Recent advancements in multi-modal large language models have propelled the development of joint probabilistic models capable of both image understanding and generation. However, we have identified that recent methods inevitably suffer from loss of image information during understanding task, due to either image discretization or diffusion denoising steps. To address this issue, we propose a novel Multi-Modal Auto-Regressive (MMAR) probabilistic modeling framework. Unlike discretization line of method, MMAR takes in continuous-valued image tokens to avoid information loss. Differing from diffusion-based approaches, we disentangle the diffusion process from auto-regressive backbone model by employing a light-weight diffusion head on top each auto-regressed image patch embedding. In this way, when the model transits from image generation to understanding through text generation, the backbone model's hidden representation of the image is not limited to the last denoising step. To successfully train our method, we also propose a theoretically proven technique that addresses the numerical stability issue and a training strategy that balances the generation and understanding task goals. Through extensive evaluations on 18 image understanding benchmarks, MMAR demonstrates much more superior performance than other joint multi-modal models, matching the method that employs pretrained CLIP vision encoder, meanwhile being able to generate high quality images at the same time. We also showed that our method is scalable with larger data and model size.
Generative models transform random noise into images; their inversion aims to transform images back to structured noise for recovery and editing. This paper addresses two key tasks: (i) inversion and (ii) editing of a real image using stochastic equivalents of rectified flow models (such as Flux). Although Diffusion Models (DMs) have recently dominated the field of generative modeling for images, their inversion presents faithfulness and editability challenges due to nonlinearities in drift and diffusion. Existing state-of-the-art DM inversion approaches rely on training of additional parameters or test-time optimization of latent variables; both are expensive in practice. Rectified Flows (RFs) offer a promising alternative to diffusion models, yet their inversion has been underexplored. We propose RF inversion using dynamic optimal control derived via a linear quadratic regulator. We prove that the resulting vector field is equivalent to a rectified stochastic differential equation. Additionally, we extend our framework to design a stochastic sampler for Flux. Our inversion method allows for state-of-the-art performance in zero-shot inversion and editing, outperforming prior works in stroke-to-image synthesis and semantic image editing, with large-scale human evaluations confirming user preference.
We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096$\times$4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8$\times$, we trained an AE that can compress images 32$\times$, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024$\times$1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.
We introduce a new paradigm for AutoRegressive (AR) image generation, termed Set AutoRegressive Modeling (SAR). SAR generalizes the conventional AR to the next-set setting, i.e., splitting the sequence into arbitrary sets containing multiple tokens, rather than outputting each token in a fixed raster order. To accommodate SAR, we develop a straightforward architecture termed Fully Masked Transformer. We reveal that existing AR variants correspond to specific design choices of sequence order and output intervals within the SAR framework, with AR and Masked AR (MAR) as two extreme instances. Notably, SAR facilitates a seamless transition from AR to MAR, where intermediate states allow for training a causal model that benefits from both few-step inference and KV cache acceleration, thus leveraging the advantages of both AR and MAR. On the ImageNet benchmark, we carefully explore the properties of SAR by analyzing the impact of sequence order and output intervals on performance, as well as the generalization ability regarding inference order and steps. We further validate the potential of SAR by training a 900M text-to-image model capable of synthesizing photo-realistic images with any resolution. We hope our work may inspire more exploration and application of AR-based modeling across diverse modalities.
We introduce Hybrid Autoregressive Transformer (HART), an autoregressive (AR) visual generation model capable of directly generating 1024x1024 images, rivaling diffusion models in image generation quality. Existing AR models face limitations due to the poor image reconstruction quality of their discrete tokenizers and the prohibitive training costs associated with generating 1024px images. To address these challenges, we present the hybrid tokenizer, which decomposes the continuous latents from the autoencoder into two components: discrete tokens representing the big picture and continuous tokens representing the residual components that cannot be represented by the discrete tokens. The discrete component is modeled by a scalable-resolution discrete AR model, while the continuous component is learned with a lightweight residual diffusion module with only 37M parameters. Compared with the discrete-only VAR tokenizer, our hybrid approach improves reconstruction FID from 2.11 to 0.30 on MJHQ-30K, leading to a 31% generation FID improvement from 7.85 to 5.38. HART also outperforms state-of-the-art diffusion models in both FID and CLIP score, with 4.5-7.7x higher throughput and 6.9-13.4x lower MACs. Our code is open sourced at https://github.com/mit-han-lab/hart.
