NeurIPS 2023 Papers

Recent advances in implicit neural representations have achieved impressive results by sampling and fusing individual points along sampling rays in the sampling space. However, due to the explosively growing sampling space, finely representing and synthesizing detailed textures remains a challenge for unbounded large-scale outdoor scenes. To alleviate the dilemma of using individual points to perceive the entire colossal space, we explore learning the surface distribution of the scene to provide structural priors and reduce the samplable space and propose a Point Diffusion implicit Function, PDF, for large-scale scene neural representation. The core of our method is a large-scale point cloud super-resolution diffusion module that enhances the sparse point cloud reconstructed from several training images into a dense point cloud as an explicit prior. Then in the rendering stage, only sampling points with prior points within the sampling radius are retained. That is, the sampling space is reduced from the unbounded space to the scene surface. Meanwhile, to fill in the background of the scene that cannot be provided by point clouds, the region sampling based on Mip-NeRF 360 is employed to model the background representation. Expensive experiments have demonstrated the effectiveness of our method for large-scale scene novel view synthesis, which outperforms relevant state-of-the-art baselines.

Denoising diffusion models have recently emerged as the predominant paradigm for generative modelling on image domains. In addition, their extension to Riemannian manifolds has facilitated a range of applications across the natural sciences. While many of these problems stand to benefit from the ability to specify arbitrary, domain-informed constraints, this setting is not covered by the existing (Riemannian) diffusion model methodology. Recent work has attempted to address this issue by constructing novel noising processes based on the reflected Brownian motion and logarithmic barrier methods. However, the associated samplers are either computationally burdensome or only apply to convex subsets of Euclidean space. In this paper, we introduce an alternative, simple noising scheme based on Metropolis sampling that affords substantial gains in computational efficiency and empirical performance compared to the earlier samplers. Of independent interest, we prove that this new process corresponds to a valid discretisation of the reflected Brownian motion. We demonstrate the scalability and flexibility of our approach on a range of problem settings with convex and non-convex constraints, including applications from geospatial modelling, robotics and protein design.

Diffusion models have risen a promising approach to data-driven planning, and have demonstrated impressive robotic control, reinforcement learning, and video planning performance. Given an effective planner, an important question to consider is replanning -- when given plans should be regenerated due to both action execution error and external environment changes. Direct plan execution, without replanning, is problematic as errors from individual actions rapidly accumulate and environments are partially observable and stochastic. Simultaneously, replanning at each timestep incurs a substantial computational cost, and may prevent successful task execution, as different generated plans prevent consistent progress to any particular goal. In this paper, we explore how we may effectively replan with diffusion models. We propose a principled approach to determine when to replan, based on the diffusion model's estimated likelihood of existing generated plans. We further present an approach to replan existing trajectories to ensure that new plans follow the same goal state as the original trajectory, which may efficiently bootstrap off previously generated plans. We illustrate how a combination of our proposed additions significantly improves the performance of diffusion planners leading to 38\% gains over past diffusion planning approaches on Maze2D and further enables handling of stochastic and long-horizon robotic control tasks.

Existing score-distilling text-to-3D generation techniques, despite their considerable promise, often encounter the view inconsistency problem. One of the most notable issues is the Janus problem, where the most canonical view of an object (\textit{e.g}., face or head) appears in other views. In this work, we explore existing frameworks for score-distilling text-to-3D generation and identify the main causes of the view inconsistency problem---the embedded bias of 2D diffusion models. Based on these findings, we propose two approaches to debias the score-distillation frameworks for view-consistent text-to-3D generation. Our first approach, called score debiasing, involves cutting off the score estimated by 2D diffusion models and gradually increasing the truncation value throughout the optimization process. Our second approach, called prompt debiasing, identifies conflicting words between user prompts and view prompts using a language model, and adjusts the discrepancy between view prompts and the viewing direction of an object. Our experimental results show that our methods improve the realism of the generated 3D objects by significantly reducing artifacts and achieve a good trade-off between faithfulness to the 2D diffusion models and 3D consistency with little overhead. Our project page is available at~\url{https://susunghong.github.io/Debiased-Score-Distillation-Sampling/}.

