NeurIPS 2024 Papers

We introduce Exo2Ego-V, a novel exocentric-to-egocentric diffusion-based video generation method for daily-life skilled human activities where sparse 4-view exocentric viewpoints are configured 360° around the scene. This task is particularly challenging due to the significant variations between exocentric and egocentric viewpoints and high complexity of dynamic motions and real-world daily-life environments. To address these challenges, we first propose a new diffusion-based multi-view exocentric encoder to extract the dense multi-scale features from multi-view exocentric videos as the appearance conditions for egocentric video generation. Then, we design an exocentric-to-egocentric view translation prior to provide spatially aligned egocentric features as a concatenation guidance for the input of egocentric video diffusion model. Finally, we introduce the temporal attention layers into our egocentric video diffusion pipeline to improve the temporal consistency cross egocentric frames. Extensive experiments demonstrate that Exo2Ego-V significantly outperforms SOTA approaches on 5 categories from the Ego-Exo4D dataset with an average of 35% in terms of LPIPS. Our code and model will be made available on https://github.com/showlab/Exo2Ego-V.

Recent advances in data-driven imitation learning and offline reinforcement learning have highlighted the use of expert data for skill acquisition and the development of hierarchical policies based on these skills. However, these approaches have not significantly advanced in adapting these skills to unseen contexts, which may involve changing environmental conditions or different user requirements. In this paper, we present a novel LLM-based policy adaptation framework LDuS which leverages an LLM to guide the generation process of a skill diffusion model upon contexts specified in language, facilitating zero-shot skill-based policy adaptation to different contexts. To implement the skill diffusion model, we adapt the loss-guided diffusion with a sequential in-painting technique, where target trajectories are conditioned by masking them with past state-action sequences, thereby enabling the robust and controlled generation of skill trajectories in test-time. To have a loss function for a given context, we employ the LLM-based code generation with iterative refinement, by which the code and controlled trajectory are validated to align with the context in a closed-loop manner. Through experiments, we demonstrate the zero-shot adaptability of LDuS to various context types including different specification levels, multi-modality, and varied temporal conditions for several robotic manipulation tasks, outperforming other language-conditioned imitation and planning methods.

The scale and quality of a dataset significantly impact the performance of deep models. However, acquiring large-scale annotated datasets is both a costly and time-consuming endeavor. To address this challenge, dataset expansion technologies aim to automatically augment datasets, unlocking the full potential of deep models. Current data expansion techniques include image transformation and image synthesis methods. Transformation-based methods introduce only local variations, leading to limited diversity. In contrast, synthesis-based methods generate entirely new content, greatly enhancing informativeness. However, existing synthesis methods carry the risk of distribution deviations, potentially degrading model performance with out-of-distribution samples. In this paper, we propose DistDiff, a training-free data expansion framework based on the distribution-aware diffusion model. DistDiff constructs hierarchical prototypes to approximate the real data distribution, optimizing latent data points within diffusion models with hierarchical energy guidance. We demonstrate its capability to generate distribution-consistent samples, significantly improving data expansion tasks. DistDiff consistently enhances accuracy across a diverse range of datasets compared to models trained solely on original data. Furthermore, our approach consistently outperforms existing synthesis-based techniques and demonstrates compatibility with widely adopted transformation-based augmentation methods. Additionally, the expanded dataset exhibits robustness across various architectural frameworks.

We introduce DiffAug, a simple and efficient diffusion-based augmentation technique to train image classifiers for the crucial yet challenging goal of improved classifier robustness. Applying DiffAug to a given example consists of one forward-diffusion step followed by one reverse-diffusion step. Using both ResNet-50 and Vision Transformer architectures, we comprehensively evaluate classifiers trained with DiffAug and demonstrate the surprising effectiveness of single-step reverse diffusion in improving robustness to covariate shifts, certified adversarial accuracy and out of distribution detection. When we combine DiffAug with other augmentations such as AugMix and DeepAugment we demonstrate further improved robustness. Finally, building on this approach, we also improve classifier-guided diffusion wherein we observe improvements in: (i) classifier-generalization, (ii) gradient quality (i.e., improved perceptual alignment) and (iii) image generation performance. We thus introduce a computationally efficient technique for training with improved robustness that does not require any additional data, and effectively complements existing augmentation approaches.