We present Deep Compression Autoencoder (DC-AE), a new family of autoencoder models for accelerating high-resolution diffusion models. Existing autoencoder models have demonstrated impressive results at a moderate spatial compression ratio (e.g., 8x), but fail to maintain satisfactory reconstruction accuracy for high spatial compression ratios (e.g., 64x). We address this challenge by introducing two key techniques: (1) Residual Autoencoding, where we design our models to learn residuals based on the space-to-channel transformed features to alleviate the optimization difficulty of high spatial-compression autoencoders; (2) Decoupled High-Resolution Adaptation, an efficient decoupled three-phases training strategy for mitigating the generalization penalty of high spatial-compression autoencoders. With these designs, we improve the autoencoder's spatial compression ratio up to 128 while maintaining the reconstruction quality. Applying our DC-AE to latent diffusion models, we achieve significant speedup without accuracy drop. For example, on ImageNet 512x512, our DC-AE provides 19.1x inference speedup and 17.9x training speedup on H100 GPU for UViT-H while achieving a better FID, compared with the widely used SD-VAE-f8 autoencoder. Our code is available at https://github.com/mit-han-lab/efficientvit.
Accurate interpretation and visualization of human instructions are crucial for text-to-image (T2I) synthesis. However, current models struggle to capture semantic variations from word order changes, and existing evaluations, relying on indirect metrics like text-image similarity, fail to reliably assess these challenges. This often obscures poor performance on complex or uncommon linguistic patterns by the focus on frequent word combinations. To address these deficiencies, we propose a novel metric called SemVarEffect and a benchmark named SemVarBench, designed to evaluate the causality between semantic variations in inputs and outputs in T2I synthesis. Semantic variations are achieved through two types of linguistic permutations, while avoiding easily predictable literal variations. Experiments reveal that the CogView-3-Plus and Ideogram 2 performed the best, achieving a score of 0.2/1. Semantic variations in object relations are less understood than attributes, scoring 0.07/1 compared to 0.17-0.19/1. We found that cross-modal alignment in UNet or Transformers plays a crucial role in handling semantic variations, a factor previously overlooked by a focus on textual encoders. Our work establishes an effective evaluation framework that advances the T2I synthesis community's exploration of human instruction understanding. Our benchmark and code are available at https://github.com/zhuxiangru/SemVarBench .
Exemplar-guided image translation, synthesizing photo-realistic images that conform to both structural control and style exemplars, is attracting attention due to its ability to enhance user control over style manipulation. Previous methodologies have predominantly depended on establishing dense correspondences across cross-domain inputs. Despite these efforts, they incur quadratic memory and computational costs for establishing dense correspondence, resulting in limited versatility and performance degradation. In this paper, we propose a novel approach termed Exemplar-guided Image Translation with Brownian-Bridge Diffusion Models (EBDM). Our method formulates the task as a stochastic Brownian bridge process, a diffusion process with a fixed initial point as structure control and translates into the corresponding photo-realistic image while being conditioned solely on the given exemplar image. To efficiently guide the diffusion process toward the style of exemplar, we delineate three pivotal components: the Global Encoder, the Exemplar Network, and the Exemplar Attention Module to incorporate global and detailed texture information from exemplar images. Leveraging Bridge diffusion, the network can translate images from structure control while exclusively conditioned on the exemplar style, leading to more robust training and inference processes. We illustrate the superiority of our method over competing approaches through comprehensive benchmark evaluations and visual results.
