Face recognition technology presents serious privacy risks due to its reliance on sensitive and immutable biometric data. To address these concerns, such systems typically convert raw facial images into embeddings, which are traditionally viewed as privacy-preserving. However, model inversion attacks challenge this assumption by reconstructing private facial images from embeddings, highlighting a critical vulnerability in face recognition systems. Most existing inversion methods require training a separate generator for each target model, making them computationally intensive. In this work, we introduce DiffUMI, a diffusion-based universal model inversion attack that requires no additional training. DiffUMI is the first approach to successfully leverage unconditional face generation without relying on model-specific generators. It surpasses state-of-the-art attacks by 15.5% and 9.82% in success rate on standard and privacy-preserving face recognition systems, respectively. Furthermore, we propose a novel use of out-of-domain detection (OODD), demonstrating for the first time that model inversion can differentiate between facial and non-facial embeddings using only the embedding space.
Recently, pre-trained text-to-image (T2I) models have been extensively adopted for real-world image restoration because of their powerful generative prior. However, controlling these large models for image restoration usually requires a large number of high-quality images and immense computational resources for training, which is costly and not privacy-friendly. In this paper, we find that the well-trained large T2I model (i.e., Flux) is able to produce a variety of high-quality images aligned with real-world distributions, offering an unlimited supply of training samples to mitigate the above issue. Specifically, we proposed a training data construction pipeline for image restoration, namely FluxGen, which includes unconditional image generation, image selection, and degraded image simulation. A novel light-weighted adapter (FluxIR) with squeeze-and-excitation layers is also carefully designed to control the large Diffusion Transformer (DiT)-based T2I model so that reasonable details can be restored. Experiments demonstrate that our proposed method enables the Flux model to adapt effectively to real-world image restoration tasks, achieving superior scores and visual quality on both synthetic and real-world degradation datasets - at only about 8.5\% of the training cost compared to current approaches.
Guidance techniques are commonly used in diffusion and flow models to improve image quality and consistency for conditional generative tasks such as class-conditional and text-to-image generation. In particular, classifier-free guidance (CFG) -- the most widely adopted guidance technique -- contrasts conditional and unconditional predictions to improve the generated images. This results, however, in trade-offs across quality, diversity and consistency, improving some at the expense of others. While recent work has shown that it is possible to disentangle these factors to some extent, such methods come with an overhead of requiring an additional (weaker) model, or require more forward passes per sampling step. In this paper, we propose Entropy Rectifying Guidance (ERG), a simple and effective guidance mechanism based on inference-time changes in the attention mechanism of state-of-the-art diffusion transformer architectures, which allows for simultaneous improvements over image quality, diversity and prompt consistency. ERG is more general than CFG and similar guidance techniques, as it extends to unconditional sampling. ERG results in significant improvements in various generation tasks such as text-to-image, class-conditional and unconditional image generation. We also show that ERG can be seamlessly combined with other recent guidance methods such as CADS and APG, further boosting generation performance.
The bridge problem is to find an SDE (or sometimes an ODE) that bridges two given distributions. The application areas of the bridge problem are enormous, among which the recent generative modeling (e.g., conditional or unconditional image generation) is the most popular. Also the famous Schr\"{o}dinger bridge problem, a widely known problem for a century, is a special instance of the bridge problem. Two most popular algorithms to tackle the bridge problems in the deep learning era are: (conditional) flow matching and iterative fitting algorithms, where the former confined to ODE solutions, and the latter specifically for the Schr\"{o}dinger bridge problem. The main contribution of this article is in two folds: i) We provide concise reviews of these algorithms with technical details to some extent; ii) We propose a novel unified perspective and framework that subsumes these seemingly unrelated algorithms (and their variants) into one. In particular, we show that our unified framework can instantiate the Flow Matching (FM) algorithm, the (mini-batch) optimal transport FM algorithm, the (mini-batch) Schr\"{o}dinger bridge FM algorithm, and the deep Schr\"{o}dinger bridge matching (DSBM) algorithm as its special cases. We believe that this unified framework will be useful for viewing the bridge problems in a more general and flexible perspective, and in turn can help researchers and practitioners to develop new bridge algorithms in their fields.
