3D Gaussian Splatting (3DGS) is an increasingly popular novel view synthesis approach due to its fast rendering time, and high-quality output. However, scaling 3DGS to large (or intricate) scenes is challenging due to its substantial memory requirement, which exceeds the memory capacity of most GPUs. In this paper, we describe CLM, a system that allows 3DGS to render large scenes using a single consumer-grade GPU, e.g., RTX4090. It does so by offloading Gaussians to CPU memory, and loading them into GPU memory only when necessary. To improve performance and reduce communication overheads, CLM uses a novel offloading strategy based on insights into 3DGS's memory access patterns. This strategy enables efficient pipelining, which overlaps GPU-to-CPU communication, GPU computation and CPU computation. Furthermore, CLM exploits these access patterns to reduce communication volume. Our evaluation shows that the resulting implementation can render a large scene that requires 102 million Gaussians on a single RTX4090 and achieve state-of-the-art reconstruction quality. The code is open-sourced at: https://github.com/nyu-systems/CLM-GS

We present a novel approach to integrating scientific knowledge into generative models, enhancing their realism and consistency in image synthesis. First, we introduce Science-T2I, an expert-annotated adversarial dataset comprising adversarial 20k image pairs with 9k prompts, covering wide distinct scientific knowledge categories. Leveraging Science-T2I, we present SciScore, an end-to-end reward model that refines the assessment of generated images based on scientific knowledge, which is achieved by augmenting both the scientific comprehension and visual capabilities of pre-trained CLIP model. Additionally, based on Science-T2I, we propose a two-stage training framework, comprising a supervised fine-tuning phase and a masked online fine-tuning phase, to incorporate scientific knowledge into existing generative models. Through comprehensive experiments, we demonstrate the effectiveness of our framework in establishing new standards for evaluating the scientific realism of generated content. Specifically, SciScore attains performance comparable to human-level, demonstrating a 5% improvement similar to evaluations conducted by experienced human evaluators. Furthermore, by applying our proposed fine-tuning method to FLUX, we achieve a performance enhancement exceeding 50% on SciScore.

Generative models have made significant impacts across various domains, largely due to their ability to scale during training by increasing data, computational resources, and model size, a phenomenon characterized by the scaling laws. Recent research has begun to explore inference-time scaling behavior in Large Language Models (LLMs), revealing how performance can further improve with additional computation during inference. Unlike LLMs, diffusion models inherently possess the flexibility to adjust inference-time computation via the number of denoising steps, although the performance gains typically flatten after a few dozen. In this work, we explore the inference-time scaling behavior of diffusion models beyond increasing denoising steps and investigate how the generation performance can further improve with increased computation. Specifically, we consider a search problem aimed at identifying better noises for the diffusion sampling process. We structure the design space along two axes: the verifiers used to provide feedback, and the algorithms used to find better noise candidates. Through extensive experiments on class-conditioned and text-conditioned image generation benchmarks, our findings reveal that increasing inference-time compute leads to substantial improvements in the quality of samples generated by diffusion models, and with the complicated nature of images, combinations of the components in the framework can be specifically chosen to conform with different application scenario.

This paper does not describe a new method; instead, it provides a thorough exploration of an important yet understudied design space related to recent advances in text-to-image synthesis—specifically, the deep fusion of large language models (LLMs) with diffusion transformers (DiTs) for multi-modal generation. Previous studies mainly focused on over-all system performance rather than detailed comparisons with alternative methods, and key design details and training recipes were often left undisclosed. These gaps create uncertainty about the real potential of this approach. To fill these gaps, we conduct an empirical study on text-to-image generation, performing controlled comparisons with established baselines, analyzing important design choices, and providing a clear, reproducible recipe for training at scale. We hope this work offers meaningful data points and practical guidelines for future research in multimodal generation.

Humans possess the visual-spatial intelligence to remember spaces from sequential visual observations. However, can Multimodal Large Language Models (MLLMs) trained on million-scale video datasets also "think in space" from videos? We present a novel video-based visual-spatial intelligence benchmark (VSI-Bench) of over 5,000 question-answer pairs, and find that MLLMs exhibit competitive—though subhuman—visual-spatial intelligence. We probe models to express how they think in space both linguistically and visually and find that while spatial reasoning capabilities remain the primary bottleneck for MLLMs to reach higher benchmark performance, local world models and spatial awareness do emerge within these models. Notably, prevailing linguistic reasoning techniques (e.g., chain-of-thought, self-consistency, tree-of-thoughts) fail to improve performance, whereas explicitly generating cognitive maps during question-answering enhances MLLMs’ spatial distance ability.

Visual Self-Supervised Learning (SSL) currently underperforms Contrastive Language-Image Pretraining (CLIP) in multimodal settings such as Visual Question Answering (VQA). This multimodal gap is often attributed to the semantics introduced by language supervision, even though visual SSL and CLIP models are often trained on different data. In this work, we ask the question: "Do visual self-supervised approaches lag behind CLIP due to the lack of language supervision, or differences in the training data?" We study this question by training both visual SSL and CLIP models on the same MetaCLIP data, and leveraging VQA as a diverse testbed for vision encoders. In this controlled setup, visual SSL models scale better than CLIP models in terms of data and model capacity, and visual SSL performance does not saturate even after scaling up to 7B parameters. Consequently, we observe visual SSL methods achieve CLIP-level performance on a wide range of VQA and classic vision benchmarks. These findings demonstrate that pure visual SSL can match language-supervised visual pretraining at scale, opening new opportunities for vision-centric representation learning.