Non-autoregressive Transformers (NATs) are recently applied in direct speech-to-speech translation systems, which convert speech across different languages without intermediate text data. Although NATs generate high-quality outputs and offer faster inference than autoregressive models, they tend to produce incoherent and repetitive results due to complex data distribution (e.g., acoustic and linguistic variations in speech). In this work, we introduce DiffNorm, a diffusion-based normalization strategy that simplifies data distributions for training NAT models. After training with a self-supervised noise estimation objective, DiffNorm constructs normalized target data by denoising synthetically corrupted speech features. Additionally, we propose to regularize NATs with classifier-free guidance, improving model robustness and translation quality by randomly dropping out source information during training. Our strategies result in a notable improvement of about $+7$ ASR-BLEU for English-Spanish (En-Es) translation and $+2$ ASR-BLEU for English-French (En-Fr) on the CVSS benchmark, while attaining over $14\times$ speedup for En-Es and $5 \times$ speedup for En-Fr translations compared to autoregressive baselines.

Text-to-image generative models have recently attracted considerable interest, enabling the synthesis of high-quality images from textual prompts. However, these models often lack the capability to generate specific subjects from given reference images or to synthesize novel renditions under varying conditions. Methods like DreamBooth and Subject-driven Text-to-Image (SuTI) have made significant progress in this area. Yet, both approaches primarily focus on enhancing similarity to reference images and require expensive setups, often overlooking the need for efficient training and avoiding overfitting to the reference images. In this work, we present the $\lambda$-Harmonic reward function, which provides a reliable reward signal and enables early stopping for faster training and effective regularization. By combining the Bradley-Terry preference model, the $\lambda$-Harmonic reward function also provides preference labels for subject-driven generation tasks. We propose Reward Preference Optimization (RPO), which offers a simpler setup (requiring only 3\% of the negative samples used by DreamBooth) and fewer gradient steps for fine-tuning. Unlike most existing methods, our approach does not require training a text encoder or optimizing text embeddings and achieves text-image alignment by fine-tuning only the U-Net component. Empirically, $\lambda$-Harmonic proves to be a reliable approach for model selection in subject-driven generation tasks. Based on preference labels and early stopping validation from the $\lambda$-Harmonic reward function, our algorithm achieves a state-of-the-art CLIP-I score of 0.833 and a CLIP-T score of 0.314 on DreamBench.

The pre-trained text-to-image diffusion models have been increasingly employed to tackle the real-world image super-resolution (Real-ISR) problem due to their powerful generative image priors. Most of the existing methods start from random noise to reconstruct the high-quality (HQ) image under the guidance of the given low-quality (LQ) image. While promising results have been achieved, such Real-ISR methods require multiple diffusion steps to reproduce the HQ image, increasing the computational cost. Meanwhile, the random noise introduces uncertainty in the output, which is unfriendly to image restoration tasks. To address these issues, we propose a one-step effective diffusion network, namely OSEDiff, for the Real-ISR problem. We argue that the LQ image contains rich information to restore its HQ counterpart, and hence the given LQ image can be directly taken as the starting point for diffusion, eliminating the uncertainty introduced by random noise sampling. We finetune the pre-trained diffusion network with trainable layers to adapt it to complex image degradations. To ensure that the one-step diffusion model could yield HQ Real-ISR output, we apply variational score distillation in the latent space to conduct KL-divergence regularization. As a result, our OSEDiff model can efficiently and effectively generate HQ images in just one diffusion step. Our experiments demonstrate that OSEDiff achieves comparable or even better Real-ISR results, in terms of both objective metrics and subjective evaluations, than previous diffusion model-based Real-ISR methods that require dozens or hundreds of steps. The source codes are released at https://github.com/cswry/OSEDiff.