While representation learning and generative modeling seek to understand visual data, unifying both domains remains unexplored. Recent Unified Self-Supervised Learning (SSL) methods have started to bridge the gap between both paradigms. However, they rely solely on semantic token reconstruction, which requires an external tokenizer during training -- introducing a significant overhead. In this work, we introduce Sorcen, a novel unified SSL framework, incorporating a synergic Contrastive-Reconstruction objective. Our Contrastive objective, "Echo Contrast", leverages the generative capabilities of Sorcen, eliminating the need for additional image crops or augmentations during training. Sorcen "generates" an echo sample in the semantic token space, forming the contrastive positive pair. Sorcen operates exclusively on precomputed tokens, eliminating the need for an online token transformation during training, thereby significantly reducing computational overhead. Extensive experiments on ImageNet-1k demonstrate that Sorcen outperforms the previous Unified SSL SoTA by 0.4%, 1.48 FID, 1.76%, and 1.53% on linear probing, unconditional image generation, few-shot learning, and transfer learning, respectively, while being 60.8% more efficient. Additionally, Sorcen surpasses previous single-crop MIM SoTA in linear probing and achieves SoTA performance in unconditional image generation, highlighting significant improvements and breakthroughs in Unified SSL models.
There is growing concern over the safety of powerful diffusion models (DMs), as they are often misused to produce inappropriate, not-safe-for-work (NSFW) content or generate copyrighted material or data of individuals who wish to be forgotten. Many existing methods tackle these issues by heavily relying on text-based negative prompts or extensively retraining DMs to eliminate certain features or samples. In this paper, we take a radically different approach, directly modifying the sampling trajectory by leveraging a negation set (e.g., unsafe images, copyrighted data, or datapoints needed to be excluded) to avoid specific regions of data distribution, without needing to retrain or fine-tune DMs. We formally derive the relationship between the expected denoised samples that are safe and those that are not safe, leading to our $\textit{safe}$ denoiser which ensures its final samples are away from the area to be negated. Inspired by the derivation, we develop a practical algorithm that successfully produces high-quality samples while avoiding negation areas of the data distribution in text-conditional, class-conditional, and unconditional image generation scenarios. These results hint at the great potential of our training-free safe denoiser for using DMs more safely.
Diffusionmodels(DMs)havedemonstratedremarkableachievements in synthesizing images of high fidelity and diversity. However, the extensive computational requirements and slow generative speed of diffusion models have limited their widespread adoption. In this paper, we propose a novel post-training quantization for diffusion models (PQD), which is a time-aware optimization framework for diffusion models based on post-training quantization. The proposed framework optimizes the inference process by selecting representative samples and conducting time-aware calibration. Experimental results show that our proposed method is able to directly quantize full-precision diffusion models into 8-bit or 4-bit models while maintaining comparable performance in a training-free manner, achieving a few FID change on ImageNet for unconditional image generation. Our approach demonstrates compatibility and can also be applied to 512x512 text-guided image generation for the first time.
Normalizing Flows (NFs) are likelihood-based models for continuous inputs. They have demonstrated promising results on both density estimation and generative modeling tasks, but have received relatively little attention in recent years. In this work, we demonstrate that NFs are more powerful than previously believed. We present TarFlow: a simple and scalable architecture that enables highly performant NF models. TarFlow can be thought of as a Transformer-based variant of Masked Autoregressive Flows (MAFs): it consists of a stack of autoregressive Transformer blocks on image patches, alternating the autoregression direction between layers. TarFlow is straightforward to train end-to-end, and capable of directly modeling and generating pixels. We also propose three key techniques to improve sample quality: Gaussian noise augmentation during training, a post training denoising procedure, and an effective guidance method for both class-conditional and unconditional settings. Putting these together, TarFlow sets new state-of-the-art results on likelihood estimation for images, beating the previous best methods by a large margin, and generates samples with quality and diversity comparable to diffusion models, for the first time with a stand-alone NF model. We make our code available at https://github.com/apple/ml-tarflow.
