work, we demonstrate that NFs are more powerful than previously believed. We present TarFlow: a simple and scalable architecture that enables highly performant NF models. TarFlow can be thought of as a Transformer-based variant of Masked Autoregressive Flows (MAFs): it consists of a stack of autoregressive Transformer blocks on image patches, alternating the autoregression direction between layers. TarFlow is straightforward to train end-to-end, and capable of directly modeling and generating pixels. We also propose three key techniques to improve sample quality: Gaussian noise augmentation during training, a post training denoising procedure, and an effective guidance method for both class-conditional and unconditional settings. Putting these together, TarFlow sets new state-of-the-art results on likelihood estimation for images, beating the previous best methods by a large margin, and generates samples with quality and diversity comparable to diffusion models, for the first time with a stand-alone NF model. We make our code available at https://github.com/apple/ml-tarflow.

Discrete diffusion is a promising framework for modeling and generating discrete data. In this work, we present Target Concrete Score Matching (TCSM), a novel and versatile objective for training and fine-tuning discrete diffusion models. TCSM provides a general framework with broad applicability. It supports pre-training discrete diffusion models directly from data samples, and many existing discrete diffusion approaches naturally emerge as special cases of our more general TCSM framework. Furthermore, the same TCSM objective extends to post-training of discrete diffusion models, including fine-tuning using reward functions or preference data, and distillation of knowledge from pre-trained autoregressive models. These new capabilities stem from the core idea of TCSM, estimating the concrete score of the target distribution, which resides in the original (clean) data space. This allows seamless integration with reward functions and pre-trained models, which inherently only operate in the clean data space rather than the noisy intermediate spaces of diffusion processes. Our experiments on language modeling tasks demonstrate that TCSM matches or surpasses current methods. Additionally, TCSM is versatile, applicable to both pre-training and post-training scenarios, offering greater flexibility and sample efficiency.

Diffusion models have demonstrated exceptional capabilities in generating high-fidelity images but typically suffer from inefficient sampling. Many solver designs and noise scheduling strategies have been proposed to dramatically improve sampling speeds. In this paper, we introduce a new sampling method that is up to 186\% faster than the current state of the art solver for comparative FID on ImageNet512. This new sampling method is training-free and uses an ordinary differential equation (ODE) solver. The key to our method resides in using higher-dimensional initial noise, allowing to produce more detailed samples with less function evaluations from existing pretrained diffusion models. In addition, by design our solver allows to control the level of detail through a simple hyper-parameter at no extra computational cost. We present how our approach leverages momentum dynamics by establishing a fundamental equivalence between momentum diffusion models and conventional diffusion models with respect to their training paradigms. Moreover, we observe the use of higher-dimensional noise naturally exhibits characteristics similar to stochastic differential equations (SDEs). Finally, we demonstrate strong performances on a set of representative pretrained diffusion models, including EDM, EDM2, and Stable-Diffusion 3, which cover models in both pixel and latent spaces, as well as class and text conditional settings. The code is available at https://github.com/apple/ml-tada.

We present STARFlow, a scalable generative model based on normalizing flows that achieves strong performance on high-resolution image synthesis. STARFlow's main building block is Transformer Autoregressive Flow (TARFlow), which combines normalizing flows with Autoregressive Transformer architectures and has recently achieved impressive results in image modeling. In this work, we first establish the theoretical universality of TARFlow for modeling continuous distributions. Building on this foundation, we introduce a set of architectural and algorithmic innovations that significantly enhance the scalability: (1) a deep-shallow design where a deep Transformer block captures most of the model’s capacity, followed by a few shallow Transformer blocks that are computationally cheap yet contribute non-negligibly, (2) learning in the latent space of pretrained autoencoders, which proves far more effective than modeling pixels directly, and (3) a novel guidance algorithm that substantially improves sample quality. Crucially, our model remains a single, end-to-end normalizing flow, allowing exact maximum likelihood training in continuous space without discretization. STARFlow achieves competitive results in both class- and text-conditional image generation, with sample quality approaching that of state-of-the-art diffusion models. To our knowledge, this is the **first** successful demonstration of normalizing flows at this scale and resolution. Code and weights available at https://github.com/apple/ml-starflow.

Autoregressive models have driven remarkable progress in language modeling. Their foundational reliance on discrete tokens, unidirectional context, and single-pass decoding, while central to their success, also inspires the exploration of a design space that could offer new axes of modeling flexibility. In this work, we explore an alternative paradigm, shifting language modeling from a discrete token space to a continuous latent space. We propose a novel framework that employs transformer-based autoregressive normalizing flows to model these continuous representations. This approach unlocks substantial flexibility, enabling the construction of models that can capture global bi-directional context through stacked, alternating-direction autoregressive transformations, support block-wise generation with flexible token patch sizes, and facilitate a hierarchical multi-pass generation process. We further propose new mixture-based coupling transformations designed to capture complex dependencies within the latent space shaped by discrete data, and demonstrate theoretical connections to conventional discrete autoregressive models. Extensive experiments on language modeling benchmarks demonstrate strong likelihood performance and highlight the flexible modeling capabilities inherent in our framework.

