ICLR 2024 Papers

While generative diffusion models excel in producing high-quality images, they can also be misused to mimic authorized images, posing a significant threat to AI systems. Efforts have been made to add calibrated perturbations to protect images from diffusion-based mimicry pipelines. However, most of the existing methods are too ineffective and even impractical to be used by individual users due to their high computation and memory requirements. In this work, we present novel findings on attacking latent diffusion models (LDM) and propose new plug-and-play strategies for more effective protection. In particular, we explore the bottleneck in attacking an LDM, discovering that the encoder module rather than the denoiser module is the vulnerable point. Based on this insight, we present our strategy using Score Distillation Sampling (SDS) to double the speed of protection and reduce memory occupation by half without compromising its strength. Additionally, we provide a robust protection strategy by counterintuitively minimizing the semantic loss, which can assist in generating more natural perturbations. Finally, we conduct extensive experiments to substantiate our findings and comprehensively evaluate our newly proposed strategies. We hope our insights and protective measures can contribute to better defense against malicious diffusion-based mimicry, advancing the development of secure AI systems.

In the battle against widespread online misinformation, a growing problem is text-image inconsistency, where images are misleadingly paired with texts with different intent or meaning. Existing classification-based methods for text-image inconsistency can identify contextual inconsistencies but fail to provide explainable justifications for their decisions that humans can understand. Although more nuanced, human evaluation is impractical at scale and susceptible to errors. To address these limitations, this study introduces D-TIIL (Diffusion-based Text-Image Inconsistency Localization), which employs text-to-image diffusion models to localize semantic inconsistencies in text and image pairs. These models, trained on large-scale datasets act as ``omniscient" agents that filter out irrelevant information and incorporate background knowledge to identify inconsistencies. In addition, D-TIIL uses text embeddings and modified image regions to visualize these inconsistencies. To evaluate D-TIIL's efficacy, we introduce a new TIIL dataset containing 14K consistent and inconsistent text-image pairs. Unlike existing datasets, TIIL enables assessment at the level of individual words and image regions and is carefully designed to represent various inconsistencies. D-TIIL offers a scalable and evidence-based approach to identifying and localizing text-image inconsistency, providing a robust framework for future research combating misinformation.

Recent advances in generative diffusion models have enabled text-controlled synthesis of realistic and diverse images with impressive quality. Despite these remarkable advances, the application of text-to-image generative models in computer vision for standard visual recognition tasks remains limited. The current de facto approach for these tasks is to design model architectures and loss functions that are tailored to the task at hand. In this paper, we develop a unified language interface for computer vision tasks that abstracts away task specific design choices and enables task execution by following natural language instructions. Our approach involves casting multiple computer vision tasks as text-to-image generation problems. Here, the text represents an instruction describing the task, and the resulting image is a visually-encoded task output. To train our model, we pool commonly-used computer vision datasets covering a range of tasks, including segmentation, object detection, depth estimation, and classification. We then use a large language model to paraphrase prompt templates that convey the specific tasks to be conducted on each image, and through this process, we create a multi-modal and multi-task training dataset comprising input and output images along with annotated instructions. Following the InstructPix2Pix architecture, we apply instruction-tuning to a text-to-image diffusion model using our constructed dataset, steering its functionality from a generative model to an instruction-guided multi-task vision learner. Experiments demonstrate that our model, dubbed InstructCV, performs competitively compared to other generalist and task-specific vision models. Moreover, it exhibits compelling generalization capabilities to unseen data, categories, and user instructions.

Diffusion models are acclaimed for generating high-quality and diverse images. However, their performance notably degrades when trained on data with a long-tailed distribution. For long tail diffusion model generation, current works focus on the calibration and enhancement of the tail generation with head-tail knowledge transfer. The transfer process relies on the abundant diversity derived from the head class and, more significantly, the condition capacity of the model prediction. However, the dependency on the conditional model prediction to realize the knowledge transfer might exhibit bias during training, leading to unsatisfactory generation results and lack of robustness. Utilizing a Bayesian framework, we develop a weighted denoising score-matching technique for knowledge transfer directly from head to tail classes. Additionally, we incorporate a gating mechanism in the knowledge transfer process. We provide statistical analysis to validate this methodology, revealing that the effectiveness of such knowledge transfer depends on both label distribution and sample similarity, providing the insight to consider sample similarity when re-balancing the label proportion in training. We extensively evaluate our approach with experiments on multiple benchmark datasets, demonstrating its effectiveness and superior performance compared to existing methods. Code: \url{https://github.com/MediaBrain-SJTU/OC_LT}.

