NeurIPS 2024 Papers

Image classification is one of the most fundamental capabilities of machine vision intelligence. In this work, we revisit the image classification task using visually-grounded language models (VLMs) such as GPT-4V and LLaVA. We find that existing proprietary and public VLMs, despite often using CLIP as a vision encoder and having many more parameters, significantly underperform CLIP on standard image classification benchmarks like ImageNet. To understand the reason, we explore several hypotheses concerning the inference algorithms, training objectives, and data processing in VLMs. Our analysis reveals that the primary cause is data-related: critical information for image classification is encoded in the VLM's latent space but can only be effectively decoded with enough training data. Specifically, there is a strong correlation between the frequency of class exposure during VLM training and instruction-tuning and the VLM's performance in those classes; when trained with sufficient data, VLMs can match the accuracy of state-of-the-art classification models. Based on these findings, we enhance a VLM by integrating classification-focused datasets into its training, and demonstrate that the enhanced classification performance of the VLM transfers to its general capabilities, resulting in an improvement of 11.8% on the newly collected ImageWikiQA dataset.

Text-to-image diffusion generative models can generate high quality images at the cost of tedious prompt engineering. Controllability can be improved by introducing layout conditioning, however existing methods lack layout editing ability and fine-grained control over object attributes. The concept of multi-layer generation holds great potential to address these limitations, however generating image instances concurrently to scene composition limits control over fine-grained object attributes, relative positioning in 3D space and scene manipulation abilities. In this work, we propose a novel multi-stage generation paradigm that is designed for fine-grained control, flexibility and interactivity. To ensure control over instance attributes, we devise a novel training paradigm to adapt a diffusion model to generate isolated scene components as RGBA images with transparency information. To build complex images, we employ these pre-generated instances and introduce a multi-layer composite generation process that smoothly assembles components in realistic scenes. Our experiments show that our RGBA diffusion model is capable of generating diverse and high quality instances with precise control over object attributes. Through multi-layer composition, we demonstrate that our approach allows to build and manipulate images from highly complex prompts with fine-grained control over object appearance and location, granting a higher degree of control than competing methods.

Diffusion Probabilistic Models (DPMs) have been recently utilized to deal with various blind image restoration (IR) tasks, where they have demonstrated outstanding performance in terms of perceptual quality. However, the task-specific nature of existing solutions and the excessive computational costs related to their training, make such models impractical and challenging to use for different IR tasks than those that were initially trained for. This hinders their wider adoption especially by those who lack access to powerful computational resources and vast amounts of training data. In this work we aim to address the above issues and enable the successful adoption of DPMs in practical IR-related applications. Towards this goal, we propose a modular diffusion probabilistic IR framework (DP-IR), which allows us to combine the performance benefits of existing pre-trained state-of-the-art IR networks and generative DPMs, while it requires only the additional training of a small module (0.7M params) related to the particular IR task of interest. Moreover, the architecture of our proposed framework allows us to employ a sampling strategy that leads to at least four times reduction of neural function evaluations without any performance loss, while it can also be combined with existing acceleration techniques (e.g. DDIM). We evaluate our model on four benchmarks for the tasks of burst JDD-SR, dynamic scene deblurring, and super-resolution. Our method outperforms existing approaches in terms of perceptual quality while retaining a competitive performance in relation to fidelity metrics.

The recent development of Sora leads to a new era in text-to-video (T2V) generation. Along with this comes the rising concern about its safety risks. The generated videos may contain illegal or unethical content, and there is a lack of comprehensive quantitative understanding of their safety, posing a challenge to their reliability and practical deployment. Previous evaluations primarily focus on the quality of video generation. While some evaluations of text-to-image models have considered safety, they cover limited aspects and do not address the unique temporal risk inherent in video generation. To bridge this research gap, we introduce T2VSafetyBench, the first comprehensive benchmark for conducting safety-critical assessments of text-to-video models. We define 4 primary categories with 14 critical aspects of video generation safety and construct a malicious prompt dataset including real-world prompts, LLM-generated prompts, and jailbreak attack-based prompts. We then conduct a thorough safety evaluation on 9 recently released T2V models. Based on our evaluation results, we draw several important findings, including: 1) no single model excels in all aspects, with different models showing various strengths; 2) the correlation between GPT-4 assessments and manual reviews is generally high; 3) there is a trade-off between the usability and safety of text-to-video generative models. This indicates that as the field of video generation rapidly advances, safety risks are set to surge, highlighting the urgency of prioritizing video safety. We hope that T2VSafetyBench can provide insights for better understanding the safety of video generation in the era of generative AIs. Our code is publicly available at \url{https://github.com/yibo-miao/T2VSafetyBench}.