Diffusion models have recently become the de-facto approach for generative modeling in the 2D domain. However, extending diffusion models to 3D is challenging, due to the difficulties in acquiring 3D ground truth data for training. On the other hand, 3D GANs that integrate implicit 3D representations into GANs have shown remarkable 3D-aware generation when trained only on single-view image datasets. However, 3D GANs do not provide straightforward ways to precisely control image synthesis. To address these challenges, We present Control3Diff, a 3D diffusion model that combines the strengths of diffusion models and 3D GANs for versatile controllable 3D-aware image synthesis for single-view datasets. Control3Diff explicitly models the underlying latent distribution (optionally conditioned on external inputs), thus enabling direct control during the diffusion process. Moreover, our approach is general and applicable to any types of controlling inputs, allowing us to train it with the same diffusion objective without any auxiliary supervision. We validate the efficacy of Control3Diff on standard image generation benchmarks including FFHQ, AFHQ, and ShapeNet, using various conditioning inputs such as images, sketches, and text prompts.

Datasets that pair Knowledge Graphs (KG) and text together (KG-T) can be used to train forward and reverse neural models that generate text from KG and vice versa. However models trained on datasets where KG and text pairs are not equivalent can suffer from more hallucination and poorer recall. In this paper, we verify this empirically by generating datasets with different levels of noise and find that noisier datasets do indeed lead to more hallucination. We argue that the ability of forward and reverse models trained on a dataset to cyclically regenerate source KG or text is a proxy for the equivalence between the KG and the text in the dataset. Using cyclic evaluation we find that manually created WebNLG is much better than automatically created TeKGen and T-REx. Informed by these observations, we construct a new, improved dataset called LAGRANGE using heuristics meant to improve equivalence between KG and text and show the impact of each of the heuristics on cyclic evaluation. We also construct two synthetic datasets using large language models (LLMs), and observe that these are conducive to models that perform significantly well on cyclic generation of text, but less so on cyclic generation of KGs, probably because of a lack of a consistent underlying ontology.

Rendering scenes observed in a monocular video from novel viewpoints is a challenging problem. For static scenes the community has studied both scene-specific optimization techniques, which optimize on every test scene, and generalized techniques, which only run a deep net forward pass on a test scene. In contrast, for dynamic scenes, scene-specific optimization techniques exist, but, to our best knowledge, there is currently no generalized method for dynamic novel view synthesis from a given monocular video. To explore whether generalized dynamic novel view synthesis from monocular videos is possible today, we establish an analysis framework based on existing techniques and work toward the generalized approach. We find a pseudo-generalized process without scene-specific \emph{appearance} optimization is possible, but geometrically and temporally consistent depth estimates are needed. Despite no scene-specific appearance optimization, the pseudo-generalized approach improves upon some scene-specific methods.For more information see project page at https://xiaoming-zhao.github.io/projects/pgdvs.

We investigate the capabilities of transformer models on relational reasoning tasks. In these tasks, models are trained on a set of strings encoding abstract relations, and are then tested out-of-distribution on data that contains symbols that did not appear in the training dataset. We prove that for any relational reasoning task in a large family of tasks, transformers learn the abstract relations and generalize to the test set when trained by gradient descent on sufficiently large quantities of training data. This is in contrast to classical fully-connected networks, which we prove fail to learn to reason. Our results inspire modifications of the transformer architecture that add only two trainable parameters per head, and that we empirically demonstrate improve data efficiency for learning to reason.

Diffusion models (DMs) represent state-of-the-art generative models for continuous inputs. DMs work by constructing a Stochastic Differential Equation (SDE) in the input space (ie, position space), and using a neural network to reverse it. In this work, we introduce a novel generative modeling framework grounded in \textbf{phase space dynamics}, where a phase space is defined as {an augmented space encompassing both position and velocity.} Leveraging insights from Stochastic Optimal Control, we construct a path measure in the phase space that enables efficient sampling. {In contrast to DMs, our framework demonstrates the capability to generate realistic data points at an early stage of dynamics propagation.} This early prediction sets the stage for efficient data generation by leveraging additional velocity information along the trajectory. On standard image generation benchmarks, our model yields favorable performance over baselines in the regime of small Number of Function Evaluations (NFEs). Furthermore, our approach rivals the performance of diffusion models equipped with efficient sampling techniques, underscoring its potential as a new tool generative modeling.