Recently, large-scale diffusion models have made impressive progress in text-to-image (T2I) generation. To further equip these T2I models with fine-grained spatial control, approaches like ControlNet introduce an extra network that learns to follow a condition image. However, for every single condition type, ControlNet requires independent training on millions of data pairs with hundreds of GPU hours, which is quite expensive and makes it challenging for ordinary users to explore and develop new types of conditions. To address this problem, we propose the CtrLoRA framework, which trains a Base ControlNet to learn the common knowledge of image-to-image generation from multiple base conditions, along with condition-specific LoRAs to capture distinct characteristics of each condition. Utilizing our pretrained Base ControlNet, users can easily adapt it to new conditions, requiring as few as 1,000 data pairs and less than one hour of single-GPU training to obtain satisfactory results in most scenarios. Moreover, our CtrLoRA reduces the learnable parameters by 90% compared to ControlNet, significantly lowering the threshold to distribute and deploy the model weights. Extensive experiments on various types of conditions demonstrate the efficiency and effectiveness of our method. Codes and model weights will be released at https://github.com/xyfJASON/ctrlora.
Diffusion models, known for their impressive image generation abilities, have played a pivotal role in the rise of visual text generation. Nevertheless, existing visual text generation methods often focus on generating entire images with text prompts, leading to imprecise control and limited practicality. A more promising direction is visual text blending, which focuses on seamlessly merging texts onto text-free backgrounds. However, existing visual text blending methods often struggle to generate high-fidelity and diverse images due to a shortage of backgrounds for synthesis and limited generalization capabilities. To overcome these challenges, we propose a new visual text blending paradigm including both creating backgrounds and rendering texts. Specifically, a background generator is developed to produce high-fidelity and text-free natural images. Moreover, a text renderer named GlyphOnly is designed for achieving visually plausible text-background integration. GlyphOnly, built on a Stable Diffusion framework, utilizes glyphs and backgrounds as conditions for accurate rendering and consistency control, as well as equipped with an adaptive text block exploration strategy for small-scale text rendering. We also explore several downstream applications based on our method, including scene text dataset synthesis for boosting scene text detectors, as well as text image customization and editing. Code and model will be available at \url{https://github.com/Zhenhang-Li/GlyphOnly}.
Most existing generative models are limited to learning a single probability distribution from the training data and cannot generalize to novel distributions for unseen data. An architecture that can generate samples from both trained datasets and unseen probability distributions would mark a significant breakthrough. Recently, score-based generative models have gained considerable attention for their comprehensive mode coverage and high-quality image synthesis, as they effectively learn an operator that maps a probability distribution to its corresponding score function. In this work, we introduce the $\emph{Score Neural Operator}$, which learns the mapping from multiple probability distributions to their score functions within a unified framework. We employ latent space techniques to facilitate the training of score matching, which tends to over-fit in the original image pixel space, thereby enhancing sample generation quality. Our trained Score Neural Operator demonstrates the ability to predict score functions of probability measures beyond the training space and exhibits strong generalization performance in both 2-dimensional Gaussian Mixture Models and 1024-dimensional MNIST double-digit datasets. Importantly, our approach offers significant potential for few-shot learning applications, where a single image from a new distribution can be leveraged to generate multiple distinct images from that distribution.
Conventional class-guided diffusion models generally succeed in generating images with correct semantic content, but often struggle with texture details. This limitation stems from the usage of class priors, which only provide coarse and limited conditional information. To address this issue, we propose Diffusion on Diffusion (DoD), an innovative multi-stage generation framework that first extracts visual priors from previously generated samples, then provides rich guidance for the diffusion model leveraging visual priors from the early stages of diffusion sampling. Specifically, we introduce a latent embedding module that employs a compression-reconstruction approach to discard redundant detail information from the conditional samples in each stage, retaining only the semantic information for guidance. We evaluate DoD on the popular ImageNet-$256 \times 256$ dataset, reducing 7$\times$ training cost compared to SiT and DiT with even better performance in terms of the FID-50K score. Our largest model DoD-XL achieves an FID-50K score of 1.83 with only 1 million training steps, which surpasses other state-of-the-art methods without bells and whistles during inference.