Advanced generative model (\textit{e.g.}, diffusion model) derived from simplified continuity assumptions of data distribution, though showing promising progress, has been difficult to apply directly to geometry generation applications due to the \textit{multi-modality} and \textit{noise-sensitive} nature of molecule geometry. This work introduces Geometric Bayesian Flow Networks (GeoBFN), which naturally fits molecule geometry by modeling diverse modalities in the differentiable parameter space of distributions. GeoBFN maintains the SE-(3) invariant density modeling property by incorporating equivariant inter-dependency modeling on parameters of distributions and unifying the probabilistic modeling of different modalities. Through optimized training and sampling techniques, we demonstrate that GeoBFN achieves state-of-the-art performance on multiple 3D molecule generation benchmarks in terms of generation quality (90.87\% molecule stability in QM9 and 85.6\% atom stability in GEOM-DRUG\footnote{The scores are reported at 1k sampling steps for fair comparison, and our scores could be further improved if sampling sufficiently longer steps.}). GeoBFN can also conduct sampling with any number of steps to reach an optimal trade-off between efficiency and quality (\textit{e.g.}, 20$\times$ speedup without sacrificing performance).

Diffusion models are the current state of the art for generating photorealistic images. Controlling the sampling process for constrained image generation tasks such as inpainting, however, remains challenging since exact conditioning on such constraints is intractable. While existing methods use various techniques to approximate the constrained posterior, this paper proposes to exploit the ability of Tractable Probabilistic Models (TPMs) to exactly and efficiently compute the constrained posterior, and to leverage this signal to steer the denoising process of diffusion models. Specifically, this paper adopts a class of expressive TPMs termed Probabilistic Circuits (PCs). Building upon prior advances, we further scale up PCs and make them capable of guiding the image generation process of diffusion models. Empirical results suggest that our approach can consistently improve the overall quality and semantic coherence of inpainted images across three natural image datasets (i.e., CelebA-HQ, ImageNet, and LSUN) with only ~10% additional computational overhead brought by the TPM. Further, with the help of an image encoder and decoder, our method can readily accept semantic constraints on specific regions of the image, which opens up the potential for more controlled image generation tasks. In addition to proposing a new framework for constrained image generation, this paper highlights the benefit of more tractable models and motivates the development of expressive TPMs.

Deep Ensemble approach is a straightforward technique used to enhance the performance of deep neural networks by training them from different initial points, converging towards various local optima. However, a limitation of this methodology lies in its high computational overhead for inference, arising from the necessity to store numerous learned parameters and execute individual forward passes for each parameter during the inference stage.We propose a novel approach called Diffusion Bridge Network to address this challenge. Based on the theory of Schr\"odinger bridge, this method directly learns to simulate an Stochastic Differential Equation (SDE) that connects the output distribution of a single ensemble member to the output distribution of the ensembled model, allowing us to obtain ensemble prediction without having to invoke forward pass through all the ensemble models. By substituting the heavy ensembles with this lightweight neural network constructing DBN, we achieved inference with reduced computational cost while maintaining accuracy and uncertainty scores on benchmark datasets such as CIFAR-10, CIFAR-100, and TinyImageNet.

Diffusion models generate highly realistic images by learning a multi-step denoising process, naturally embodying the principles of multi-task learning (MTL). Despite the inherent connection between diffusion models and MTL, there remains an unexplored area in designing neural architectures that explicitly incorporate MTL into the framework of diffusion models. In this paper, we present Denoising Task Routing (DTR), a simple add-on strategy for existing diffusion model architectures to establish distinct information pathways for individual tasks within a single architecture by selectively activating subsets of channels in the model. What makes DTR particularly compelling is its seamless integration of prior knowledge of denoising tasks into the framework: (1) Task Affinity: DTR activates similar channels for tasks at adjacent timesteps and shifts activated channels as sliding windows through timesteps, capitalizing on the inherent strong affinity between tasks at adjacent timesteps. (2) Task Weights: During the early stages (higher timesteps) of the denoising process, DTR assigns a greater number of task-specific channels, leveraging the insight that diffusion models prioritize reconstructing global structure and perceptually rich contents in earlier stages, and focus on simple noise removal in later stages. Our experiments reveal that DTR not only consistently boosts diffusion models' performance across different evaluation protocols without adding extra parameters but also accelerates training convergence. Finally, we show the complementarity between our architectural approach and existing MTL optimization techniques, providing a more complete view of MTL in the context of diffusion training. Significantly, by leveraging this complementarity, we attain matched performance of DiT-XL using the smaller DiT-L with a reduction in training iterations from 7M to 2M. Our project page is available at https://byeongjun-park.github.io/DTR/