We present Manifold Diffusion Fields (MDF), an approach that unlocks learning of diffusion models of data in general non-euclidean geometries. Leveraging insights from spectral geometry analysis, we define an intrinsic coordinate system on the manifold via the eigen-functions of the Laplace-Beltrami Operator. MDF represents functions using an explicit parametrization formed by a set of multiple input-output pairs. Our approach allows to sample continuous functions on manifolds and is invariant with respect to rigid and isometric transformations of the manifold. In addition, we show that MDF generalizes to the case where the training set contains functions on different manifolds. Empirical results on multiple datasets and manifolds including challenging scientific problems like weather prediction or molecular conformation show that MDF can capture distributions of such functions with better diversity and fidelity than previous approaches.

Diffusion models are the de-facto approach for generating high-quality images and videos, but learning high-dimensional models remains a formidable task due to computational and optimization challenges. Existing methods often resort to training cascaded models in pixel space, or using a downsampled latent space of a separately trained auto-encoder. In this paper, we introduce Matryoshka Diffusion (MDM), an end-to-end framework for high-resolution image and video synthesis. We propose a diffusion process that denoises inputs at multiple resolutions jointly and uses a NestedUNet architecture where features and parameters for small-scale inputs are nested within those of large scales. In addition, MDM enables a progressive training schedule from lower to higher resolutions, which leads to significant improvements in optimization for high-resolution generation. We demonstrate the effectiveness of our approach on various benchmarks, including class-conditioned image generation, high-resolution text-to-image, and text-to-video applications. Remarkably, we can train a single pixel-space model at resolutions of up to 1024x1024 pixels, demonstrating strong zero-shot generalization using the CC12M dataset, which contains only 12 million images. Code and pre-trained checkpoints are released at https://github.com/apple/ml-mdm.

Pretrained language models are commonly aligned with human preferences and downstream tasks via reinforcement finetuning (RFT), which refers to maximizing a (possibly learned) reward function using policy gradient algorithms. This work identifies a fundamental optimization obstacle in RFT: we prove that the expected gradient for an input vanishes when its reward standard deviation under the model is small, even if the expected reward is far from optimal. Through experiments on an RFT benchmark and controlled environments, as well as a theoretical analysis, we then demonstrate that vanishing gradients due to small reward standard deviation are prevalent and detrimental, leading to extremely slow reward maximization. Lastly, we explore ways to overcome vanishing gradients in RFT. We find the common practice of an initial supervised finetuning (SFT) phase to be the most promising candidate, which sheds light on its importance in an RFT pipeline. Moreover, we show that a relatively small number of SFT optimization steps on as few as 1% of the input samples can suffice, indicating that the initial SFT phase need not be expensive in terms of compute and data labeling efforts. Overall, our results emphasize that being mindful for inputs whose expected gradient vanishes, as measured by the reward standard deviation, is crucial for successful execution of RFT.

Joint embedding (JE) architectures have emerged as a promising avenue for ac- quiring transferable data representations. A key obstacle to using JE methods, however, is the inherent challenge of evaluating learned representations without access to a downstream task, and an annotated dataset. Without efficient and re- liable evaluation, it is difficult to iterate on architectural and training choices for JE methods. In this paper, we introduce LiDAR (Linear Discriminant Analysis Rank), a metric designed to measure the quality of representations within JE archi- tectures. Our metric addresses several shortcomings of recent approaches based on feature covariance rank by discriminating between informative and uninforma- tive features. In essence, LiDAR quantifies the rank of the Linear Discriminant Analysis (LDA) matrix associated with the surrogate SSL task—a measure that intuitively captures the information content as it pertains to solving the SSL task. We empirically demonstrate that LiDAR significantly surpasses naive rank based approaches in its predictive power of optimal hyperparameters. Our proposed cri- terion presents a more robust and intuitive means of assessing the quality of rep- resentations within JE architectures, which we hope facilitates broader adoption of these powerful techniques in various domains.

Existing vision-language models exhibit strong generalization on a variety of visual domains and tasks. However, such models mainly perform zero-shot recognition in a closed-set manner, and thus struggle to handle open-domain visual concepts by design. There are recent finetuning methods, such as prompt learning, that not only study the discrimination between in-distribution (ID) and out-of-distribution (OOD) samples, but also show some improvements in both ID and OOD accuracies. In this paper, we first demonstrate that vision-language models, after long enough finetuning but without proper regularization, tend to overfit the known classes in the given dataset, with degraded performance on unknown classes. Then we propose a novel approach OGEN to address this pitfall, with the main focus on improving the OOD GENeralization of finetuned models. Specifically, a class-conditional feature generator is introduced to synthesize OOD features using just the class name of any unknown class. Such synthesized features will provide useful knowledge about unknowns and help regularize the decision boundary between ID and OOD data when optimized jointly. Equally important is our adaptive self-distillation mechanism to regularize our feature generation model during joint optimization, i.e., adaptively transferring knowledge between model states to further prevent overfitting. Experiments validate that our method yields convincing gains in OOD generalization performance in different settings. Code: https://github.com/apple/ml-ogen.