Determining whether two sets of images belong to the same or different distributions or domains is a crucial task in modern medical image analysis and deep learning; for example, to evaluate the output quality of image generative models. Currently, metrics used for this task either rely on the (potentially biased) choice of some downstream task, such as segmentation, or adopt task-independent perceptual metrics (e.g., Fr\'echet Inception Distance/FID) from natural imaging, which we show insufficiently capture anatomical features. To this end, we introduce a new perceptual metric tailored for medical images, FRD (Fr\'echet Radiomic Distance), which utilizes standardized, clinically meaningful, and interpretable image features. We show that FRD is superior to other image distribution metrics for a range of medical imaging applications, including out-of-domain (OOD) detection, the evaluation of image-to-image translation (by correlating more with downstream task performance as well as anatomical consistency and realism), and the evaluation of unconditional image generation. Moreover, FRD offers additional benefits such as stability and computational efficiency at low sample sizes, sensitivity to image corruptions and adversarial attacks, feature interpretability, and correlation with radiologist-perceived image quality. Additionally, we address key gaps in the literature by presenting an extensive framework for the multifaceted evaluation of image similarity metrics in medical imaging -- including the first large-scale comparative study of generative models for medical image translation -- and release an accessible codebase to facilitate future research. Our results are supported by thorough experiments spanning a variety of datasets, modalities, and downstream tasks, highlighting the broad potential of FRD for medical image analysis.
Diffusion models have significant impact on wide range of generative tasks, especially on image inpainting and restoration. Although the improvements on aiming for decreasing number of function evaluations (NFE), the iterative results are still computationally expensive. Consistency models are as a new family of generative models, enable single-step sampling of high quality data without the need for adversarial training. In this paper, we introduce the beta noise distribution, which provides flexibility in adjusting noise levels. This is combined with a sinusoidal curriculum that enhances the learning of the trajectory between the noise distribution and the posterior distribution of interest, allowing High Noise Improved Consistency Training (HN-iCT) to be trained in a supervised fashion. Additionally, High Noise Improved Consistency Training with Image Condition (HN-iCT-CN) architecture is introduced, enables to take Low Dose images as a condition for extracting significant features by Weighted Attention Gates (WAG).Our results indicate that unconditional image generation using HN-iCT significantly outperforms basic CT and iCT training techniques with NFE=1 on the CIFAR10 and CelebA datasets. Moreover, our image-conditioned model demonstrates exceptional performance in enhancing low-dose (LD) CT scans.
Vision Transformers and U-Net architectures have been widely adopted in the implementation of Diffusion Models. However, each architecture presents specific challenges while realizing them on-device. Vision Transformers require positional embedding to maintain correspondence between the tokens processed by the transformer, although they offer the advantage of using fixed-size, reusable repetitive blocks following tokenization. The U-Net architecture lacks these attributes, as it utilizes variable-sized intermediate blocks for down-convolution and up-convolution in the noise estimation backbone for the diffusion process. To address these issues, we propose an architecture that utilizes a fixed-size, reusable transformer block as a core structure, making it more suitable for hardware implementation. Our architecture is characterized by low complexity, token-free design, absence of positional embeddings, uniformity, and scalability, making it highly suitable for deployment on mobile and resource-constrained devices. The proposed model exhibit competitive and consistent performance across both unconditional and conditional image generation tasks. The model achieved a state-of-the-art FID score of 1.6 on unconditional image generation with the CelebA.
Score identity Distillation (SiD) is a data-free method that has achieved SOTA performance in image generation by leveraging only a pretrained diffusion model, without requiring any training data. However, its ultimate performance is constrained by how accurate the pretrained model captures the true data scores at different stages of the diffusion process. In this paper, we introduce SiDA (SiD with Adversarial Loss), which not only enhances generation quality but also improves distillation efficiency by incorporating real images and adversarial loss. SiDA utilizes the encoder from the generator's score network as a discriminator, allowing it to distinguish between real images and those generated by SiD. The adversarial loss is batch-normalized within each GPU and then combined with the original SiD loss. This integration effectively incorporates the average "fakeness" per GPU batch into the pixel-based SiD loss, enabling SiDA to distill a single-step generator. SiDA converges significantly faster than its predecessor when distilled from scratch, and swiftly improves upon the original model's performance during fine-tuning from a pre-distilled SiD generator. This one-step adversarial distillation method establishes new benchmarks in generation performance when distilling EDM diffusion models, achieving FID scores of 1.110 on ImageNet 64x64. When distilling EDM2 models trained on ImageNet 512x512, our SiDA method surpasses even the largest teacher model, EDM2-XXL, which achieved an FID of 1.81 using classifier-free guidance (CFG) and 63 generation steps. In contrast, SiDA achieves FID scores of 2.156 for size XS, 1.669 for S, 1.488 for M, 1.413 for L, 1.379 for XL, and 1.366 for XXL, all without CFG and in a single generation step. These results highlight substantial improvements across all model sizes. Our code is available at https://github.com/mingyuanzhou/SiD/tree/sida.