In this paper we tackle a fundamental question: "Can we train latent diffusion models together with the variational auto-encoder (VAE) tokenizer in an end-to-end manner?" Traditional deep-learning wisdom dictates that end-to-end training is often preferable when possible. However, for latent diffusion transformers, it is observed that end-to-end training both VAE and diffusion-model using standard diffusion-loss is ineffective, even causing a degradation in final performance. We show that while diffusion loss is ineffective, end-to-end training can be unlocked through the representation-alignment (REPA) loss -- allowing both VAE and diffusion model to be jointly tuned during the training process. Despite its simplicity, the proposed training recipe (REPA-E) shows remarkable performance; speeding up diffusion model training by over 17x and 45x over REPA and vanilla training recipes, respectively. Interestingly, we observe that end-to-end tuning with REPA-E also improves the VAE itself; leading to improved latent space structure and downstream generation performance. In terms of final performance, our approach sets a new state-of-the-art; achieving FID of 1.12 and 1.69 with and without classifier-free guidance on ImageNet 256 x 256. Code is available at https://end2end-diffusion.github.io.

In this work, we propose Visual-Predictive Instruction Tuning (VPiT) - a simple and effective extension to visual instruction tuning that enables a pretrained LLM to quickly morph into an unified autoregressive model capable of generating both text and visual tokens. VPiT teaches an LLM to predict discrete text tokens and continuous visual tokens from any input sequence of image and text data curated in an instruction-following format. Our empirical investigation reveals several intriguing properties of VPiT: (1) visual generation ability emerges as a natural byproduct of improved visual understanding, and can be unlocked efficiently with a small amount of generation data; (2) while we find understanding and generation to be mutually beneficial, understanding data contributes to both capabilities more effectively than generation data. Building upon these findings, we train our MetaMorph model and achieve competitive performance on both visual understanding and generation. In visual generation, MetaMorph can leverage the world knowledge and reasoning abilities gained from LLM pretraining, and overcome common failure modes exhibited by other generation models. Our results suggest that LLMs may have strong "prior" vision capabilities that can be efficiently adapted to both visual understanding and generation with a relatively simple instruction tuning process.

Video detailed captioning is a key task which aims to generate comprehensive and coherent textual descriptions of video content, benefiting both video understanding and generation. In this paper, we propose AuroraCap, a video captioner based on a large multimodal model. We follow the simplest architecture design without additional parameters for temporal modeling. To address the overhead caused by lengthy video sequences, we implement the token merging strategy, reducing the number of input visual tokens. Surprisingly, we found that this strategy results in little performance loss. AuroraCap shows superior performance on various video and image captioning benchmarks, for example, obtaining a CIDEr of 88.9 on Flickr30k, beating GPT-4V (55.3) and Gemini-1.5 Pro (82.2). However, existing video caption benchmarks only include simple descriptions, consisting of a few dozen words, which limits research in this field. Therefore, we develop VDC, a video detailed captioning benchmark with over one thousand carefully annotated structured captions. In addition, we propose a new LLM-assisted metric VDCscore for bettering evaluation, which adopts a divide-and-conquer strategy to transform long caption evaluation into multiple short question-answer pairs. With the help of human Elo ranking, our experiments show that this benchmark better correlates with human judgments of video detailed captioning quality.

In this study, we examine the representation learning abilities of Denoising Diffusion Models (DDM) that were originally purposed for image generation. Our philosophy is to deconstruct a DDM, gradually transforming it into a classical Denoising Autoencoder (DAE). This deconstructive process allows us to explore how various components of modern DDMs influence self-supervised representation learning. We observe that only a very few modern components are critical for learning good representations, while many others are nonessential. Our study ultimately arrives at an approach that is highly simplified and to a large extent resembles a classical DAE. We hope our study will rekindle interest in a family of classical methods within the realm of modern self-supervised learning.

Recent advances in diffusion models have introduced a new era of text-guided image manipulation, enabling users to create realistic edited images with simple textual prompts. However, there is significant concern about the potential misuse of these methods, especially in creating misleading or harmful content. Although recent defense strategies, which introduce imperceptible adversarial noise to induce model failure, have shown promise, they remain ineffective against more sophisticated manipulations, such as editing with a mask. In this work, we propose DiffusionGuard, a robust and effective defense method against unauthorized edits by diffusion-based image editing models, even in challenging setups. Through a detailed analysis of these models, we introduce a novel objective that generates adversarial noise targeting the early stage of the diffusion process. This approach significantly improves the efficiency and effectiveness of adversarial noises. We also introduce a mask-augmentation technique to enhance robustness against various masks during test time. Finally, we introduce a comprehensive benchmark designed to evaluate the effectiveness and robustness of methods in protecting against privacy threats in realistic scenarios. Through extensive experiments, we show that our method achieves stronger protection and improved mask robustness with lower computational costs compared to the strongest baseline. Additionally, our method exhibits superior transferability and better resilience to noise removal techniques compared to all baseline methods. Our source code is publicly available at https://choi403.github.io/diffusionguard.