Large language models exhibit surprising emergent generalization properties, yet also struggle on many simple reasoning tasks such as arithmetic and parity. In this work, we focus on length generalization, and we propose a unifying framework to understand when and how Transformers can be expected to length generalize on a given task. First, we show that there exist algorithmic tasks for which standard decoder-only Transformers trained from scratch naturally exhibit strong length generalization. For these tasks, we leverage the RASP programming language (Weiss et al., 2021) to show that the correct algorithmic solution which solves the task can be represented by a simple Transformer. We thus propose the RASP-Generalization Conjecture: Transformers tend to learn a length-generalizing solution if there exists a short RASP-L program that works for all input lengths. We present empirical evidence to support the correlation between RASP-simplicity and generalization. We leverage our insights to give new scratchpad formats which yield strong length generalization on traditionally hard tasks (such as parity and addition), and we illustrate how scratchpad can hinder generalization when it increases the complexity of the corresponding RASP-L program. Overall, our work provides a novel perspective on the mechanisms of length generalization and the algorithmic capabilities of Transformers.

This paper introduces AIM, a collection of vision models pre-trained with an autoregressive objective. These models are inspired by their textual counterparts, i.e., Large Language Models (LLMs), and exhibit similar scaling properties. Specifically, we highlight two key findings: (1) the performance of the visual features scale with both the model capacity and the quantity of data, (2) the value of the objective function correlates with the performance of the model on downstream tasks. We illustrate the practical implication of these findings by pre-training a 7 billion parameter AIM on 2 billion images, that achieves 84.0% on ImageNet-1k with a frozen trunk. Interestingly, even at this scale, we observe no sign of saturation in performance, suggesting that AIM potentially represents a new frontier for training large-scale vision models. The pre-training of AIM is similar to the pre-training of LLMs, and does not require any image-specific strategy to stabilize the training at scale.

Diffusion models have demonstrated great potential for generating diverse images. However, their performance often suffers from slow generation due to iterative denoising. Knowledge distillation has been recently proposed as a remedy which can reduce the number of inference steps to one or a few, without significant quality degradation. However, existing distillation methods either require significant amounts of offline computation for generating synthetic training data from the teacher model, or need to perform expensive online learning with the help of real data. In this work, we present a novel technique called BOOT, that overcomes these limitations with an efficient data-free distillation algorithm. The core idea is to learn a time-conditioned model that predicts the output of a pre-trained diffusion model teacher given any time-step. Such a model can be efficiently trained based on bootstrapping from two consecutive sampled steps. Furthermore, our method can be easily adapted to large-scale text-to-image diffusion models, which are challenging for previous methods given the fact that the training sets are often large and difficult to access. We demonstrate the effectiveness of our approach on several benchmark datasets in the DDIM setting, achieving comparable generation quality while being orders of magnitude faster than the diffusion teacher. The text-to-image results show that the proposed approach is able to handle highly complex distributions, shedding light on more efficient generative modeling.

We present a novel way to predict molecular conformers through a simple formulation that sidesteps many of the heuristics of prior works and achieves state of the art results by using the advantages of scale. By training a diffusion generative model directly on 3D atomic positions without making assumptions about the explicit structure of molecules (e.g. modeling torsional angles) we are able to radically simplify structure learning, and make it trivial to scale up the model sizes. This model, called Molecular Conformer Fields (MCF), works by parameterizing conformer structures as functions that map elements from a molecular graph directly to their 3D location in space. This formulation allows us to boil down the essence of structure prediction to learning a distribution over functions. Experimental results show that scaling up the model capacity leads to large gains in generalization performance without enforcing inductive biases like rotational equivariance. MCF represents an advance in extending diffusion models to handle complex scientific problems in a conceptually simple, scalable and effective manner.

Large pretrained vision-language models like CLIP have shown promising generalization capability, but may struggle in specialized domains (e.g., satellite imagery) or fine-grained classification (e.g., car models) where the visual concepts are unseen or under-represented during pretraining. Prompt learning offers a parameter-efficient finetuning framework that can adapt CLIP to downstream tasks even when limited annotation data are available. In this paper, we improve prompt learning by distilling the textual knowledge from natural language prompts (either human- or LLM-generated) to provide rich priors for those under-represented concepts. We first obtain a prompt ``summary'' aligned to each input image via a learned prompt aggregator. Then we jointly train a prompt generator, optimized to produce a prompt embedding that stays close to the aggregated summary while minimizing task loss at the same time. We dub such prompt embedding as Aggregate-and-Adapted Prompt Embedding (AAPE). AAPE is shown to be able to generalize to different downstream data distributions and tasks, including vision-language understanding tasks (e.g., few-shot classification, VQA) and generation tasks (image captioning) where AAPE achieves competitive performance. We also show AAPE is particularly helpful to handle non-canonical and OOD examples. Furthermore, AAPE learning eliminates LLM-based inference cost as required by baselines, and scales better with data and LLM model size.