We present Meissonic, which elevates non-autoregressive masked image modeling (MIM) text-to-image to a level comparable with state-of-the-art diffusion models like SDXL. By incorporating a comprehensive suite of architectural innovations, advanced positional encoding strategies, and optimized sampling conditions, Meissonic substantially improves MIM's performance and efficiency. Additionally, we leverage high-quality training data, integrate micro-conditions informed by human preference scores, and employ feature compression layers to further enhance image fidelity and resolution. Our model not only matches but often exceeds the performance of existing models like SDXL in generating high-quality, high-resolution images. Extensive experiments validate Meissonic's capabilities, demonstrating its potential as a new standard in text-to-image synthesis. We release a model checkpoint capable of producing $1024 \times 1024$ resolution images.
The creation of complex 3D scenes tailored to user specifications has been a tedious and challenging task with traditional 3D modeling tools. Although some pioneering methods have achieved automatic text-to-3D generation, they are generally limited to small-scale scenes with restricted control over the shape and texture. We introduce SceneCraft, a novel method for generating detailed indoor scenes that adhere to textual descriptions and spatial layout preferences provided by users. Central to our method is a rendering-based technique, which converts 3D semantic layouts into multi-view 2D proxy maps. Furthermore, we design a semantic and depth conditioned diffusion model to generate multi-view images, which are used to learn a neural radiance field (NeRF) as the final scene representation. Without the constraints of panorama image generation, we surpass previous methods in supporting complicated indoor space generation beyond a single room, even as complicated as a whole multi-bedroom apartment with irregular shapes and layouts. Through experimental analysis, we demonstrate that our method significantly outperforms existing approaches in complex indoor scene generation with diverse textures, consistent geometry, and realistic visual quality. Code and more results are available at: https://orangesodahub.github.io/SceneCraft
Sonar image synthesis is crucial for advancing applications in underwater exploration, marine biology, and defence. Traditional methods often rely on extensive and costly data collection using sonar sensors, jeopardizing data quality and diversity. To overcome these limitations, this study proposes a new sonar image synthesis framework, Synth-SONAR leveraging diffusion models and GPT prompting. The key novelties of Synth-SONAR are threefold: First, by integrating Generative AI-based style injection techniques along with publicly available real/simulated data, thereby producing one of the largest sonar data corpus for sonar research. Second, a dual text-conditioning sonar diffusion model hierarchy synthesizes coarse and fine-grained sonar images with enhanced quality and diversity. Third, high-level (coarse) and low-level (detailed) text-based sonar generation methods leverage advanced semantic information available in visual language models (VLMs) and GPT-prompting. During inference, the method generates diverse and realistic sonar images from textual prompts, bridging the gap between textual descriptions and sonar image generation. This marks the application of GPT-prompting in sonar imagery for the first time, to the best of our knowledge. Synth-SONAR achieves state-of-the-art results in producing high-quality synthetic sonar datasets, significantly enhancing their diversity and realism.
3D-aware image generation necessitates extensive training data to ensure stable training and mitigate the risk of overfitting. This paper first considers a novel task known as One-shot 3D Generative Domain Adaptation (GDA), aimed at transferring a pre-trained 3D generator from one domain to a new one, relying solely on a single reference image. One-shot 3D GDA is characterized by the pursuit of specific attributes, namely, high fidelity, large diversity, cross-domain consistency, and multi-view consistency. Within this paper, we introduce 3D-Adapter, the first one-shot 3D GDA method, for diverse and faithful generation. Our approach begins by judiciously selecting a restricted weight set for fine-tuning, and subsequently leverages four advanced loss functions to facilitate adaptation. An efficient progressive fine-tuning strategy is also implemented to enhance the adaptation process. The synergy of these three technological components empowers 3D-Adapter to achieve remarkable performance, substantiated both quantitatively and qualitatively, across all desired properties of 3D GDA. Furthermore, 3D-Adapter seamlessly extends its capabilities to zero-shot scenarios, and preserves the potential for crucial tasks such as interpolation, reconstruction, and editing within the latent space of the pre-trained generator. Code will be available at https://github.com/iceli1007/3D-Adapter.