Text-to-Image (T2I) models have made significant advancements in recent years, but they still struggle to accurately capture intricate details specified in complex compositional prompts. While fine-tuning T2I models with reward objectives has shown promise, it suffers from "reward hacking" and may not generalize well to unseen prompt distributions. In this work, we propose Reward-based Noise Optimization (ReNO), a novel approach that enhances T2I models at inference by optimizing the initial noise based on the signal from one or multiple human preference reward models. Remarkably, solving this optimization problem with gradient ascent for 50 iterations yields impressive results on four different one-step models across two competitive benchmarks, T2I-CompBench and GenEval. Within a computational budget of 20-50 seconds, ReNO-enhanced one-step models consistently surpass the performance of all current open-source Text-to-Image models. Extensive user studies demonstrate that our model is preferred nearly twice as often compared to the popular SDXL model and is on par with the proprietary Stable Diffusion 3 with 8B parameters. Moreover, given the same computational resources, a ReNO-optimized one-step model outperforms widely-used open-source models such as SDXL and PixArt-alpha, highlighting the efficiency and effectiveness of ReNO in enhancing T2I model performance at inference time.

Compositionality is a critical capability in Text-to-Image (T2I) models, as it reflects their ability to understand and combine multiple concepts from text descriptions. Existing evaluations of compositional capability rely heavily on human-designed text prompts or fixed templates, limiting their diversity and complexity, and yielding low discriminative power. We propose ConceptMix, a scalable, controllable, and customizable benchmark which automatically evaluates compositional generation ability of T2I models. This is done in two stages. First, ConceptMix generates the text prompts: concretely, using categories of visual concepts (e.g., objects, colors, shapes, spatial relationships), it randomly samples an object and k-tuples of visual concepts, then uses GPT-4o to generate text prompts for image generation based on these sampled concepts. Second, ConceptMix evaluates the images generated in response to these prompts: concretely, it checks how many of the k concepts actually appeared in the image by generating one question per visual concept and using a strong VLM to answer them. Through administering ConceptMix to a diverse set of T2I models (proprietary as well as open ones) using increasing values of k, we show that our ConceptMix has higher discrimination power than earlier benchmarks. Specifically, ConceptMix reveals that the performance of several models, especially open models, drops dramatically with increased k. Importantly, it also provides insight into the lack of prompt diversity in widely-used training datasets. Additionally, we conduct extensive human studies to validate the design of ConceptMix and compare our automatic grading with human judgement. We hope it will guide future T2I model development.

As generative large language models (LLMs) such as ChatGPT gain widespread adoption in various domains, their potential to propagate and amplify social biases, particularly in high-stakes areas such as the labor market, has become a pressing concern. AI algorithms are not only widely used in the selection of job applicants, individual job seekers may also make use of generative LLMs to help develop their job application materials. Against this backdrop, this research builds on a novel experimental design to examine social biases within ChatGPT-generated job applications in response to real job advertisements. By simulating the process of job application creation, we examine the language patterns and biases that emerge when the model is prompted with diverse job postings. Notably, we present a novel bias evaluation framework based on Masked Language Models to quantitatively assess social bias based on validated inventories of social cues/words, enabling a systematic analysis of the language used. Our findings show that the increasing adoption of generative AI, not only by employers but also increasingly by individual job seekers, can reinforce and exacerbate gender and social inequalities in the labor market through the use of biased and gendered language.