In this paper, we present a novel diffusion model called SyncDreamer that generates multiview-consistent images from a single-view image. Using pretrained large-scale 2D diffusion models, recent work Zero123 demonstrates the ability to generate plausible novel views from a single-view image of an object. However, maintaining consistency in geometry and colors for the generated images remains a challenge. To address this issue, we propose a synchronized multiview diffusion model that models the joint probability distribution of multiview images, enabling the generation of multiview-consistent images in a single reverse process. SyncDreamer synchronizes the intermediate states of all the generated images at every step of the reverse process through a 3D-aware feature attention mechanism that correlates the corresponding features across different views. Experiments show that SyncDreamer generates images with high consistency across different views, thus making it well-suited for various 3D generation tasks such as novel-view-synthesis, text-to-3D, and image-to-3D. Project page: https://liuyuan-pal.github.io/SyncDreamer/.

We present Symphony, an $E(3)$ equivariant autoregressive generative model for 3D molecular geometries that iteratively builds a molecule from molecular fragments. Existing autoregressive models such as G-SchNet and G-SphereNet for molecules utilize rotationally invariant features to respect the 3D symmetries of molecules. In contrast, Symphony uses message-passing with higher-degree $E(3)$-equivariant features. This allows a novel representation of probability distributions via spherical harmonic signals to efficiently model the 3D geometry of molecules. We show that Symphony is able to accurately generate small molecules from the QM9 dataset, outperforming existing autoregressive models and approaching the performance of diffusion models. Our code is available at https://github.com/atomicarchitects/symphony.

Comprehending natural language instructions is a charming property for 3D indoor scene synthesis systems. Existing methods directly model object joint distributions and express object relations implicitly within a scene, thereby hindering the controllability of generation. We introduce InstructScene, a novel generative framework that integrates a semantic graph prior and a layout decoder to improve controllability and fidelity for 3D scene synthesis. The proposed semantic graph prior jointly learns scene appearances and layout distributions, exhibiting versatility across various downstream tasks in a zero-shot manner. To facilitate the benchmarking for text-driven 3D scene synthesis, we curate a high-quality dataset of scene-instruction pairs with large language and multimodal models. Extensive experimental results reveal that the proposed method surpasses existing state-of-the-art approaches by a large margin. Thorough ablation studies confirm the efficacy of crucial design components. Project page: https://chenguolin.github.io/projects/InstructScene.

We introduce a generative model with an intrinsically interpretable layer---a concept bottleneck layer---that constrains the model to encode human-understandable concepts. The concept bottleneck layer partitions the generative model into three parts: the pre-concept bottleneck portion, the CB layer, and the post-concept bottleneck portion. To train CB generative models, we complement the traditional task-based loss function for training generative models with a concept loss and an orthogonality loss. The CB layer and these loss terms are model agnostic, which we demonstrate by applying the CB layer to three different families of generative models: generative adversarial networks, variational autoencoders, and diffusion models. On multiple datasets across different types of generative models, steering a generative model, with the CB layer, outperforms all baselines---in some cases, it is \textit{10 times} more effective. In addition, we show how the CB layer can be used to interpret the output of the generative model and debug the model during or post training.

In light of the widespread success of generative models, a significant amount of research has gone into speeding up their sampling time. However, generative models are often sampled multiple times to obtain a diverse set incurring a cost that is orthogonal to sampling time. We tackle the question of how to improve diversity and sample efficiency by moving beyond the common assumption of independent samples. We propose particle guidance, an extension of diffusion-based generative sampling where a joint-particle time-evolving potential enforces diversity. We analyze theoretically the joint distribution that particle guidance generates, how to learn a potential that achieves optimal diversity, and the connections with methods in other disciplines. Empirically, we test the framework both in the setting of conditional image generation, where we are able to increase diversity without affecting quality, and molecular conformer generation, where we reduce the state-of-the-art median error by 13% on average.

Generative models have gained more and more attention in recent years for their remarkable success in tasks that required estimating and sampling data distribution to generate high-fidelity synthetic data. In speech, text-to-speech synthesis and neural vocoder are good examples where generative models have shined. While generative models have been applied to different applications in speech, there exists no general-purpose generative model that models speech directly. In this work, we take a step toward this direction by showing a single pre-trained generative model can be adapted to different downstream tasks with strong performance. Specifically, we pre-trained a generative model, named SpeechFlow, on 60k hours of untranscribed speech with Flow Matching and masked conditions. Experiment results show the pre-trained generative model can be fine-tuned with task-specific data to match or surpass existing expert models on speech enhancement, separation, and synthesis. Our work suggested a foundational model for generation tasks in speech can be built with generative pre-training.