Diffusion models have emerged as a powerful tool for generating high-quality images from textual descriptions. Despite their successes, these models often exhibit limited diversity in the sampled images, particularly when sampling with a high classifier-free guidance weight. To address this issue, we present Kaleido, a novel approach that enhances the diversity of samples by incorporating autoregressive latent priors. Kaleido integrates an autoregressive language model that encodes the original caption and generates latent variables, serving as abstract and intermediary representations for guiding and facilitating the image generation process. In this paper, we explore a variety of discrete latent representations, including textual descriptions, detection bounding boxes, object blobs, and visual tokens. These representations diversify and enrich the input conditions to the diffusion models, enabling more diverse outputs. Our experimental results demonstrate that Kaleido effectively broadens the diversity of the generated image samples from a given textual description while maintaining high image quality. Furthermore, we show that Kaleido adheres closely to the guidance provided by the generated latent variables, demonstrating its capability to effectively control and direct the image generation process.

Two competing paradigms exist for self-supervised learning of data representations. Joint Embedding Predictive Architectures (JEPAs) is a class of architectures in which semantically similar inputs are encoded into representations that are predictive of each other. A recent successful approach that falls under the JEPA framework is self-distillation, where an online encoder is trained to predict the output of the target encoder, sometimes with a lightweight predictor network. This is contrasted with the Masked Auto Encoder (MAE) paradigm, where an encoder and decoder are trained to reconstruct missing parts of the input in ambient space rather than its latent representation. A common motivation for using the JEPA approach over MAE is that the JEPA objective prioritizes abstract features over fine-grained pixel information (which can be unpredictable and uninformative). In this work, we seek to understand the mechanism behind this empirical observation by analyzing deep linear models. We uncover a surprising mechanism: in a simplified linear setting where both approaches learn similar representations, JEPAs are biased to learn high influence features, or features characterized by having high regression coefficients. Our results point to a distinct implicit bias of predicting in latent space that may shed light on its success in practice.

Novel view synthesis from a single image requires inferring occluded regions of objects and scenes whilst simultaneously maintaining semantic and physical consistency with the input. Existing approaches condition neural radiance fields (NeRF) on local image features, projecting points to the input image plane, and aggregating 2D features to perform volume rendering. However, under severe occlusion, this projection fails to resolve uncertainty, resulting in blurry renderings that lack details. In this work, we propose NerfDiff, which addresses this issue by distilling the knowledge of a 3D-aware conditional diffusion model (CDM) into NeRF through synthesizing and refining a set of virtual views at test-time. We further propose a novel NeRF-guided distillation algorithm that simultaneously generates 3D consistent virtual views from the CDM samples, and finetunes the NeRF based on the improved virtual views. Our approach significantly outperforms existing NeRF-based and geometry-free approaches on challenging datasets including ShapeNet, ABO, and Clevr3D.

Training stability is of great importance to Transformers. In this work, we investigate the training dynamics of Transformers by examining the evolution of the attention layers. In particular, we track the attention entropy for each attention head during the course of training, which is a proxy for model sharpness. We identify a common pattern across different architectures and tasks, where low attention entropy is accompanied by high training instability, which can take the form of oscillating loss or divergence. We denote the pathologically low attention entropy, corresponding to highly concentrated attention scores, as $\textit{entropy collapse}$. As a remedy, we propose $\sigma$Reparam, a simple and efficient solution where we reparametrize all linear layers with spectral normalization and an additional learned scalar. We demonstrate that $\sigma$Reparam successfully prevents entropy collapse in the attention layers, promoting more stable training. Additionally, we prove a tight lower bound of the attention entropy, which decreases exponentially fast with the spectral norm of the attention logits, providing additional motivation for our approach. We conduct experiments with $\sigma$Reparam on image classification, image self-supervised learning, machine translation, speech recognition, and language modeling tasks. We show that $\sigma$Reparam provides stability and robustness with respect to the choice of hyperparameters, going so far as enabling training (a) a Vision Transformer to competitive performance without warmup, weight decay, layer normalization or adaptive optimizers; (b) deep architectures in machine translation and (c) speech recognition to competitive performance without warmup and adaptive optimizers. Code is available at https://github.com/apple/ml-sigma-reparam.