Fine-grained visual classification (FGVC) involves classifying closely related subcategories. This task is inherently difficult due to the subtle differences between classes and the high intra-class variance. Moreover, FGVC datasets are typically small and challenging to gather, thus highlighting a significant need for effective data augmentation. Recent advancements in text-to-image diffusion models have introduced new possibilities for data augmentation in image classification. While these models have been used to generate training data for classification tasks, their effectiveness in full-dataset training of FGVC models remains under-explored. Recent techniques that rely on text-to-image generation or Img2Img methods, such as SDEdit, often struggle to generate images that accurately represent the class while modifying them to a degree that significantly increases the dataset's diversity. To address these challenges, we present SaSPA: Structure and Subject Preserving Augmentation. Contrary to recent methods, our method does not use real images as guidance, thereby increasing generation flexibility and promoting greater diversity. To ensure accurate class representation, we employ conditioning mechanisms, specifically by conditioning on image edges and subject representation. We conduct extensive experiments and benchmark SaSPA against both traditional and generative data augmentation techniques. SaSPA consistently outperforms all established baselines across multiple settings, including full dataset training and contextual bias. Additionally, our results reveal interesting patterns in using synthetic data for FGVC models; for instance, we find a relationship between the amount of real data used and the optimal proportion of synthetic data.

Sign language translation (SLT) addresses the problem of translating information from a sign language in video to a spoken language in text. Existing studies, while showing progress, are often limited to narrow domains and/or few sign languages and struggle with open-domain tasks. In this paper, we push forward the frontier of SLT by scaling pretraining data, model size, and number of translation directions. We perform large-scale SLT pretraining on different data including 1) noisy multilingual Youtube SLT data, 2) parallel text corpora, and 3) SLT data augmented by translating video captions to other languages with off-the-shelf machine translation models. We unify different pretraining tasks with task-specific prompts under the encoder-decoder architecture, and initialize the SLT model with pretrained (m/By)T5 models across model sizes. SLT pretraining results on How2Sign and FLEURS-ASL\#0 (ASL to 42 spoken languages) demonstrate the significance of data/model scaling and cross-lingual cross-modal transfer, as well as the feasibility of zero-shot SLT. We finetune the pretrained SLT models on 5 downstream open-domain SLT benchmarks covering 5 sign languages. Experiments show substantial quality improvements over the vanilla baselines, surpassing the previous state-of-the-art (SOTA) by wide margins.

We present Meta 3D AssetGen (AssetGen), a significant advancement in text-to-3D generation which produces faithful, high-quality meshes with texture and material control. Compared to works that bake shading in the 3D object’s appearance, AssetGen outputs physically-based rendering (PBR) materials, supporting realistic relighting. AssetGen generates first several views of the object with separate shaded and albedo appearance channels, and then reconstructs colours, metalness and roughness in 3D, using a deferred shading loss for efficient supervision. It also uses a sign-distance function to represent 3D shape more reliably and introduces a corresponding loss for direct shape supervision. This is implemented using fused kernels for high memory efficiency. After mesh extraction, a texture refinement transformer operating in UV space significantly improves sharpness and details. AssetGen achieves 17% improvement in Chamfer Distance and 40% in LPIPS over the best concurrent work for few-view reconstruction, and a human preference of 72% over the best industry competitors of comparable speed, including those that support PBR. Project page with generated assets: https://assetgen.github.io

While following different technical routes, both low-rank and orthogonal adaptation techniques can efficiently adapt large-scale pre-training models in specific tasks or domains based on a small piece of trainable parameters. In this study, we bridge the gap between these two techniques, proposing a simple but effective adaptation method based on Householder reflections. Given a pre-trained model, our method fine-tunes its layers by multiplying each frozen weight matrix with an orthogonal matrix constructed by a chain of learnable Householder reflections (HRs). This HR-based orthogonal fine-tuning is equivalent to an adaptive low-rank adaptation. Moreover, we show that the orthogonality of the reflection planes corresponding to the HRs impacts the model capacity and regularity. The analysis motivates us to regularize the orthogonality of the HRs, leading to different implementations of the proposed Householder reflection adaptation (HRA) method. Compared with state-of-the-art methods, HRA achieves superior performance with fewer learnable parameters when adapting large language models and conditional image generators. The code of the experiments is available at https://github.com/DaShenZi721/HRA, and the method has been merged into the [PEFT](https://github.com/huggingface/peft) package.