Text-to-image generation has recently witnessed remarkable achievements. We introduce a text-conditional image diffusion model, termed RAPHAEL, to generate highly artistic images, which accurately portray the text prompts, encompassing multiple nouns, adjectives, and verbs. This is achieved by stacking tens of mixture-of-experts (MoEs) layers, i.e., space-MoE and time-MoE layers, enabling billions of diffusion paths (routes) from the network input to the output. Each path intuitively functions as a "painter" for depicting a particular textual concept onto a specified image region at a diffusion timestep. Comprehensive experiments reveal that RAPHAEL outperforms recent cutting-edge models, such as Stable Diffusion, ERNIE-ViLG 2.0, DeepFloyd, and DALL-E 2, in terms of both image quality and aesthetic appeal. Firstly, RAPHAEL exhibits superior performance in switching images across diverse styles, such as Japanese comics, realism, cyberpunk, and ink illustration. Secondly, a single model with three billion parameters, trained on 1,000 A100 GPUs for two months, achieves a state-of-the-art zero-shot FID score of 6.61 on the COCO dataset. Furthermore, RAPHAEL significantly surpasses its counterparts in human evaluation on the ViLG-300 benchmark. We believe that RAPHAEL holds the potential to propel the frontiers of image generation research in both academia and industry, paving the way for future breakthroughs in this rapidly evolving field. More details can be found on a webpage: https://raphael-painter.github.io/.

Denoising diffusion probabilistic models have transformed image generation with their impressive fidelity and diversity. We show that they also excel in estimating optical flow and monocular depth, surprisingly without task-specific architectures and loss functions that are predominant for these tasks. Compared to the point estimates of conventional regression-based methods, diffusion models also enable Monte Carlo inference, e.g., capturing uncertainty and ambiguity in flow and depth. With self-supervised pre-training, the combined use of synthetic and real data for supervised training, and technical innovations (infilling and step-unrolled denoising diffusion training) to handle noisy-incomplete training data, one can train state-of-the-art diffusion models for depth and optical flow estimation, with additional zero-shot coarse-to-fine refinement for high resolution estimates. Extensive experiments focus on quantitative performance against benchmarks, ablations, and the model's ability to capture uncertainty and multimodality, and impute missing values. Our model obtains a state-of-the-art relative depth error of 0.074 on the indoor NYU benchmark and an Fl-all score of 3.26\% on the KITTI optical flow benchmark, about 25\% better than the best published method.

Diffusion models are powerful, but they require a lot of time and data to train. We propose Patch Diffusion, a generic patch-wise training framework, to significantly reduce the training time costs while improving data efficiency, which thus helps democratize diffusion model training to broader users. At the core of our innovations is a new conditional score function at the patch level, where the patch location in the original image is included as additional coordinate channels, while the patch size is randomized and diversified throughout training to encode the cross-region dependency at multiple scales. Sampling with our method is as easy as in the original diffusion model. Through Patch Diffusion, we could achieve $\mathbf{\ge 2\times}$ faster training, while maintaining comparable or better generation quality. Patch Diffusion meanwhile improves the performance of diffusion models trained on relatively small datasets, $e.g.$, as few as 5,000 images to train from scratch. We achieve outstanding FID scores in line with state-of-the-art benchmarks: 1.77 on CelebA-64$\times$64, 1.93 on AFHQv2-Wild-64$\times$64, and 2.72 on ImageNet-256$\times$256. We share our code and pre-trained models at https://github.com/Zhendong-Wang/Patch-Diffusion.