Autoregressive models for text sometimes generate repetitive and low-quality output because errors accumulate during the steps of generation. This issue is often attributed to exposure bias -- the difference between how a model is trained, and how it is used during inference. Denoising diffusion models provide an alternative approach in which a model can revisit and revise its output. However, they can be computationally expensive and prior efforts on text have led to models that produce less fluent output compared to autoregressive models, especially for longer text and paragraphs. In this paper, we propose PLANNER, a model that combines latent semantic diffusion with autoregressive generation, to generate fluent text while exercising global control over paragraphs. The model achieves this by combining an autoregressive "decoding" module with a "planning" module that uses latent diffusion to generate semantic paragraph embeddings in a coarse-to-fine manner. The proposed method is evaluated on various conditional generation tasks, and results on semantic generation, text completion and summarization show its effectiveness in generating high-quality long-form text in an efficient manner.

We identify incremental learning dynamics in transformers, where the difference between trained and initial weights progressively increases in rank. We rigorously prove this occurs under the simplifying assumptions of diagonal weight matrices and small initialization. Our experiments support the theory and also show that phenomenon can occur in practice without the simplifying assumptions.

Self-attention mechanisms model long-range context by using pairwise attention between all input tokens. In doing so, they assume a fixed attention granularity defined by the individual tokens (e.g., text characters or image pixels), which may not be optimal for modeling complex dependencies at higher levels. In this paper, we propose ContextPool to address this problem by adapting the attention granularity for each token. Inspired by the success of ConvNets that are combined with pooling to capture long-range dependencies, we learn to pool neighboring features for each token before computing attention in a given attention layer. The pooling weights and support size are adaptively determined, allowing the pooled features to encode meaningful context with varying scale. We show that ContextPool makes attention models more expressive, achieving strong performance often with fewer layers and thus significantly reduced cost. Experiments validate that our ContextPool module, when plugged into transformer models, matches or surpasses state-of-the-art performance using less compute on several language and image benchmarks, outperforms recent works with learned context sizes or sparse attention patterns, and is also applicable to ConvNets for efficient feature learning.

Transformers \cite{transformer} have gained increasing popularity in a wide range of applications, including Natural Language Processing (NLP), Computer Vision and Speech Recognition, because of their powerful representational capacity. However, harnessing this representational capacity effectively requires a large amount of data, strong regularization, or both, to mitigate overfitting. Recently, the power of the Transformer has been unlocked by self-supervised pretraining strategies based on masked autoencoders which rely on reconstructing masked inputs, directly, or contrastively from unmasked content. This pretraining strategy which has been used in BERT models in NLP \cite{bert}, Wav2Vec models in Speech \cite{wv2v2} and, recently, in MAE models in Vision \cite{beit, mae}, forces the model to learn about relationships between the content in different parts of the input using autoencoding related objectives. In this paper, we propose a novel, but surprisingly simple alternative to content reconstruction – that of predicting locations from content, without providing positional information for it. Doing so requires the Transformer to understand the positional relationships between different parts of the input, from their content alone. This amounts to an efficient implementation where the pretext task is a classification problem among all possible positions for each input token. We experiment on both Vision and Speech benchmarks, where our approach brings improvements over strong supervised training baselines and is comparable to modern unsupervised/self-supervised pretraining methods. Our method also enables Transformers trained without position embeddings to outperform ones trained with full position information.

We introduce GAUDI, a generative model capable of capturing the distribution of complex and realistic 3D scenes that can be rendered immersively from a moving camera. We tackle this challenging problem with a scalable yet powerful approach, where we first optimize a latent representation that disentangles radiance fields and camera poses. This latent representation is then used to learn a generative model that enables both unconditional and conditional generation of 3D scenes. Our model generalizes previous works that focus on single objects by removing the assumption that the camera pose distribution can be shared across samples. We show that GAUDI obtains state-of-the-art performance in the unconditional generative setting across multiple datasets and allows for conditional generation of 3D scenes given conditioning variables like sparse image observations or text that describes the scene.

We study the problem of novel view synthesis from sparse source observations of a scene comprised of 3D objects. We propose a simple yet effective approach that is neither continuous nor implicit, challenging recent trends on view synthesis. Our approach explicitly encodes observations into a volumetric representation that enables amortized rendering. We demonstrate that although continuous radiance field representations have gained a lot of attention due to their expressive power, our simple approach obtains comparable or even better novel view reconstruction quality comparing with state-of-the-art baselines [49] while increasing rendering speed by over 400x. Our model is trained in a category-agnostic manner and does not require scene-specific optimization. Therefore, it is able to generalize novel view synthesis to object categories not seen during training. In addition, we show that with our simple formulation, we can use view synthesis as a self-supervision signal for efficient learning of 3D geometry without explicit 3D supervision.