Latent-based image generative models, such as Latent Diffusion Models (LDMs) and Mask Image Models (MIMs), have achieved notable success in image generation tasks. These models typically leverage reconstructive autoencoders like VQGAN or VAE to encode pixels into a more compact latent space and learn the data distribution in the latent space instead of directly from pixels. However, this practice raises a pertinent question: Is it truly the optimal choice? In response, we begin with an intriguing observation: despite sharing the same latent space, autoregressive models significantly lag behind LDMs and MIMs in image generation. This finding contrasts sharply with the field of NLP, where the autoregressive model GPT has established a commanding presence. To address this discrepancy, we introduce a unified perspective on the relationship between latent space and generative models, emphasizing the stability of latent space in image generative modeling. Furthermore, we propose a simple but effective discrete image tokenizer to stabilize the latent space for image generative modeling by applying K-Means on the latent features of self-supervised learning models. Experimental results show that image autoregressive modeling with our tokenizer (DiGIT) benefits both image understanding and image generation with the next token prediction principle, which is inherently straightforward for GPT models but challenging for other generative models. Remarkably, for the first time, a GPT-style autoregressive model for images outperforms LDMs, which also exhibits substantial improvement akin to GPT when scaling up model size. Our findings underscore the potential of an optimized latent space and the integration of discrete tokenization in advancing the capabilities of image generative models. The code is available at \url{https://github.com/DAMO-NLP-SG/DiGIT}.
Classical generative diffusion models learn an isotropic Gaussian denoising process, treating all spatial regions uniformly, thus neglecting potentially valuable structural information in the data. Inspired by the long-established work on anisotropic diffusion in image processing, we present a novel edge-preserving diffusion model that generalizes over existing isotropic models by considering a hybrid noise scheme. In particular, we introduce an edge-aware noise scheduler that varies between edge-preserving and isotropic Gaussian noise. We show that our model's generative process converges faster to results that more closely match the target distribution. We demonstrate its capability to better learn the low-to-mid frequencies within the dataset, which plays a crucial role in representing shapes and structural information. Our edge-preserving diffusion process consistently outperforms state-of-the-art baselines in unconditional image generation. It is also particularly more robust for generative tasks guided by a shape-based prior, such as stroke-to-image generation. We present qualitative and quantitative results (FID and CLIP score) showing consistent improvements of up to 30% for both tasks.
Energy based models (EBMs) are appealing for their generality and simplicity in data likelihood modeling, but have conventionally been difficult to train due to the unstable and time-consuming implicit MCMC sampling during contrastive divergence training. In this paper, we present a novel energy-based generative framework, Variational Potential Flow (VAPO), that entirely dispenses with implicit MCMC sampling and does not rely on complementary latent models or cooperative training. The VAPO framework aims to learn a potential energy function whose gradient (flow) guides the prior samples, so that their density evolution closely follows an approximate data likelihood homotopy. An energy loss function is then formulated to minimize the Kullback-Leibler divergence between density evolution of the flow-driven prior and the data likelihood homotopy. Images can be generated after training the potential energy, by initializing the samples from Gaussian prior and solving the ODE governing the potential flow on a fixed time interval using generic ODE solvers. Experiment results show that the proposed VAPO framework is capable of generating realistic images on various image datasets. In particular, our proposed framework achieves competitive FID scores for unconditional image generation on the CIFAR-10 and CelebA datasets.
We propose a gradient flow procedure for generative modeling by transporting particles from an initial source distribution to a target distribution, where the gradient field on the particles is given by a noise-adaptive Wasserstein Gradient of the Maximum Mean Discrepancy (MMD). The noise-adaptive MMD is trained on data distributions corrupted by increasing levels of noise, obtained via a forward diffusion process, as commonly used in denoising diffusion probabilistic models. The result is a generalization of MMD Gradient Flow, which we call Diffusion-MMD-Gradient Flow or DMMD. The divergence training procedure is related to discriminator training in Generative Adversarial Networks (GAN), but does not require adversarial training. We obtain competitive empirical performance in unconditional image generation on CIFAR10, MNIST, CELEB-A (64 x64) and LSUN Church (64 x 64). Furthermore, we demonstrate the validity of the approach when MMD is replaced by a lower bound on the KL divergence.