Creating stable, controllable videos is a complex task due to the need for significant variation in temporal dynamics and cross-frame temporal consistency. To address this, we enhance the spatial-temporal capability and introduce a versatile video generation model, VersVideo, which leverages textual, visual, and stylistic conditions. Current video diffusion models typically extend image diffusion architectures by supplementing 2D operations (such as convolutions and attentions) with temporal operations. While this approach is efficient, it often restricts spatial-temporal performance due to the oversimplification of standard 3D operations. To counter this, we incorporate two key elements: (1) multi-excitation paths for spatial-temporal convolutions with dimension pooling across different axes, and (2) multi-expert spatial-temporal attention blocks. These enhancements boost the model's spatial-temporal performance without significantly escalating training and inference costs. We also tackle the issue of information loss that arises when a variational autoencoder is used to transform pixel space into latent features and then back into pixel frames. To mitigate this, we incorporate temporal modules into the decoder to maintain inter-frame consistency. Lastly, by utilizing the innovative denoising UNet and decoder, we develop a unified ControlNet model suitable for various conditions, including image, Canny, HED, depth, and style. Examples of the videos generated by our model can be found at https://jinxixiang.github.io/versvideo/.

Text-to-video editing aims to edit the visual appearance of a source video conditional on textual prompts. A major challenge in this task is to ensure that all frames in the edited video are visually consistent. Most recent works apply advanced text-to-image diffusion models to this task by inflating 2D spatial attention in the U-Net into spatio-temporal attention. Although temporal context can be added through spatio-temporal attention, it may introduce some irrelevant information for each patch and therefore cause inconsistency in the edited video. In this paper, for the first time, we introduce optical flow into the attention module in diffusion model's U-Net to address the inconsistency issue for text-to-video editing. Our method, FLATTEN, enforces the patches on the same flow path across different frames to attend to each other in the attention module, thus improving the visual consistency in the edited videos. Additionally, our method is training-free and can be seamlessly integrated into any diffusion based text-to-video editing methods and improve their visual consistency. Experiment results on existing text-to-video editing benchmarks show that our proposed method achieves the new state-of-the-art performance. In particular, our method excels in maintaining the visual consistency in the edited videos.

Deep generative models have made tremendous progress in modeling complex data, often exhibiting generation quality that surpasses a typical human's ability to discern the authenticity of samples. Undeniably, a key driver of this success is enabled by the massive amounts of web-scale data consumed by these models. Due to these models' striking performance and ease of availability, the web will inevitably be increasingly populated with synthetic content. Such a fact directly implies that future iterations of generative models will be trained on both clean and artificially generated data from past models. In this paper, we develop a framework to rigorously study the impact of training generative models on mixed datasets---from classical training on real data to self-consuming generative models trained on purely synthetic data. We first prove the stability of iterative training under the condition that the initial generative models approximate the data distribution well enough and the proportion of clean training data (w.r.t. synthetic data) is large enough. We empirically validate our theory on both synthetic and natural images by iteratively training normalizing flows and state-of-the-art diffusion models on CIFAR10 and FFHQ.

Score matching methods -- estimate probability densities without computing the normalization constant -- are particularly useful in deep learning. However, computational and memory costs of score matching methods can be prohibitive for high-dimensional data or complex models, particularly due to the derivatives or Hessians of the log density function appearing in the objective function. Some existing approaches modify the objective function to reduce the quadratic computational complexity for Hessian computation. However, the memory bottleneck of score matching methods remains for deep learning. This study improves the memory efficiency of score matching by leveraging deep equilibrium models. We provide a theoretical analysis of deep equilibrium models for scoring matching and applying implicit differentiation to higher-order derivatives. Empirical evaluations demonstrate that our approach enables the development of deep and expressive models with improved performance and comparable computational and memory costs over shallow architectures.

Flow matching is a powerful framework for generating high-quality samples in various applications, especially image synthesis. However, the intensive computational demands of these models, especially during the finetuning process and sampling processes, pose significant challenges for low-resource scenarios. This paper introduces Bellman Optimal Stepsize Straightening (BOSS) technique for distilling flow-matching generative models: it aims specifically for a few-step efficient image sampling while adhering to a computational budget constraint. First, this technique involves a dynamic programming algorithm that optimizes the stepsizes of the pretrained network. Then, it refines the velocity network to match the optimal step sizes, aiming to straighten the generation paths. Extensive experimental evaluations across image generation tasks demonstrate the efficacy of BOSS in terms of both resource utilization and image quality. Our results reveal that BOSS achieves substantial gains in efficiency while maintaining competitive sample quality, effectively bridging the gap between low-resource constraints and the demanding requirements of flow-matching generative models. Our paper also fortifies the responsible development of artificial intelligence, offering a more sustainable generative model that reduces computational costs and environmental footprints. Our code can be found at https://github.com/nguyenngocbaocmt02/BOSS.