Reconstructing human vision from brain activities has been an appealing task that helps to understand our cognitive process. Even though recent research has seen great success in reconstructing static images from non-invasive brain recordings, work on recovering continuous visual experiences in the form of videos is limited. In this work, we propose Mind-Video that learns spatiotemporal information from continuous fMRI data of the cerebral cortex progressively through masked brain modeling, multimodal contrastive learning with spatiotemporal attention, and co-training with an augmented Stable Diffusion model that incorporates network temporal inflation. We show that high-quality videos of arbitrary frame rates can be reconstructed with Mind-Video using adversarial guidance. The recovered videos were evaluated with various semantic and pixel-level metrics. We achieved an average accuracy of 85% in semantic classification tasks and 0.19 in structural similarity index (SSIM), outperforming the previous state-of-the-art by 45%. We also show that our model is biologically plausible and interpretable, reflecting established physiological processes.

Diffusion probabilistic models (DPMs) have demonstrated a very promising ability in high-resolution image synthesis. However, sampling from a pre-trained DPM is time-consuming due to the multiple evaluations of the denoising network, making it more and more important to accelerate the sampling of DPMs. Despite recent progress in designing fast samplers, existing methods still cannot generate satisfying images in many applications where fewer steps (e.g., $<$10) are favored. In this paper, we develop a unified corrector (UniC) that can be applied after any existing DPM sampler to increase the order of accuracy without extra model evaluations, and derive a unified predictor (UniP) that supports arbitrary order as a byproduct. Combining UniP and UniC, we propose a unified predictor-corrector framework called UniPC for the fast sampling of DPMs, which has a unified analytical form for any order and can significantly improve the sampling quality over previous methods, especially in extremely few steps. We evaluate our methods through extensive experiments including both unconditional and conditional sampling using pixel-space and latent-space DPMs. Our UniPC can achieve 3.87 FID on CIFAR10 (unconditional) and 7.51 FID on ImageNet 256$\times$256 (conditional) with only 10 function evaluations. Code is available at https://github.com/wl-zhao/UniPC.

Generating animation of physics-based characters with intuitive control has long been a desirable task with numerous applications. However, generating physically simulated animations that reflect high-level human instructions remains a difficult problem due to the complexity of physical environments and the richness of human language. In this paper, we present $\textbf{InsActor}$, a principled generative framework that leverages recent advancements in diffusion-based human motion models to produce instruction-driven animations of physics-based characters. Our framework empowers InsActor to capture complex relationships between high-level human instructions and character motions by employing diffusion policies for flexibly conditioned motion planning. To overcome invalid states and infeasible state transitions in planned motions, InsActor discovers low-level skills and maps plans to latent skill sequences in a compact latent space. Extensive experiments demonstrate that InsActor achieves state-of-the-art results on various tasks, including instruction-driven motion generation and instruction-driven waypoint heading. Notably, the ability of InsActor to generate physically simulated animations using high-level human instructions makes it a valuable tool, particularly in executing long-horizon tasks with a rich set of instructions. Our project page is available at [jiawei-ren.github.io/projects/insactor/index.html](https://jiawei-ren.github.io/projects/insactor/index.html)

The existence of adversarial examples has been a mystery for years and attracted much interest. A well-known theory by \citet{ilyas2019adversarial} explains adversarial vulnerability from a data perspective by showing that one can extract non-robust features from adversarial examples and these features alone are useful for classification. However, the explanation remains quite counter-intuitive since non-robust features are mostly noise features to humans. In this paper, we re-examine the theory from a larger context by incorporating multiple learning paradigms. Notably, we find that contrary to their good usefulness under supervised learning, non-robust features attain poor usefulness when transferred to other self-supervised learning paradigms, such as contrastive learning, masked image modeling, and diffusion models. It reveals that non-robust features are not really as useful as robust or natural features that enjoy good transferability between these paradigms. Meanwhile, for robustness, we also show that naturally trained encoders from robust features are largely non-robust under AutoAttack. Our cross-paradigm examination suggests that the non-robust features are not really useful but more like paradigm-wise shortcuts, and robust features alone might be insufficient to attain reliable model robustness. Code is available at \url{https://github.com/PKU-ML/AdvNotRealFeatures}.