Diffusion generative modelling (DGM) based on stochastic differential equations (SDEs) with score matching has achieved unprecedented results in data generation. In this paper, we propose a novel fast high-quality generative modelling method based on high-order Langevin dynamics (HOLD) with score matching. This motive is proved by third-order Langevin dynamics. By augmenting the previous SDEs, e.g. variance exploding or variance preserving SDEs for single-data variable processes, HOLD can simultaneously model position, velocity, and acceleration, thereby improving the quality and speed of the data generation at the same time. HOLD is composed of one Ornstein-Uhlenbeck process and two Hamiltonians, which reduce the mixing time by two orders of magnitude. Empirical experiments for unconditional image generation on the public data set CIFAR-10 and CelebA-HQ show that the effect is significant in both Frechet inception distance (FID) and negative log-likelihood, and achieves the state-of-the-art FID of 1.85 on CIFAR-10.
Latent Diffusion Models (LDMs) have emerged as powerful generative models, known for delivering remarkable results under constrained computational resources. However, deploying LDMs on resource-limited devices remains a complex issue, presenting challenges such as memory consumption and inference speed. To address this issue, we introduce LD-Pruner, a novel performance-preserving structured pruning method for compressing LDMs. Traditional pruning methods for deep neural networks are not tailored to the unique characteristics of LDMs, such as the high computational cost of training and the absence of a fast, straightforward and task-agnostic method for evaluating model performance. Our method tackles these challenges by leveraging the latent space during the pruning process, enabling us to effectively quantify the impact of pruning on model performance, independently of the task at hand. This targeted pruning of components with minimal impact on the output allows for faster convergence during training, as the model has less information to re-learn, thereby addressing the high computational cost of training. Consequently, our approach achieves a compressed model that offers improved inference speed and reduced parameter count, while maintaining minimal performance degradation. We demonstrate the effectiveness of our approach on three different tasks: text-to-image (T2I) generation, Unconditional Image Generation (UIG) and Unconditional Audio Generation (UAG). Notably, we reduce the inference time of Stable Diffusion (SD) by 34.9% while simultaneously improving its FID by 5.2% on MS-COCO T2I benchmark. This work paves the way for more efficient pruning methods for LDMs, enhancing their applicability.
Recently, diffusion transformers have gained wide attention with its excellent performance in text-to-image and text-to-vidoe models, emphasizing the need for transformers as backbone for diffusion models. Transformer-based models have shown better generalization capability compared to CNN-based models for general vision tasks. However, much less has been explored in the existing literature regarding the capabilities of transformer-based diffusion backbones and expanding their generative prowess to other datasets. This paper focuses on enabling a single pre-trained diffusion transformer model to scale across multiple datasets swiftly, allowing for the completion of diverse generative tasks using just one model. To this end, we propose DiffScaler, an efficient scaling strategy for diffusion models where we train a minimal amount of parameters to adapt to different tasks. In particular, we learn task-specific transformations at each layer by incorporating the ability to utilize the learned subspaces of the pre-trained model, as well as the ability to learn additional task-specific subspaces, which may be absent in the pre-training dataset. As these parameters are independent, a single diffusion model with these task-specific parameters can be used to perform multiple tasks simultaneously. Moreover, we find that transformer-based diffusion models significantly outperform CNN-based diffusion models methods while performing fine-tuning over smaller datasets. We perform experiments on four unconditional image generation datasets. We show that using our proposed method, a single pre-trained model can scale up to perform these conditional and unconditional tasks, respectively, with minimal parameter tuning while performing as close as fine-tuning an entire diffusion model for that particular task.
Transformers have catalyzed advancements in computer vision and natural language processing (NLP) fields. However, substantial computational complexity poses limitations for their application in long-context tasks, such as high-resolution image generation. This paper introduces a series of architectures adapted from the RWKV model used in the NLP, with requisite modifications tailored for diffusion model applied to image generation tasks, referred to as Diffusion-RWKV. Similar to the diffusion with Transformers, our model is designed to efficiently handle patchnified inputs in a sequence with extra conditions, while also scaling up effectively, accommodating both large-scale parameters and extensive datasets. Its distinctive advantage manifests in its reduced spatial aggregation complexity, rendering it exceptionally adept at processing high-resolution images, thereby eliminating the necessity for windowing or group cached operations. Experimental results on both condition and unconditional image generation tasks demonstrate that Diffison-RWKV achieves performance on par with or surpasses existing CNN or Transformer-based diffusion models in FID and IS metrics while significantly reducing total computation FLOP usage.