Advances in 3D reconstruction have enabled high-quality 3D capture, but require a user to collect hundreds to thousands of images to create a 3D scene. We present CAT3D, a method for creating anything in 3D by simulating this real-world capture process with a multi-view diffusion model. Given any number of input images and a set of target novel viewpoints, our model generates highly consistent novel views of a scene. These generated views can be used as input to robust 3D reconstruction techniques to produce 3D representations that can be rendered from any viewpoint in real-time. CAT3D can create entire 3D scenes in as little as one minute, and outperforms existing methods for single image and few-view 3D scene creation.

Certified defenses against adversarial attacks offer formal guarantees on the robustness of a model, making them more reliable than empirical methods such as adversarial training, whose effectiveness is often later reduced by unseen attacks. Still, the limited certified robustness that is currently achievable has been a bottleneck for their practical adoption. Gowal et al. and Wang et al. have shown that generating additional training data using state-of-the-art diffusion models can considerably improve the robustness of adversarial training. In this work, we demonstrate that a similar approach can substantially improve deterministic certified defenses but also reveal notable differences in the scaling behavior between certified and empirical methods. In addition, we provide a list of recommendations to scale the robustness of certified training approaches. Our approach achieves state-of-the-art deterministic robustness certificates on CIFAR-10 for the $\ell_2$ ($\epsilon = 36/255$) and $\ell_{\infty}$ ($\epsilon = 8/255$) threat models, outperforming the previous results by $+3.95$ and $+1.39$ percentage points, respectively. Furthermore, we report similar improvements for CIFAR-100.

Disentangled representation learning strives to extract the intrinsic factors within the observed data. Factoring these representations in an unsupervised manner is notably challenging and usually requires tailored loss functions or specific structural designs. In this paper, we introduce a new perspective and framework, demonstrating that diffusion models with cross-attention itself can serve as a powerful inductive bias to facilitate the learning of disentangled representations. We propose to encode an image into a set of concept tokens and treat them as the condition of the latent diffusion model for image reconstruction, where cross attention over the concept tokens is used to bridge the encoder and the U-Net of the diffusion model. We analyze that the diffusion process inherently possesses the time-varying information bottlenecks. Such information bottlenecks and cross attention act as strong inductive biases for promoting disentanglement. Without any regularization term in the loss function, this framework achieves superior disentanglement performance on the benchmark datasets, surpassing all previous methods with intricate designs. We have conducted comprehensive ablation studies and visualization analyses, shedding a light on the functioning of this model. We anticipate that our findings will inspire more investigation on exploring diffusion model for disentangled representation learning towards more sophisticated data analysis and understanding.

Prevalent text-to-video retrieval methods represent multimodal text-video data in a joint embedding space, aiming at bridging the relevant text-video pairs and pulling away irrelevant ones. One main challenge in state-of-the-art retrieval methods lies in the modality gap, which stems from the substantial disparities between text and video and can persist in the joint space. In this work, we leverage the potential of Diffusion models to address the text-video modality gap by progressively aligning text and video embeddings in a unified space. However, we identify two key limitations of existing Diffusion models in retrieval tasks: The L2 loss does not fit the ranking problem inherent in text-video retrieval, and the generation quality heavily depends on the varied initial point drawn from the isotropic Gaussian, causing inaccurate retrieval. To this end, we introduce a new Diffusion-Inspired Truncated Sampler (DITS) that jointly performs progressive alignment and modality gap modeling in the joint embedding space. The key innovation of DITS is to leverage the inherent proximity of text and video embeddings, defining a truncated diffusion flow from the fixed text embedding to the video embedding, enhancing controllability compared to adopting the isotropic Gaussian. Moreover, DITS adopts the contrastive loss to jointly consider the relevant and irrelevant pairs, not only facilitating alignment but also yielding a discriminatively structured embedding. Experiments on five benchmark datasets suggest the state-of-the-art performance of DITS. We empirically find that DITS can also improve the structure of the CLIP embedding space. Code is available at https://github.com/Jiamian- Wang/DITS-text-video-retrieval