Image generation has shown remarkable results in generating high-fidelity realistic images, in particular with the advancement of diffusion-based models. However, the prevalence of AI-generated images may have side effects for the machine learning community that are not clearly identified. Meanwhile, the success of deep learning in computer vision is driven by the massive dataset collected on the Internet. The extensive quantity of synthetic data being added to the Internet would become an obstacle for future researchers to collect "clean" datasets without AI-generated content. Prior research has shown that using datasets contaminated by synthetic images may result in performance degradation when used for training. In this paper, we investigate the potential impact of contaminated datasets on Online Continual Learning (CL) research. We experimentally show that contaminated datasets might hinder the training of existing online CL methods. Also, we propose Entropy Selection with Real-synthetic similarity Maximization (ESRM), a method to alleviate the performance deterioration caused by synthetic images when training online CL models. Experiments show that our method can significantly alleviate performance deterioration, especially when the contamination is severe. For reproducibility, the source code of our work is available at https://github.com/maorong-wang/ESRM.

Using image models naively for solving inverse video problems often suffers from flickering, texture-sticking, and temporal inconsistency in generated videos. To tackle these problems, in this paper, we view frames as continuous functions in the 2D space, and videos as a sequence of continuous warping transformations between different frames. This perspective allows us to train function space diffusion models only on **images** and utilize them to solve temporally correlated inverse problems. The function space diffusion models need to be equivariant with respect to the underlying spatial transformations. To ensure temporal consistency, we introduce a simple post-hoc test-time guidance towards (self)-equivariant solutions. Our method allows us to deploy state-of-the-art latent diffusion models such as Stable Diffusion XL to solve video inverse problems. We demonstrate the effectiveness of our method for video inpainting and $8\times$ video super-resolution, outperforming existing techniques based on noise transformations. We provide generated video results in the following URL: https://giannisdaras.github.io/warped_diffusion.github.io/.

We introduce a novel state-space architecture for diffusion models, effectively harnessing spatial and frequency information to enhance the inductive bias towards local features in input images for image generation tasks. While state-space networks, including Mamba, a revolutionary advancement in recurrent neural networks, typically scan input sequences from left to right, they face difficulties in designing effective scanning strategies, especially in the processing of image data. Our method demonstrates that integrating wavelet transformation into Mamba enhances the local structure awareness of visual inputs and better captures long-range relations of frequencies by disentangling them into wavelet subbands, representing both low- and high-frequency components. These wavelet-based outputs are then processed and seamlessly fused with the original Mamba outputs through a cross-attention fusion layer, combining both spatial and frequency information to optimize the order awareness of state-space models which is essential for the details and overall quality of image generation. Besides, we introduce a globally-shared transformer to supercharge the performance of Mamba, harnessing its exceptional power to capture global relationships. Through extensive experiments on standard benchmarks, our method demonstrates superior results compared to DiT and DIFFUSSM, achieving faster training convergence and delivering high-quality outputs. The codes and pretrained models are released at https://github.com/VinAIResearch/DiMSUM.git.

Discrete diffusion models have emerged as powerful tools for high-quality data generation. Despite their success in discrete spaces, such as text generation tasks, the acceleration of discrete diffusion models remains under-explored. In this paper, we propose discrete non-Markov diffusion models (DNDM), which naturally induce the predetermined transition time set. This enables a training-free sampling algorithm that significantly reduces the number of function evaluations (i.e., calls to the neural network), making the sampling process much faster. Furthermore, we study the transition from finite to infinite step sampling, offering new insights into bridging the gap between discrete and continuous-time processes for discrete diffusion models. Extensive experiments on natural language generation and machine translation tasks demonstrate the superior performance of our method in terms of both generation speed and sample quality compared to existing methods for discrete diffusion models. Codes are available at \url{https://github.com/uclaml/DNDM}.

Visual text design plays a critical role in conveying themes, emotions, and atmospheres in multimodal formats such as film posters and album covers. Translating these visual and textual elements across languages extends the concept of translation beyond mere text, requiring the adaptation of aesthetic and stylistic features. To address this, we introduce a novel task of Multimodal Style Translation (MuST-Bench), a benchmark designed to evaluate the ability of visual text generation models to perform translation across different writing systems while preserving design intent. Our initial experiments on MuST-Bench reveal that existing visual text generation models struggle with the proposed task due to the inadequacy of textual descriptions in conveying visual design. In response, we introduce SIGIL, a framework for multimodal style translation that eliminates the need for style descriptions. SIGIL enhances image generation models through three innovations: glyph latent for multilingual settings, pre-trained VAEs for stable style guidance, and an OCR model with reinforcement learning feedback for optimizing readable character generation. SIGIL outperforms existing baselines by achieving superior style consistency and legibility while maintaining visual fidelity, setting itself apart from traditional description-based approaches. We release MuST-Bench publicly for broader use and exploration https://huggingface.co/datasets/yejinc/MuST-Bench.