We study the problem of directly optimizing arbitrary non-differentiable task evaluation metrics such as misclassification rate and recall. Our method, named MetricOpt, operates in a black-box setting where the computational details of the target metric are unknown. We achieve this by learning a differentiable value function, which maps compact task-specific model parameters to metric observations. The learned value function is easily pluggable into existing optimizers like SGD and Adam, and is effective for rapidly finetuning a pre-trained model. This leads to consistent improvements since the value function provides effective metric supervision during finetuning, and helps to correct the potential bias of loss-only supervision. MetricOpt achieves state-of-the-art performance on a variety of metrics for (image) classification, image retrieval and object detection. Solid benefits are found over competing methods, which often involve complex loss design or adaptation. MetricOpt also generalizes well to new tasks and model architectures.

We tackle the challenge of learning a distribution over complex, realistic, indoor scenes. In this paper, we introduce Generative Scene Networks (GSN), which learns to decompose scenes into a collection of many local radiance fields that can be rendered from a free moving camera. Our model can be used as a prior to generate new scenes, or to complete a scene given only sparse 2D observations. Recent work has shown that generative models of radiance fields can capture properties such as multi-view consistency and view-dependent lighting. However, these models are specialized for constrained viewing of single objects, such as cars or faces. Due to the size and complexity of realistic indoor environments, existing models lack the representational capacity to adequately capture them. Our decomposition scheme scales to larger and more complex scenes while preserving details and diversity, and the learned prior enables high-quality rendering from viewpoints that are significantly different from observed viewpoints. When compared to existing models, GSN produces quantitatively higher-quality scene renderings across several different scene datasets.

For many fundamental scene understanding tasks, it is difficult or impossible to obtain per-pixel ground truth labels from real images. We address this challenge by introducing Hypersim, a photorealistic synthetic dataset for holistic indoor scene understanding. To create our dataset, we leverage a large repository of synthetic scenes created by professional artists, and we generate 77,400 images of 461 indoor scenes with detailed per-pixel labels and corresponding ground truth geometry. Our dataset: (1) relies exclusively on publicly available 3D assets; (2) includes complete scene geometry, material information, and lighting information for every scene; (3) includes dense per-pixel semantic instance segmentations and complete camera information for every image; and (4) factors every image into diffuse reflectance, diffuse illumination, and a non-diffuse residual term that captures view-dependent lighting effects.We analyze our dataset at the level of scenes, objects, and pixels, and we analyze costs in terms of money, computation time, and annotation effort. Remarkably, we find that it is possible to generate our entire dataset from scratch, for roughly half the cost of training a popular open-source natural language processing model. We also evaluate sim-to-real transfer performance on two real-world scene understanding tasks – semantic segmentation and 3D shape prediction – where we find that pre-training on our dataset significantly improves performance on both tasks, and achieves state-of-the-art performance on the most challenging Pix3D test set. All of our rendered image data, as well as all the code we used to generate our dataset and perform our experiments, is available online.

Offline Reinforcement Learning promises to learn effective policies from previously-collected, static datasets without the need for exploration. However, existing Q-learning and actor-critic based off-policy RL algorithms fail when bootstrapping from out-of-distribution (OOD) actions or states. We hypothesize that a key missing ingredient from the existing methods is a proper treatment of uncertainty in the offline setting. We propose Uncertainty Weighted Actor-Critic (UWAC), an algorithm that detects OOD state-action pairs and down-weights their contribution in the training objectives accordingly. Implementation-wise, we adopt a practical and effective dropout-based uncertainty estimation method that introduces very little overhead over existing RL algorithms. Empirically, we observe that UWAC substantially improves model stability during training. In addition, UWAC out-performs existing offline RL methods on a variety of competitive tasks, and achieves significant performance gains over the state-of-the-art baseline on datasets with sparse demonstrations collected from human experts.

State-of-the-art learning-based monocular 3D reconstruction methods learn priors over object categories on the training set, and as a result struggle to achieve reasonable generalization to object categories unseen during training. In this paper we study the inductive biases encoded in the model architecture that impact the generalization of learning-based 3D reconstruction methods. We find that 3 inductive biases impact performance: the spatial extent of the encoder, the use of the underlying geometry of the scene to describe point features, and the mechanism to aggregate information from multiple views. Additionally, we propose mechanisms to enforce those inductive biases: a point representation that is aware of camera position, and a variance cost to aggregate information across views. Our model achieves state-of-the-art results on the standard ShapeNet 3D reconstruction benchmark in various settings.

We propose a framework for learning neural scene representations directly from images, without 3D supervision. Our key insight is that 3D structure can be imposed by ensuring that the learned representation transforms like a real 3D scene. Specifically, we introduce a loss which enforces equivariance of the scene representation with respect to 3D transformations. Our formulation allows us to infer and render scenes in real time while achieving comparable results to models requiring minutes for inference. In addition, we introduce two challenging new datasets for scene representation and neural rendering, including scenes with complex lighting and backgrounds. Through experiments, we show that our model achieves compelling results on these datasets as well as on standard ShapeNet benchmarks.