Recent studies have shown that the denoising process in (generative) diffusion models can induce meaningful (discriminative) representations inside the model, though the quality of these representations still lags behind those learned through recent self-supervised learning methods. We argue that one main bottleneck in training large-scale diffusion models for generation lies in effectively learning these representations. Moreover, training can be made easier by incorporating high-quality external visual representations, rather than relying solely on the diffusion models to learn them independently. We study this by introducing a straightforward regularization called REPresentation Alignment (REPA), which aligns the projections of noisy input hidden states in denoising networks with clean image representations obtained from external, pretrained visual encoders. The results are striking: our simple strategy yields significant improvements in both training efficiency and generation quality when applied to popular diffusion and flow-based transformers, such as DiTs and SiTs. For instance, our method can speed up SiT training by over 17.5$\times$, matching the performance (without classifier-free guidance) of a SiT-XL model trained for 7M steps in less than 400K steps. In terms of final generation quality, our approach achieves state-of-the-art results of FID=1.42 using classifier-free guidance with the guidance interval.

3D Gaussian Splatting (3DGS) is increasingly popular for 3D reconstruction due to its superior visual quality and rendering speed. However, 3DGS training currently occurs on a single GPU, limiting its ability to handle high-resolution and large-scale 3D reconstruction tasks due to memory constraints. We introduce Grendel, a distributed system designed to partition 3DGS parameters and parallelize computation across multiple GPUs. As each Gaussian affects a small, dynamic subset of rendered pixels, Grendel employs sparse all-to-all communication to transfer the necessary Gaussians to pixel partitions and performs dynamic load balancing. Unlike existing 3DGS systems that train using one camera view image at a time, Grendel supports batched training with multiple views. We explore various optimization hyperparameter scaling strategies and find that a simple sqrt(batch-size) scaling rule is highly effective. Evaluations using large-scale, high-resolution scenes show that Grendel enhances rendering quality by scaling up 3DGS parameters across multiple GPUs. On the 4K ``Rubble'' dataset, we achieve a test PSNR of 27.28 by distributing 40.4 million Gaussians across 16 GPU, compared to a PSNR of 26.28 using 11.2 million Gaussians on a single GPU. Grendel is an open-source project available at: https://github.com/nyu-systems/Grendel-GS

Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models. However, their roles in enhancing model generalization capabilities remain unclear. This paper studies the difference between SFT and RL on generalization and memorization, focusing on text-based rule variants and visual variants. We introduce GeneralPoints, an arithmetic reasoning card game, and adopt V-IRL, a real-world navigation environment, to assess how models trained with SFT and RL generalize to unseen variants in both textual and visual domains. We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants. SFT, in contrast, tends to memorize training data and struggles to generalize out-of-distribution scenarios. Further analysis reveals that RL improves the model's underlying visual recognition capabilities, contributing to its enhanced generalization in the visual domain. Despite RL's superior generalization, we show that SFT remains essential for effective RL training; SFT stabilizes the model's output format, enabling subsequent RL to achieve its performance gains. These findings demonstrates the capability of RL for acquiring generalizable knowledge in complex, multi-modal tasks.

Large-scale pre-trained video generation models excel in content creation but are not reliable as physically accurate world simulators out of the box. This work studies the process of post-training these models for accurate world modeling through the lens of the simple, yet fundamental, physics task of modeling object freefall. We show state-of-the-art video generation models struggle with this basic task, despite their visually impressive outputs. To remedy this problem, we find that fine-tuning on a relatively small amount of simulated videos is effective in inducing the dropping behavior in the model, and we can further improve results through a novel reward modeling procedure we introduce. Our study also reveals key limitations of post-training in generalization and distribution modeling. Additionally, we release a benchmark for this task that may serve as a useful diagnostic tool for tracking physical accuracy in large-scale video generative model development. Code is available at this repository: https://github.com/vision-x-nyu/pisa-experiments.

Contrastive Language-Image Pretraining (CLIP) is a popular foundation model, supporting from zero-shot classification, retrieval to encoders for multimodal large language models (MLLMs). Although CLIP is successfully trained on billion-scale image-text pairs from the English world, scaling CLIP's training further to learning from the worldwide web data is still challenging: (1) no curation method is available to handle data points from non-English world; (2) the English performance from existing multilingual CLIP is worse than its English-only counterpart, i.e., "curse of multilinguality" that is common in LLMs. Here, we present Meta CLIP 2, the first recipe training CLIP from scratch on worldwide web-scale image-text pairs. To generalize our findings, we conduct rigorous ablations with minimal changes that are necessary to address the above challenges and present a recipe enabling mutual benefits from English and non-English world data. In zero-shot ImageNet classification, Meta CLIP 2 ViT-H/14 surpasses its English-only counterpart by 0.8% and mSigLIP by 0.7%, and surprisingly sets new state-of-the-art without system-level confounding factors (e.g., translation, bespoke architecture changes) on multilingual benchmarks, such as CVQA with 57.4%, Babel-ImageNet with 50.2% and XM3600 with 64.3% on image-to-text retrieval. Code and model are available at https://github.com/facebookresearch/MetaCLIP.

Diffusion- and flow-based policies deliver state-of-the-art performance on long-horizon robotic manipulation and imitation-learning tasks. However, these controllers employ a fixed inference budget at every control step, regardless of task complexity, leading to computational inefficiency for simple subtasks while potentially underperforming on challenging ones. To address these issues, we introduce Difficulty-Aware Stochastic Interpolant Policy (DA-SIP), a framework that enables robotic controllers to adaptively adjust their integration horizon in real-time based on task difficulty. Our approach employs a difficulty classifier that analyzes RGB-D observations to dynamically select the step budget, the optimal solver variant, and ODE/SDE integration at each control cycle. DA-SIP builds upon the stochastic interpolant formulation to provide a unified framework that unlocks diverse training and inference configurations for diffusion- and flow-based policies. Through comprehensive benchmarks across diverse manipulation tasks, DA-SIP achieves 2.6-4.4× reduction in total computation time while maintaining task-success rates comparable to fixed maximum-computation baselines. By implementing adaptive computation within this framework, DA-SIP transforms generative robot controllers into efficient, task-aware systems that intelligently allocate inference resources where they provide the greatest benefit.