In domains of cybersecurity, recent advancements in Machine Learning (ML) and Deep Learning (DL) have significantly enhanced Network Intrusion Detection Systems (NIDS), improving the effectiveness of cybersecurity operations. However, attackers have also leveraged ML/DL to develop sophisticated models that generate adversarial packets capable of evading NIDS detection. Consequently, defenders must study and analyze these models to prepare for the evasion attacks that exploit NIDS detection mechanisms. Unfortunately, conventional generation models often rely on unrealistic assumptions about attackers' knowledge of NIDS components, making them impractical for real-world scenarios. To address this issue, we present DiffuPac, a first-of-its-kind generation model designed to generate adversarial packets that evade detection without relying on specific NIDS components. DiffuPac integrates a pre-trained Bidirectional Encoder Representations from Transformers (BERT) with diffusion model, which, through its capability for conditional denoising and classifier-free guidance, effectively addresses the real-world constraint of limited attacker knowledge. By concatenating malicious packets with contextually relevant normal packets and applying targeted noising only to the malicious packets, DiffuPac seamlessly blends adversarial packets into genuine network traffic. Through evaluations on real-world datasets, we demonstrate that DiffuPac achieves strong evasion capabilities against sophisticated NIDS, outperforming conventional methods by an average of 6.69 percentage points, while preserving the functionality and practicality of the generated adversarial packets.

Embodied AI agents require a fine-grained understanding of the physical world mediated through visual and language inputs. Such capabilities are difficult to learn solely from task-specific data. This has led to the emergence of pre-trained vision-language models as a tool for transferring representations learned from internet-scale data to downstream tasks and new domains. However, commonly used contrastively trained representations such as in CLIP have been shown to fail at enabling embodied agents to gain a sufficiently fine-grained scene understanding—a capability vital for control. To address this shortcoming, we consider representations from pre-trained text-to-image diffusion models, which are explicitly optimized to generate images from text prompts and as such, contain text-conditioned representations that reflect highly fine-grained visuo-spatial information. Using pre-trained text-to-image diffusion models, we construct Stable Control Representations which allow learning downstream control policies that generalize to complex, open-ended environments. We show that policies learned using Stable Control Representations are competitive with state-of-the-art representation learning approaches across a broad range of simulated control settings, encompassing challenging manipulation and navigation tasks. Most notably, we show that Stable Control Representations enable learning policies that exhibit state-of-the-art performance on OVMM, a difficult open-vocabulary navigation benchmark.

As text-to-image (T2I) synthesis models increase in size, they demand higher inference costs due to the need for more expensive GPUs with larger memory, which makes it challenging to reproduce these models in addition to the restricted access to training datasets. Our study aims to reduce these inference costs and explores how far the generative capabilities of T2I models can be extended using only publicly available datasets and open-source models. To this end, by using the de facto standard text-to-image model, Stable Diffusion XL (SDXL), we present three key practices in building an efficient T2I model: (1) Knowledge distillation: we explore how to effectively distill the generation capability of SDXL into an efficient U-Net and find that self-attention is the most crucial part. (2) Data: despite fewer samples, high-resolution images with rich captions are more crucial than a larger number of low-resolution images with short captions. (3) Teacher: Step-distilled Teacher allows T2I models to reduce the noising steps. Based on these findings, we build two types of efficient text-to-image models, called KOALA-Turbo & -Lightning, with two compact U-Nets (1B & 700M), reducing the model size up to 54% and 69% of the SDXL U-Net. In particular, the KOALA-Lightning-700M is 4 times faster than SDXL while still maintaining satisfactory generation quality. Moreover, unlike SDXL, our KOALA models can generate 1024px high-resolution images on consumer-grade GPUs with 8GB of VRAMs (3060Ti). We believe that our KOALA models will have a significant practical impact, serving as cost-effective alternatives to SDXL for academic researchers and general users in resource-constrained environments.