Modern neural network performance typically improves as model size increases. A recent line of research on the Neural Tangent Kernel (NTK) of over-parameterized networks indicates that the improvement with size increase is a product of a better conditioned loss landscape. In this work, we investigate a form of over-parameterization achieved through ensembling, where we define collegial ensembles (CE) as the aggregation of multiple independent models with identical architectures, trained as a single model. We show that the optimization dynamics of CE simplify dramatically when the number of models in the ensemble is large, resembling the dynamics of wide models, yet scale much more favorably. We use recent theoretical results on the finite width corrections of the NTK to perform efficient architecture search in a space of finite width CE that aims to either minimize capacity, or maximize trainability under a set of constraints. The resulting ensembles can be efficiently implemented in practical architectures using group convolutions and block diagonal layers. Finally, we show how our framework can be used to analytically derive optimal group convolution modules originally found using expensive grid searches, without having to train a single model.

In most machine learning training paradigms a fixed, often handcrafted, loss function is assumed to be a good proxy for an underlying evaluation metric. In this work we assess this assumption by meta-learning an adaptive loss function to directly optimize the evaluation metric. We propose a sample efficient reinforcement learning approach for adapting the loss dynamically during training. We empirically show how this formulation improves performance by simultaneously optimizing the evaluation metric and smoothing the loss landscape. We verify our method in metric learning and classification scenarios, showing considerable improvements over the state-of-the-art on a diverse set of tasks. Importantly, our method is applicable to a wide range of loss functions and evaluation metrics. Furthermore, the learned policies are transferable across tasks and data, demonstrating the versatility of the method.

We examine Generative Adversarial Networks (GANs) through the lens of deep Energy Based Models (EBMs), with the goal of exploiting the density model that follows from this formulation. In contrast to a traditional view where the discriminator learns a constant function when reaching convergence, here we show that it can provide useful information for downstream tasks, e.g., feature extraction for classification. To be concrete, in the EBM formulation, the discriminator learns an unnormalized density function (i.e., the negative energy term) that characterizes the data manifold. We propose to evaluate both the generator and the discriminator by deriving corresponding Fisher Score and Fisher Information from the EBM. We show that by assuming that the generated examples form an estimate of the learned density, both the Fisher Information and the normalized Fisher Vectors are easy to compute. We also show that we are able to derive a distance metric between examples and between sets of examples. We conduct experiments showing that the GAN-induced Fisher Vectors demonstrate competitive performance as unsupervised feature extractors for classification and perceptual similarity tasks. Code is available at \url{https://github.com/apple/ml-afv}.

This paper describes a Markov Random Field for real-valued image modeling that has two sets of latent variables. One set is used to gate the interactions between all pairs of pixels, while the second set determines the mean intensities of each pixel. This is a powerful model with a conditional distribution over the input that is Gaussian, with both mean and covariance determined by the configuration of latent variables, which is unlike previous models that were restricted to using Gaussians with either a fixed mean or a diagonal covariance matrix. Thanks to the increased flexibility, this gated MRF can generate more realistic samples after training on an unconstrained distribution of high-resolution natural images. Furthermore, the latent variables of the model can be inferred efficiently and can be used as very effective descriptors in recognition tasks. Both generation and discrimination drastically improve as layers of binary latent variables are added to the model, yielding a hierarchical model called a Deep Belief Network.

The most popular way to use probabilistic models in vision is first to extract some descriptors of small image patches or object parts using well-engineered features, and then to use statistical learning tools to model the dependencies among these features and eventual labels. Learning probabilistic models directly on the raw pixel values has proved to be much more difficult and is typically only used for regularizing discriminative methods. In this work, we use one of the best, pixel-level, generative models of natural images-a gated MRF-as the lowest level of a deep belief network (DBN) that has several hidden layers. We show that the resulting DBN is very good at coping with occlusion when predicting expression categories from face images, and it can produce features that perform comparably to SIFT descriptors for discriminating different types of scene. The generative ability of the model also makes it easy to see what information is captured and what is lost at each level of representation.

We describe a generative model of the relationship between two images. The model is defined as a factored three-way Boltzmann machine, in which hidden variables collaborate to define the joint correlation matrix for image pairs. Modeling the joint distribution over pairs makes it possible to efficiently match images that are the same according to a learned measure of similarity. We apply the model to several face matching tasks, and show that it learns to represent the input images using task-specific basis functions. Matching performance is superior to previous similar generative models, including recent conditional models of transformations. We also show that the model can be used as a plug-in matching score to perform invariant classification.