This paper investigates score-based diffusion models when the underlying target distribution is concentrated on or near low-dimensional manifolds within the higher-dimensional space in which they formally reside, a common characteristic of natural image distributions. Despite previous efforts to understand the data generation process of diffusion models, existing theoretical support remains highly suboptimal in the presence of low-dimensional structure, which we strengthen in this paper. For the popular Denoising Diffusion Probabilistic Model (DDPM), we find that the dependency of the error incurred within each denoising step on the ambient dimension $d$ is in general unavoidable. We further identify a unique design of coefficients that yields a converges rate at the order of $O(k^{2}/\sqrt{T})$ (up to log factors), where $k$ is the intrinsic dimension of the target distribution and $T$ is the number of steps. This represents the first theoretical demonstration that the DDPM sampler can adapt to unknown low-dimensional structures in the target distribution, highlighting the critical importance of coefficient design. All of this is achieved by a novel set of analysis tools that characterize the algorithmic dynamics in a more deterministic manner.

In this work, we study the generalizability of diffusion models by looking into the hidden properties of the learned score functions, which are essentially a series of deep denoisers trained on various noise levels. We observe that as diffusion models transition from memorization to generalization, their corresponding nonlinear diffusion denoisers exhibit increasing linearity. This discovery leads us to investigate the linear counterparts of the nonlinear diffusion models, which are a series of linear models trained to match the function mappings of the nonlinear diffusion denoisers. Surprisingly, these linear denoisers are approximately the optimal denoisers for a multivariate Gaussian distribution characterized by the empirical mean and covariance of the training dataset. This finding implies that diffusion models have the inductive bias towards capturing and utilizing the Gaussian structure (covariance information) of the training dataset for data generation. We empirically demonstrate that this inductive bias is a unique property of diffusion models in the generalization regime, which becomes increasingly evident when the model's capacity is relatively small compared to the training dataset size. In the case that the model is highly overparameterized, this inductive bias emerges during the initial training phases before the model fully memorizes its training data. Our study provides crucial insights into understanding the notable strong generalization phenomenon recently observed in real-world diffusion models.

Large-scale diffusion models are adept at generating high-fidelity images and facilitating image editing and interpolation. However, they have limitations when tasked with generating images in dynamic, evolving domains. In this paper, we introduce Terra, a novel Time-varying low-rank adapter that offers a fine-tuning framework specifically tailored for domain flow generation. The key innovation of Terra lies in its construction of a continuous parameter manifold through a time variable, with its expressive power analyzed theoretically. This framework not only enables interpolation of image content and style but also offers a generation-based approach to address the domain shift problems in unsupervised domain adaptation and domain generalization. Specifically, Terra transforms images from the source domain to the target domain and generates interpolated domains with various styles to bridge the gap between domains and enhance the model generalization, respectively. We conduct extensive experiments on various benchmark datasets, empirically demonstrate the effectiveness of Terra. Our source code is publicly available on https://github.com/zwebzone/terra.

We introduce GrounDiT, a novel training-free spatial grounding technique for text-to-image generation using Diffusion Transformers (DiT). Spatial grounding with bounding boxes has gained attention for its simplicity and versatility, allowing for enhanced user control in image generation. However, prior training-free approaches often rely on updating the noisy image during the reverse diffusion process via backpropagation from custom loss functions, which frequently struggle to provide precise control over individual bounding boxes. In this work, we leverage the flexibility of the Transformer architecture, demonstrating that DiT can generate noisy patches corresponding to each bounding box, fully encoding the target object and allowing for fine-grained control over each region. Our approach builds on an intriguing property of DiT, which we refer to as semantic sharing. Due to semantic sharing, when a smaller patch is jointly denoised alongside a generatable-size image, the two become semantic clones. Each patch is denoised in its own branch of the generation process and then transplanted into the corresponding region of the original noisy image at each timestep, resulting in robust spatial grounding for each bounding box. In our experiments on the HRS and DrawBench benchmarks, we achieve state-of-the-art performance compared to previous training-free approaches. Project Page: https://groundit-diffusion.github.io/.