Recent reports claim that large language models (LLMs) now outperform elite humans in competitive programming. Drawing on knowledge from a group of medalists in international algorithmic contests, we revisit this claim, examining how LLMs differ from human experts and where limitations still remain. We introduce LiveCodeBench Pro, a benchmark composed of problems from Codeforces, ICPC, and IOI that are continuously updated to reduce the likelihood of data contamination. A team of Olympiad medalists annotates every problem for algorithmic categories and conducts a line-by-line analysis of failed model-generated submissions. Using this new data and benchmark, we find that frontier models still have significant limitations: without external tools, the best model achieves only 53\% pass@1 on medium-difficulty problems and 0\% on hard problems, domains where expert humans still excel. We also find that LLMs succeed at implementation-heavy problems but struggle with nuanced algorithmic reasoning and complex case analysis, often generating confidently incorrect justifications. High performance appears largely driven by implementation precision and tool augmentation, not superior reasoning. LiveCodeBench Pro thus highlights the significant gap to human grandmaster levels, while offering fine-grained diagnostics to steer future improvements in code-centric LLM reasoning.

The success of contrastive language-image pretraining (CLIP) relies on the supervision from the pairing between images and captions, which tends to be noisy in web- crawled data. We present Mixture of Data Experts (MoDE) and learn a system of CLIP data experts via clustering. Each data expert is trained on one data cluster, being less sensitive to false negative noises in other clusters. At inference time, we ensemble their outputs by applying weights determined through the correlation between task metadata and cluster conditions. To estimate the correlation pre-cisely, the samples in one cluster should be semantically similar, but the number of data experts should still be rea-sonable for training and inference. As such, we consider the ontology in human language and propose to use fine- grained cluster centers to represent each data expert at a coarse-grained level. Experimental studies show that four CLIP data experts on ViT-B/16 outperform the ViT-L/14 by OpenAI CLIP and OpenCLIP on zero-shot image classification but with less (<35%) training cost. Meanwhile, MoDE can train all data expert asynchronously and can flexibly include new data experts. The code is available here.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However, the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent MultiModal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors, we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify “CLIP-blind pairs”- images that CLIP perceives as similar despite their clear visual differences. With these pairs, we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems, including GPT-4V, struggle with straightforward questions across nine basic visual patterns, often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues, we propose a Mixture of Features (MoF) approach, demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together, our research suggests visual representation learning remains an open challenge, and accurate visual grounding is crucial for future successful multimodal systems.

When we look around and perform complex tasks, how we see and selectively process what we see is crucial. How-ever, the lack of this visual search mechanism in current multimodal LLMs (MLLMs) hinders their ability to focus on important visual details, especially when handling high-resolution and visually crowded images. To address this, we introduce V*, an LLM- guided visual search mechanism that employs the world knowledge in LLMs for efficient visual querying. When combined with an MLLM, this mechanism enhances collaborative reasoning, contextual understanding, and precise visual grounding. This integration results in a new MLLM meta-architecture, named Show, sEArch, and TelL (SEAL). We further create V* Bench, a benchmark specifically designed to evaluate MLLMs in their ability to process high-resolution images and focus on visual details. Our study highlights the necessity of incorporating visual search capabilities into multimodal systems. The code is available here.

We present Image Sculpting, a new framework for editing 2D images by incorporating tools from 3D geometry and graphics. This approach differs markedly from existing methods, which are confined to 2D spaces and typically rely on textual instructions, leading to ambiguity and limited control. Image Sculpting converts 2D objects into 3D, enabling direct interaction with their 3D geometry. Post-editing, these objects are re-rendered into 2D, merging into the original image to produce high-fidelity results through a coarse-to-fine enhancement process. The framework supports precise, quantifiable, and physically-plausible editing options such as pose editing, rotation, translation, 3D composition, carving, and serial addition. It marks an initial step towards combining the creative freedom of generative models with the precision of graphics pipelines.

Hu Xu, Po-Yao Huang, Xiaoqing Tan, Ching-Feng Yeh, Jacob Kahn, Christine Jou, Gargi Ghosh, Omer Levy, Luke Zettlemoyer, Wen-tau Yih, Shang-Wen Li, Saining Xie, Christoph Feichtenhofer. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its \textit{data} and \textit{not} the \textit{model} architecture or pre-training {objective}. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8\% accuracy, surpassing CLIP's 68.3\% on \mbox{ViT-B} models. Scaling to 1B data, while maintaining the same training budget, attains \textbf{72.4\%}. Our observations hold across various model sizes, exemplified by ViT-H achieving \textbf{80.5\%}, without any bells-and-whistles. Curation code and training data distribution over metadata will be made available.