Although the diffusion model has achieved remarkable performance in the field of image generation, its high inference delay hinders its wide application in edge devices with scarce computing resources. Therefore, many training-free sampling methods have been proposed to reduce the number of sampling steps required for diffusion models. However, they perform poorly under a very small number of sampling steps. Thanks to the emergence of knowledge distillation technology, the existing training scheme methods have achieved excellent results at very low step numbers. However, the current methods mainly focus on designing novel diffusion model sampling methods with knowledge distillation. How to transfer better diffusion knowledge from teacher models is a more valuable problem but rarely studied. Therefore, we propose Relational Diffusion Distillation (RDD), a novel distillation method tailored specifically for distilling diffusion models. Unlike existing methods that simply align teacher and student models at pixel level or feature distributions, our method introduces cross-sample relationship interaction during the distillation process and alleviates the memory constraints induced by multiple sample interactions. Our RDD significantly enhances the effectiveness of the progressive distillation framework within the diffusion model. Extensive experiments on several datasets (e.g., CIFAR-10 and ImageNet) demonstrate that our proposed RDD leads to 1.47 FID decrease under 1 sampling step compared to state-of-the-art diffusion distillation methods and achieving 256x speed-up compared to DDIM strategy. Code is available at https://github.com/cantbebetter2/RDD.
As the boosting development of large vision-language models like Contrastive Language-Image Pre-training (CLIP), many CLIP-like methods have shown impressive abilities on visual recognition, especially in low-data regimes scenes. However, we have noticed that most of these methods are limited to introducing new modifications on text and image encoder. Recently, latent diffusion models (LDMs) have shown good ability on image generation. The potent capabilities of LDMs direct our focus towards the latent representations sampled by UNet. Inspired by the conjecture in CoOp that learned prompts encode meanings beyond the existing vocabulary, we assume that, for deep models, the latent representations are concise and accurate understanding of images, in which high-frequency, imperceptible details are abstracted away. In this paper, we propose a Few-shot Language Image model Embedded with latent Representations (FLIER) for image recognition by introducing a latent encoder jointly trained with CLIP's image encoder, it incorporates pre-trained vision-language knowledge of CLIP and the latent representations from Stable Diffusion. We first generate images and corresponding latent representations via Stable Diffusion with the textual inputs from GPT-3. With latent representations as "models-understandable pixels", we introduce a flexible convolutional neural network with two convolutional layers to be the latent encoder, which is simpler than most encoders in vision-language models. The latent encoder is jointly trained with CLIP's image encoder, transferring pre-trained knowledge to downstream tasks better. Experiments and extensive ablation studies on various visual classification tasks demonstrate that FLIER performs state-of-the-art on 11 datasets for most few-shot classification.
Discrete diffusion models have achieved success in tasks like image generation and masked language modeling but face limitations in controlled content editing. We introduce DICE (Discrete Inversion for Controllable Editing), the first approach to enable precise inversion for discrete diffusion models, including multinomial diffusion and masked generative models. By recording noise sequences and masking patterns during the reverse diffusion process, DICE enables accurate reconstruction and flexible editing of discrete data without the need for predefined masks or attention manipulation. We demonstrate the effectiveness of DICE across both image and text domains, evaluating it on models such as VQ-Diffusion, Paella, and RoBERTa. Our results show that DICE preserves high data fidelity while enhancing editing capabilities, offering new opportunities for fine-grained content manipulation in discrete spaces.
Diffusion transformers (DiT) have already achieved appealing synthesis and scaling properties in content recreation, e.g., image and video generation. However, scaling laws of DiT are less explored, which usually offer precise predictions regarding optimal model size and data requirements given a specific compute budget. Therefore, experiments across a broad range of compute budgets, from 1e17 to 6e18 FLOPs are conducted to confirm the existence of scaling laws in DiT for the first time. Concretely, the loss of pretraining DiT also follows a power-law relationship with the involved compute. Based on the scaling law, we can not only determine the optimal model size and required data but also accurately predict the text-to-image generation loss given a model with 1B parameters and a compute budget of 1e21 FLOPs. Additionally, we also demonstrate that the trend of pre-training loss matches the generation performances (e.g., FID), even across various datasets, which complements the mapping from compute to synthesis quality and thus provides a predictable benchmark that assesses model performance and data quality at a reduced cost.