Diffusion Transformers (DiTs) introduce the transformer architecture to diffusion tasks for latent-space image generation. With an isotropic architecture that chains a series of transformer blocks, DiTs demonstrate competitive performance and good scalability; but meanwhile, the abandonment of U-Net by DiTs and their following improvements is worth rethinking. To this end, we conduct a simple toy experiment by comparing a U-Net architectured DiT with an isotropic one. It turns out that the U-Net architecture only gain a slight advantage amid the U-Net inductive bias, indicating potential redundancies within the U-Net-style DiT. Inspired by the discovery that U-Net backbone features are low-frequency-dominated, we perform token downsampling on the query-key-value tuple for self-attention and bring further improvements despite a considerable amount of reduction in computation. Based on self-attention with downsampled tokens, we propose a series of U-shaped DiTs (U-DiTs) in the paper and conduct extensive experiments to demonstrate the extraordinary performance of U-DiT models. The proposed U-DiT could outperform DiT-XL with only 1/6 of its computation cost. Codes are available at https://github.com/YuchuanTian/U-DiT.

Existing dynamic scene generation methods mostly rely on distilling knowledge from pre-trained 3D generative models, which are typically fine-tuned on synthetic object datasets. As a result, the generated scenes are often object-centric and lack photorealism. To address these limitations, we introduce a novel pipeline designed for photorealistic text-to-4D scene generation, discarding the dependency on multi-view generative models and instead fully utilizing video generative models trained on diverse real-world datasets. Our method begins by generating a reference video using the video generation model. We then learn the canonical 3D representation of the video using a freeze-time video, delicately generated from the reference video. To handle inconsistencies in the freeze-time video, we jointly learn a per-frame deformation to model these imperfections. We then learn the temporal deformation based on the canonical representation to capture dynamic interactions in the reference video. The pipeline facilitates the generation of dynamic scenes with enhanced photorealism and structural integrity, viewable from multiple perspectives, thereby setting a new standard in 4D scene generation.

We propose ID-to-3D, a method to generate identity- and text-guided 3D human heads with disentangled expressions, starting from even a single casually captured ‘in-the-wild’ image of a subject. The foundation of our approach is anchored in compositionality, alongside the use of task-specific 2D diffusion models as priors for optimization. First, we extend a foundational model with a lightweight expression-aware and ID-aware architecture, and create 2D priors for geometric and texture generation, via fine-tuning only 0.2% of its available training parameters. Then, we jointly leverage a neural parametric representation for the expression of each subject and a multi-stage generation of highly detailed geometry and albedo texture. This combination of strong face identity embeddings and our neural representation enables accurate reconstruction of not only facial features but also accessories and hair, and can be meshed to provide render-ready assets for gaming and telepresence. Our results achieve an unprecedented level of id-consistent and high-quality texture and geometry generation, generalizing to a ‘world’ of unseen 3D identities, without relying on large 3D captured datasets of human assets.

In a great number of tasks in science and engineering, the goal is to infer an unknown image from a small number of noisy measurements collected from a known forward model describing certain sensing or imaging modality. Due to resource constraints, this image reconstruction task is often extremely ill-posed, which necessitates the adoption of expressive prior information to regularize the solution space. Score-based diffusion models, thanks to its impressive empirical success, have emerged as an appealing candidate of an expressive prior in image reconstruction. In order to accommodate diverse tasks at once, it is of great interest to develop efficient, consistent and robust algorithms that incorporate unconditional score functions of an image prior distribution in conjunction with flexible choices of forward models. This work develops an algorithmic framework for employing score-based diffusion models as an expressive data prior in nonlinear inverse problems with general forward models. Motivated by the plug-and-play framework in the imaging community, we introduce a diffusion plug-and-play method (DPnP) that alternatively calls two samplers, a proximal consistency sampler based solely on the likelihood function of the forward model, and a denoising diffusion sampler based solely on the score functions of the image prior. The key insight is that denoising under white Gaussian noise can be solved rigorously via both stochastic (i.e., DDPM-type) and deterministic (i.e., DDIM-type) samplers using the same set of score functions trained for generation. We establish both asymptotic and non-asymptotic performance guarantees of DPnP, and provide numerical experiments to illustrate its promise in solving both linear and nonlinear image reconstruction tasks. To the best of our knowledge, DPnP is the first provably-robust posterior sampling method for nonlinear inverse problems using unconditional diffusion priors.