We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning research. This gap hinders accurate sensory grounding in real-world scenarios. Our study uses LLMs and visual instruction tuning as an interface to evaluate various visual representations, offering new insights into different models and architectures—self-supervised, strongly supervised, or combinations thereof—based on experiments with over 15 vision models. We critically examine existing MLLM benchmarks, addressing the difficulties involved in consolidating and interpreting results from various tasks. To further improve visual grounding, we propose spatial vision aggregator (SVA), a dynamic and spatially-aware connector that integrates vision features with LLMs while reducing the number of tokens. Additionally, we discuss the curation of high-quality visual instruction-tuning data from publicly available sources, emphasizing the importance of distribution balancing. Collectively, Cambrian-1 not only achieves state-of-the-art performances but also serves as a comprehensive, open cookbook for instruction-tuned MLLMs. We provide model weights, code, supporting tools, datasets, and detailed instruction-tuning and evaluation recipes. We hope our release will inspire and accelerate advancements in multimodal systems and visual representation learning.

Large vision-language models (VLMs) fine-tuned on specialized visual instruction-following data have exhibited impressive language reasoning capabilities across various scenarios. However, this fine-tuning paradigm may not be able to efficiently learn optimal decision-making agents in multi-step goal-directed tasks from interactive environments. To address this challenge, we propose an algorithmic framework that fine-tunes VLMs with reinforcement learning (RL). Specifically, our framework provides a task description and then prompts the VLM to generate chain-of-thought (CoT) reasoning, enabling the VLM to efficiently explore intermediate reasoning steps that lead to the final text-based action. Next, the open-ended text output is parsed into an executable action to interact with the environment to obtain goal-directed task rewards. Finally, our framework uses these task rewards to fine-tune the entire VLM with RL. Empirically, we demonstrate that our proposed framework enhances the decision-making capabilities of VLM agents across various tasks, enabling 7b models to outperform commercial models such as GPT4-V or Gemini. Furthermore, we find that CoT reasoning is a crucial component for performance improvement, as removing the CoT reasoning results in a significant decrease in the overall performance of our method.

Driven by improved architectures and better representation learning frameworks, the field of visual recognition has enjoyed rapid modernization and performance boost in the early 2020s. For example, modern ConvNets, represented by ConvNeXt [33], have demonstrated strong performance in various scenarios. While these models were originally designed for supervised learning with ImageNet labels, they can also potentially benefit from self-supervised learning techniques such as masked autoencoders (MAE) [14]. However, we found that simply combining these two approaches leads to subpar performance. In this paper, we propose a fully convolutional masked autoencoder framework and a new Global Response Normalization (GRN) layer that can be added to the ConvNeXt architecture to enhance inter-channel feature competition. This co-design of self-supervised learning techniques and architectural improvement results in a new model family called ConvNeXt V2, which significantly improves the performance of pure ConvNets on various recognition benchmarks, including ImageNet classification, COCO detection, and ADE20K segmentation. We also provide pre-trained ConvNeXt V2 models of various sizes, ranging from an efficient 3.7M-parameter Atto model with 76.7% top-1 accuracy on ImageNet, to a 650M Huge model that achieves a state-of-the-art 88.9% accuracy using only public training data.

Large vision-language models are generally applicable to many downstream tasks, but come at an exorbitant training cost that only large institutions can afford. This paper trades generality for efficiency and presents Curation in Training (CiT), a simple and efficient vision-text learning algorithm that couples a data objective into training. CiT automatically yields quality data to speed-up contrastive image-text training and alleviates the need for an offline data filtering pipeline, allowing broad data sources (including raw image-text pairs from the web). CiT contains two loops: an outer loop curating the training data and an inner loop consuming the curated training data. The text encoder connects the two loops. Given metadata for tasks of interest, e.g., class names, and a large pool of image-text pairs, CiT alternatively selects relevant training data from the pool by measuring the similarity of their text embeddings and embeddings of the metadata. In our experiments, we observe that CiT can speed up training by over an order of magnitude, especially if the raw data size is large.

We explore a new class of diffusion models based on the transformer architecture. We train latent diffusion models of images, replacing the commonly-used U-Net backbone with a transformer that operates on latent patches. We analyze the scalability of our Diffusion Transformers (DiTs) through the lens of forward pass complexity as measured by Gflops. We find that DiTs with higher Gflops—through increased transformer depth/width or increased number of input tokens—consistently have lower FID. In addition to possessing good scalability properties, our largest DiT-XL/2 models outperform all prior diffusion models on the class-conditional ImageNet 512×512 and 256×256 benchmarks, achieving a state-of-the-art FID of 2.27 on the latter.

Object detection has been expanded from a limited number of categories to open vocabulary. Moving forward, a complete intelligent vision system requires understanding more fine-grained object descriptions, object parts. In this paper, we propose a detector with the ability to predict both open-vocabulary objects and their part segmentation. This ability comes from two designs. First, we train the detector on the joint of part-level, object-level and image-level data to build the multi-granularity alignment between language and image. Second, we parse the novel object into its parts by its dense semantic correspondence with the base object. These two designs enable the detector to largely benefit from various data sources and foundation models. In open-vocabulary part segmentation experiments, our method outperforms the baseline by 3.3∼7.3 mAP in cross-dataset generalization on PartImageNet, and improves the baseline by 7.3 novel AP50in cross-category generalization on Pascal Part. Finally, we train a detector that generalizes to a wide range of part segmentation datasets while achieving better performance than dataset-specific training.

The “Roaring 20s” of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually “modernize” a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.