In this work, we investigate the task of text-to-image (T2I) synthesis under the abstract-to-intricate setting, i.e., generating intricate visual content from simple abstract text prompts. Inspired by human imagination intuition, we propose a novel scene-graph hallucination (SGH) mechanism for effective abstract-to-intricate T2I synthesis. SGH carries out scene hallucination by expanding the initial scene graph (SG) of the input prompt with more feasible specific scene structures, in which the structured semantic representation of SG ensures high controllability of the intrinsic scene imagination. To approach the T2I synthesis, we deliberately build an SG-based hallucination diffusion system. First, we implement the SGH module based on the discrete diffusion technique, which evolves the SG structure by iteratively adding new scene elements. Then, we utilize another continuous-state diffusion model as the T2I synthesizer, where the overt image-generating process is navigated by the underlying semantic scene structure induced from the SGH module. On the benchmark COCO dataset, our system outperforms the existing best-performing T2I model by a significant margin, especially improving on the abstract-to-intricate T2I generation. Further in-depth analyses reveal how our methods advance.

The generation of 3D molecules requires simultaneously deciding the categorical features (atom types) and continuous features (atom coordinates). Deep generative models, especially Diffusion Models (DMs), have demonstrated effectiveness in generating feature-rich geometries. However, existing DMs typically suffer from unstable probability dynamics with inefficient sampling speed. In this paper, we introduce geometric flow matching, which enjoys the advantages of both equivariant modeling and stabilized probability dynamics. More specifically, we propose a hybrid probability path where the coordinates probability path is regularized by an equivariant optimal transport, and the information between different modalities is aligned. Experimentally, the proposed method could consistently achieve better performance on multiple molecule generation benchmarks with 4.75$\times$ speed up of sampling on average.

Diffusion models are a class of generative models that serve to establish a stochastic transport map between an empirically observed, yet unknown, target distribution and a known prior. Despite their remarkable success in real-world applications, a theoretical understanding of their generalization capabilities remains underdeveloped. This work embarks on a comprehensive theoretical exploration of the generalization attributes of diffusion models. We establish the theoretical estimates of the generalization gap that evolves in tandem with the training dynamics of score-based diffusion models, suggesting a polynomially small generalization error ($O(n^{-2/5}+m^{-4/5})$) on both the sample size $n$ and the model capacity $m$, evading the curse of dimensionality (i.e., independent of the data dimension) when *early-stopped*. Furthermore, we extend our quantitative analysis to a *data-dependent* scenario, wherein target distributions are portrayed as a succession of densities with progressively increasing distances between modes. This precisely elucidates the *adverse* effect of "*modes shift*'' in ground truths on the model generalization. Furthermore, these estimates are not solely theoretical constructs but have also been confirmed through numerical simulations. Our findings contribute to the rigorous understanding of diffusion models' generalization properties and provide insights that may guide practical applications.

The pursuit of controllability as a higher standard of visual content creation has yielded remarkable progress in customizable image synthesis. However, achieving controllable video synthesis remains challenging due to the large variation of temporal dynamics and the requirement of cross-frame temporal consistency. Based on the paradigm of compositional generation, this work presents VideoComposer that allows users to flexibly compose a video with textual conditions, spatial conditions, and more importantly temporal conditions. Specifically, considering the characteristic of video data, we introduce the motion vector from compressed videos as an explicit control signal to provide guidance regarding temporal dynamics. In addition, we develop a Spatio-Temporal Condition encoder (STC-encoder) that serves as a unified interface to effectively incorporate the spatial and temporal relations of sequential inputs, with which the model could make better use of temporal conditions and hence achieve higher inter-frame consistency. Extensive experimental results suggest that VideoComposer is able to control the spatial and temporal patterns simultaneously within a synthesized video in various forms, such as text description, sketch sequence, reference video, or even simply hand-crafted motions. The code and models are publicly available at https://videocomposer.github.io.