We present Masked Feature Prediction (MaskFeat) for self-supervised pre-training of video models. Our approach first randomly masks out a portion of the input sequence and then predicts the feature of the masked regions. We study five different types of features and find Histograms of Oriented Gradients (HOG), a hand-crafted feature descriptor, works particularly well in terms of both performance and efficiency. We observe that the local contrast normalization in HOG is essential for good results, which is in line with earlier work using HOG for visual recognition. Our approach can learn abundant visual knowledge and drive large-scale Transformer based models. Without using extra model weights or supervision, MaskFeat pretrained on unlabeled videos achieves unprecedented results of 86.7% with MViTv2-L on Kinetics-400, 88.3% on Kinetics 600, 80.4% on Kinetics-700, 38.8 mAP on AVA, and 75.0% on SSv2. MaskFeat further generalizes to image input, which can be interpreted as a video with a single frame and obtains competitive results on ImageN et.

This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3× or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pretraining and shows promising scaling behavior.

Neural Architecture Search (NAS) has emerged as a promising technique for automatic neural network design. However, existing MCTS based NAS approaches often utilize manually designed action space, which is not directly related to the performance metric to be optimized (e.g., accuracy), leading to sample-inefficient explorations of architectures. To improve the sample efficiency, this paper proposes Latent Action Neural Architecture Search (LaNAS), which learns actions to recursively partition the search space into good or bad regions that contain networks with similar performance metrics. During the search phase, as different action sequences lead to regions with different performance, the search efficiency can be significantly improved by biasing towards the good regions. On three NAS tasks, empirical results demonstrate that LaNAS is at least an order more sample efficient than baseline methods including evolutionary algorithms, Bayesian optimizations, and random search. When applied in practice, both one-shot and regular LaNAS consistently outperform existing results. Particularly, LaNAS achieves 99.0 percent accuracy on CIFAR-10 and 80.8 percent top1 accuracy at 600 MFLOPS on ImageNet in only 800 samples, significantly outperforming AmoebaNet with 33× fewer samples. Our code is publicly available at https://github.com/facebookresearch/LaMCTS.

The rapid progress in 3D scene understanding has come with growing demand for data; however, collecting and an-notating 3D scenes (e.g. point clouds) are notoriously hard. For example, the number of scenes (e.g. indoor rooms) that can be accessed and scanned might be limited; even given sufficient data, acquiring 3D labels (e.g. instance masks) requires intensive human labor. In this paper, we explore data-efficient learning for 3D point cloud. As a first step towards this direction, we propose Contrastive Scene Contexts, a 3D pre-training method that makes use of both point-level correspondences and spatial contexts in a scene. Our method achieves state-of-the-art results on a suite of benchmarks where training data or labels are scarce. Our study reveals that exhaustive labelling of 3D point clouds might be unnecessary; and remarkably, on ScanNet, even using 0.1% of point labels, we still achieve 89% (instance segmentation) and 96% (semantic segmentation) of the baseline performance that uses full annotations.

This paper does not describe a novel method. Instead, it studies a straightforward, incremental, yet must-know baseline given the recent progress in computer vision: self-supervised learning for Vision Transformers (ViT). While the training recipes for standard convolutional networks have been highly mature and robust, the recipes for ViT are yet to be built, especially in the self-supervised scenarios where training becomes more challenging. In this work, we go back to basics and investigate the effects of several fundamental components for training self-supervised ViT. We observe that instability is a major issue that degrades accuracy, and it can be hidden by apparently good results. We reveal that these results are indeed partial failure, and they can be improved when training is made more stable. We benchmark ViT results in MoCo v3 and several other self-supervised frameworks, with ablations in various aspects. We discuss the currently positive evidence as well as challenges and open questions. We hope that this work will provide useful data points and experience for future research.

Recent advances in 3D perception have shown impressive progress in understanding geometric structures of 3D shapes and even scenes. Inspired by these advances in geometric understanding, we aim to imbue image-based perception with representations learned under geometric constraints. We introduce an approach to learn view-invariant, geometry-aware representations for network pre-training, based on multi-view RGB-D data, that can then be effectively transferred to downstream 2D tasks. We propose to employ contrastive learning under both multi-view image constraints and image-geometry constraints to encode 3D priors into learned 2D representations. This results not only in improvement over 2D-only representation learning on the image-based tasks of semantic segmentation, instance segmentation and object detection on real-world indoor datasets, but moreover, provides significant improvement in the low data regime. We show significant improvement of 6.0% on semantic segmentation on full data as well as 11.9% on 20% data against baselines on ScanNet. Our code is open sourced at https://github.com/Sekunde/Pri3D.

Invariance to a broad array of image corruptions, such as warping, noise, or color shifts, is an important aspect of building robust models in computer vision. Recently, several new data augmentations have been proposed that significantly improve performance on ImageNet-C, a benchmark of such corruptions. However, there is still a lack of basic understanding on the relationship between data augmentations and test-time corruptions. To this end, we develop a feature space for image transforms, and then use a new measure in this space between augmentations and corruptions called the Minimal Sample Distance to demonstrate there is a strong correlation between similarity and performance. We then investigate recent data augmentations and observe a significant degradation in corruption robustness when the test-time corruptions are sampled to be perceptually dissimilar from ImageNet-C in this feature space. Our results suggest that test error can be improved by training on perceptually similar augmentations, and data augmentations may not generalize well beyond the existing benchmark. We hope our results and tools will allow for more robust progress towards improving robustness to image corruptions. We provide code at https://github.com/facebookresearch/augmentation-corruption.

We present Momentum Contrast (MoCo) for unsupervised visual representation learning. From a perspective on contrastive learning as dictionary look-up, we build a dynamic dictionary with a queue and a moving-averaged encoder. This enables building a large and consistent dictionary on-the-fly that facilitates contrastive unsupervised learning. MoCo provides competitive results under the common linear protocol on ImageNet classification. More importantly, the representations learned by MoCo transfer well to downstream tasks. MoCo can outperform its supervised pre-training counterpart in 7 detection/segmentation tasks on PASCAL VOC, COCO, and other datasets, sometimes surpassing it by large margins. This suggests that the gap between unsupervised and supervised representation learning has been largely closed in many vision tasks.

Differentiable Neural Architecture Search (DNAS) has demonstrated great success in designing state-of-the-art, efficient neural networks. However, DARTS-based DNAS's search space is small when compared to other search methods', since all candidate network layers must be explicitly instantiated in memory. To address this bottleneck, we propose a memory and computationally efficient DNAS variant: DMaskingNAS. This algorithm expands the search space by up to 10^14x over conventional DNAS, supporting searches over spatial and channel dimensions that are otherwise prohibitively expensive: input resolution and number of filters. We propose a masking mechanism for feature map reuse, so that memory and computational costs stay nearly constant as the search space expands. Furthermore, we employ effective shape propagation to maximize per-FLOP or per-parameter accuracy. The searched FBNetV2s yield state-of-the-art performance when compared with all previous architectures. With up to 421x less search cost, DMaskingNAS finds models with 0.9% higher accuracy, 15% fewer FLOPs than MobileNetV3-Small; and with similar accuracy but 20% fewer FLOPs than Efficient-B0. Furthermore, our FBNetV2 outperforms MobileNetV3 by 2.6% in accuracy, with equivalent model size. FBNetV2 models are open-sourced at https://github.com/facebookresearch/mobile-vision.

Neural networks are often represented as graphs of connections between neurons. However, despite their wide use, there is currently little understanding of the relationship between the graph structure of the neural network and its predictive performance. Here we systematically investigate how does the graph structure of neural networks affect their predictive performance. To this end, we develop a novel graph-based representation of neural networks called relational graph, where layers of neural network computation correspond to rounds of message exchange along the graph structure. Using this representation we show that: (1) a “sweet spot” of relational graphs leads to neural networks with significantly improved predictive performance; (2) neural network’s performance is approximately a smooth function of the clustering coefficient and average path length of its relational graph; (3) our findings are consistent across many different tasks and datasets; (4) the sweet spot can be identified efficiently; (5) top-performing neural networks have graph structure surprisingly similar to those of real biological neural networks. Our work opens new directions for the design of neural architectures and the understanding on neural networks in general.

Research on 2D and 3D generative models typically focuses on the final artifact being created, e.g., an image or a 3D structure. Unlike 2D image generation, the generation of 3D objects in the real world is commonly constrained by the process and order in which the object is constructed. For instance, gravity needs to be taken into account when building a block tower. In this paper, we explore the prediction of ordered actions to construct 3D objects. Instead of predicting actions based on physical constraints, we propose learning through observing human actions. To enable large-scale data collection, we use the Minecraft1 environment. We introduce 3D-Craft, a new dataset of 2,500 Minecraft houses each built by human players sequentially from scratch. To learn from these human action sequences, we propose an order-aware 3D generative model called VoxelCNN. In contrast to other 3D generative models which either have no explicit order (e.g. holistic generation with 3DGAN [35]), or follow a simple heuristic order (e.g. raster-scan), VoxelCNN is trained to imitate human building order with spatial awareness. We also transferred the order to other dataset such as ShapeNet[10]. The 3D-Craft dataset, models, and benchmark system will be made publicly available, which may inspire new directions for future research exploration. https://github.com/facebookresearch/VoxelCNN.

Over the past several years progress in designing better neural network architectures for visual recognition has been substantial. To help sustain this rate of progress, in this work we propose to reexamine the methodology for comparing network architectures. In particular, we introduce a new comparison paradigm of distribution estimates, in which network design spaces are compared by applying statistical techniques to populations of sampled models, while controlling for confounding factors like network complexity. Compared to current methodologies of comparing point and curve estimates of model families, distribution estimates paint a more complete picture of the entire design landscape. As a case study, we examine design spaces used in neural architecture search (NAS). We find significant statistical differences between recent NAS design space variants that have been largely overlooked. Furthermore, our analysis reveals that the design spaces for standard model families like ResNeXt can be comparable to the more complex ones used in recent NAS work. We hope these insights into distribution analysis will enable more robust progress toward discovering better networks for visual recognition.

Neural networks for image recognition have evolved through extensive manual design from simple chain-like models to structures with multiple wiring paths. The success of ResNets and DenseNets is due in large part to their innovative wiring plans. Now, neural architecture search (NAS) studies are exploring the joint optimization of wiring and operation types, however, the space of possible wirings is constrained and still driven by manual design despite being searched. In this paper, we explore a more diverse set of connectivity patterns through the lens of randomly wired neural networks. To do this, we first define the concept of a stochastic network generator that encapsulates the entire network generation process. Encapsulation provides a unified view of NAS and randomly wired networks. Then, we use three classical random graph models to generate randomly wired graphs for networks. The results are surprising: several variants of these random generators yield network instances that have competitive accuracy on the ImageNet benchmark. These results suggest that new efforts focusing on designing better network generators may lead to new breakthroughs by exploring less constrained search spaces with more room for novel design.