Synthesizing images with user-specified subjects has received growing attention due to its practical applications. Despite the recent success in single subject customization, existing algorithms suffer from high training cost and low success rate along with increased number of subjects. Towards controllable image synthesis with multiple subjects as the constraints, this work studies how to efficiently represent a particular subject as well as how to appropriately compose different subjects. We find that the text embedding regarding the subject token already serves as a simple yet effective representation that supports arbitrary combinations without any model tuning. Through learning a residual on top of the base embedding, we manage to robustly shift the raw subject to the customized subject given various text conditions. We then propose to employ layout, a very abstract and easy-to-obtain prior, as the spatial guidance for subject arrangement. By rectifying the activations in the cross-attention map, the layout appoints and separates the location of different subjects in the image, significantly alleviating the interference across them. Using cross-attention map as the intermediary, we could strengthen the signal of target subjects and weaken the signal of irrelevant subjects within a certain region, significantly alleviating the interference across subjects. Both qualitative and quantitative experimental results demonstrate our superiority over state-of-the-art alternatives under a variety of settings for multi-subject customization.

Large-scale generative models such as GPT and DALL-E have revolutionized the research community. These models not only generate high fidelity outputs, but are also generalists which can solve tasks not explicitly taught. In contrast, speech generative models are still primitive in terms of scale and task generalization. In this paper, we present Voicebox, the most versatile text-guided generative model for speech at scale. Voicebox is a non-autoregressive flow-matching model trained to infill speech, given audio context and text, trained on over 50K hours of speech that are not filtered or enhanced. Similar to GPT, Voicebox can perform many different tasks through in-context learning, but is more flexible as it can also condition on future context. Voicebox can be used for mono or cross-lingual zero-shot text-to-speech synthesis, noise removal, content editing, style conversion, and diverse sample generation. In particular, Voicebox outperforms the state-of-the-art zero-shot TTS model VALL-E on both intelligibility (5.9\% vs 1.9\% word error rates) and audio similarity (0.580 vs 0.681) while being up to 20 times faster. Audio samples can be found in \url{https://voicebox.metademolab.com}.

Denoising diffusion models are a powerful type of generative models used to capture complex distributions of real-world signals. However, their applicability is limited to scenarios where training samples are readily available, which is not always the case in real-world applications. For example, in inverse graphics, the goal is to generate samples from a distribution of 3D scenes that align with a given image, but ground-truth 3D scenes are unavailable and only 2D images are accessible. To address this limitation, we propose a novel class of denoising diffusion probabilistic models that learn to sample from distributions of signals that are never directly observed. Instead, these signals are measured indirectly through a known differentiable forward model, which produces partial observations of the unknown signal. Our approach involves integrating the forward model directly into the denoising process. A key contribution of our work is the integration of a differentiable forward model into the denoising process. This integration effectively connects the generative modeling of observations with the generative modeling of the underlying signals, allowing for end-to-end training of a conditional generative model over signals. During inference, our approach enables sampling from the distribution of underlying signals that are consistent with a given partial observation. We demonstrate the effectiveness of our method on three challenging computer vision tasks. For instance, in the context of inverse graphics, our model enables direct sampling from the distribution of 3D scenes that align with a single 2D input image.

Semi-implicit variational inference (SIVI) has been introduced to expand the analytical variational families by defining expressive semi-implicit distributions in a hierarchical manner. However, the single-layer architecture commonly used in current SIVI methods can be insufficient when the target posterior has complicated structures. In this paper, we propose hierarchical semi-implicit variational inference, called HSIVI, which generalizes SIVI to allow more expressive multi-layer construction of semi-implicit distributions. By introducing auxiliary distributions that interpolate between a simple base distribution and the target distribution, the conditional layers can be trained by progressively matching these auxiliary distributions one layer after another. Moreover, given pre-trained score networks, HSIVI can be used to accelerate the sampling process of diffusion models with the score matching objective. We show that HSIVI significantly enhances the expressiveness of SIVI on several Bayesian inference problems with complicated target distributions. When used for diffusion model acceleration, we show that HSIVI can produce high quality samples comparable to or better than the existing fast diffusion model based samplers with a small number of function evaluations on various datasets.