We tackle the problem of point cloud recognition. Unlike previous approaches where a point cloud is either converted into a volume/image or represented independently in a permutation-invariant set, we develop a new representation by adopting the concept of shape context as the building block in our network design. The resulting model, called ShapeContextNet, consists of a hierarchy with modules not relying on a fixed grid while still enjoying properties similar to those in convolutional neural networks - being able to capture and propagate the object part information. In addition, we find inspiration from self-attention based models to include a simple yet effective contextual modeling mechanism - making the contextual region selection, the feature aggregation, and the feature transformation process fully automatic. ShapeContextNet is an end-to-end model that can be applied to the general point cloud classification and segmentation problems. We observe competitive results on a number of benchmark datasets.

We present a simple, highly modularized network architecture for image classification. Our network is constructed by repeating a building block that aggregates a set of transformations with the same topology. Our simple design results in a homogeneous, multi-branch architecture that has only a few hyper-parameters to set. This strategy exposes a new dimension, which we call cardinality (the size of the set of transformations), as an essential factor in addition to the dimensions of depth and width. On the ImageNet-1K dataset, we empirically show that even under the restricted condition of maintaining complexity, increasing cardinality is able to improve classification accuracy. Moreover, increasing cardinality is more effective than going deeper or wider when we increase the capacity. Our models, named ResNeXt, are the foundations of our entry to the ILSVRC 2016 classification task in which we secured 2nd place. We further investigate ResNeXt on an ImageNet-5K set and the COCO detection set, also showing better results than its ResNet counterpart. The code and models are publicly available online.

We propose deeply-supervised nets (DSN), a method that simultaneously minimizes classification error and improves the directness and transparency of the hidden layer learning process. We focus our attention on three aspects of traditional convolutional-neural-network-type (CNN-type) architectures: (1) transparency in the effect intermediate layers have on overall classification; (2) discriminativeness and robustness of learned features, especially in early layers; (3) training effectiveness in the face of “vanishing” gradients. To combat these issues, we introduce “companion” objective functions at each hidden layer, in addition to the overall objective function at the output layer (an integrated strategy distinct from layer-wise pre-training). We also analyze our algorithm using techniques extended from stochastic gradient methods. The advantages provided by our method are evident in our experimental results, showing state-of-the-art performance on MNIST, CIFAR-10, CIFAR-100, and SVHN.

Deep convolutional neural networks (CNN) have seen tremendous success in large-scale generic object recognition. In comparison with generic object recognition, fine-grained image classification (FGIC) is much more challenging because (i) fine-grained labeled data is much more expensive to acquire (usually requiring domain expertise); (ii) there exists large intra-class and small inter-class variance. Most recent work exploiting deep CNN for image recognition with small training data adopts a simple strategy: pre-train a deep CNN on a large-scale external dataset (e.g., ImageNet) and fine-tune on the small-scale target data to fit the specific classification task. In this paper, beyond the fine-tuning strategy, we propose a systematic framework of learning a deep CNN that addresses the challenges from two new perspectives: (i) identifying easily annotated hyper-classes inherent in the fine-grained data and acquiring a large number of hyper-class-labeled images from readily available external sources (e.g., image search engines), and formulating the problem into multitask learning; (ii) a novel learning model by exploiting a regularization between the fine-grained recognition model and the hyper-class recognition model. We demonstrate the success of the proposed framework on two small-scale fine-grained datasets (Stanford Dogs and Stanford Cars) and on a large-scale car dataset that we collected.

We develop a new edge detection algorithm that addresses two critical issues in this long-standing vision problem: (1) holistic image training, and (2) multi-scale feature learning. Our proposed method, holistically-nested edge detection (HED), turns pixel-wise edge classification into image-to-image prediction by means of a deep learning model that leverages fully convolutional neural networks and deeply-supervised nets. HED automatically learns rich hierarchical representations (guided by deep supervision on side responses) that are crucially important in order to approach the human ability to resolve the challenging ambiguity in edge and object boundary detection. We significantly advance the state-of-the-art on the BSD500 dataset (ODS F-score of 0.782) and the NYU Depth dataset (ODS F-score of 0.746), and do so with an improved speed (0.4 second per image) that is orders of magnitude faster than recent CNN-based edge detection algorithms.

Linear projection for reducing data dimensionality is a common practice in various data processing applications. Among the existing projection methods, Principal Component Analysis (PCA) is arguably the most popular one. Standard PCA used in image preprocessing pursues the projection directions by minimizing the reconstruction error in a least square sense. However, since PCA does not adapt to the data or any specific domains, it may lead to severe loss of certain discriminative features during the projection, and damage the performance of either human perception (e.g. stimulus in the visual cortex, as modeled by Gabor wavelets), or machine perceptions (e.g. recognizing the images based on a certain type of visual features), or both. In this paper, we propose a novel Perception Preserving Projections (PPP) method to preserve the information for specific perception systems. In particular, PPP incorporates domain-specific feature extractor into the standard PCA formulation for the projection learning procedure. This enables PPP to make more sensible projections for feature based perception systems while retaining the simplicity and unsupervised manner of PCA. In experimental studies, PPP shows clear effectiveness and improvement over PCA in terms of two performance metrics: feature extraction deviation and the pattern recognition accuracy.