Richard Zhang

Sheng-Yu Wang, Alexei A. Efros, Jun-Yan Zhu et al. · berkeley

While large text-to-image models are able to synthesize "novel" images, these images are necessarily a reflection of the training data. The problem of data attribution in such models -- which of the images in the training set are most responsible for the appearance of a given generated image -- is a difficult yet important one. As an initial step toward this problem, we evaluate attribution through "customization" methods, which tune an existing large-scale model toward a given exemplar object or style. Our key insight is that this allows us to efficiently create synthetic images that are computationally influenced by the exemplar by construction. With our new dataset of such exemplar-influenced images, we are able to evaluate various data attribution algorithms and different possible feature spaces. Furthermore, by training on our dataset, we can tune standard models, such as DINO, CLIP, and ViT, toward the attribution problem. Even though the procedure is tuned towards small exemplar sets, we show generalization to larger sets. Finally, by taking into account the inherent uncertainty of the problem, we can assign soft attribution scores over a set of training images.

Dave Epstein, Taesung Park, Richard Zhang et al. · berkeley

We propose an unsupervised, mid-level representation for a generative model of scenes. The representation is mid-level in that it is neither per-pixel nor per-image; rather, scenes are modeled as a collection of spatial, depth-ordered "blobs" of features. Blobs are differentiably placed onto a feature grid that is decoded into an image by a generative adversarial network. Due to the spatial uniformity of blobs and the locality inherent to convolution, our network learns to associate different blobs with different entities in a scene and to arrange these blobs to capture scene layout. We demonstrate this emergent behavior by showing that, despite training without any supervision, our method enables applications such as easy manipulation of objects within a scene (e.g., moving, removing, and restyling furniture), creation of feasible scenes given constraints (e.g., plausible rooms with drawers at a particular location), and parsing of real-world images into constituent parts. On a challenging multi-category dataset of indoor scenes, BlobGAN outperforms StyleGAN2 in image quality as measured by FID. See our project page for video results and interactive demo: https://www.dave.ml/blobgan

Nupur Kumari, Bingliang Zhang, Richard Zhang et al.

While generative models produce high-quality images of concepts learned from a large-scale database, a user often wishes to synthesize instantiations of their own concepts (for example, their family, pets, or items). Can we teach a model to quickly acquire a new concept, given a few examples? Furthermore, can we compose multiple new concepts together? We propose Custom Diffusion, an efficient method for augmenting existing text-to-image models. We find that only optimizing a few parameters in the text-to-image conditioning mechanism is sufficiently powerful to represent new concepts while enabling fast tuning (~6 minutes). Additionally, we can jointly train for multiple concepts or combine multiple fine-tuned models into one via closed-form constrained optimization. Our fine-tuned model generates variations of multiple new concepts and seamlessly composes them with existing concepts in novel settings. Our method outperforms or performs on par with several baselines and concurrent works in both qualitative and quantitative evaluations while being memory and computationally efficient.

Minguk Kang, Jun-Yan Zhu, Richard Zhang et al.

The recent success of text-to-image synthesis has taken the world by storm and captured the general public's imagination. From a technical standpoint, it also marked a drastic change in the favored architecture to design generative image models. GANs used to be the de facto choice, with techniques like StyleGAN. With DALL-E 2, auto-regressive and diffusion models became the new standard for large-scale generative models overnight. This rapid shift raises a fundamental question: can we scale up GANs to benefit from large datasets like LAION? We find that naÏvely increasing the capacity of the StyleGAN architecture quickly becomes unstable. We introduce GigaGAN, a new GAN architecture that far exceeds this limit, demonstrating GANs as a viable option for text-to-image synthesis. GigaGAN offers three major advantages. First, it is orders of magnitude faster at inference time, taking only 0.13 seconds to synthesize a 512px image. Second, it can synthesize high-resolution images, for example, 16-megapixel pixels in 3.66 seconds. Finally, GigaGAN supports various latent space editing applications such as latent interpolation, style mixing, and vector arithmetic operations.

Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang et al.

Large-scale text-to-image diffusion models can generate high-fidelity images with powerful compositional ability. However, these models are typically trained on an enormous amount of Internet data, often containing copyrighted material, licensed images, and personal photos. Furthermore, they have been found to replicate the style of various living artists or memorize exact training samples. How can we remove such copyrighted concepts or images without retraining the model from scratch? To achieve this goal, we propose an efficient method of ablating concepts in the pretrained model, i.e., preventing the generation of a target concept. Our algorithm learns to match the image distribution for a target style, instance, or text prompt we wish to ablate to the distribution corresponding to an anchor concept. This prevents the model from generating target concepts given its text condition. Extensive experiments show that our method can successfully prevent the generation of the ablated concept while preserving closely related concepts in the model.

Stephanie Fu, Netanel Tamir, Shobhita Sundaram et al.

Current perceptual similarity metrics operate at the level of pixels and patches. These metrics compare images in terms of their low-level colors and textures, but fail to capture mid-level similarities and differences in image layout, object pose, and semantic content. In this paper, we develop a perceptual metric that assesses images holistically. Our first step is to collect a new dataset of human similarity judgments over image pairs that are alike in diverse ways. Critical to this dataset is that judgments are nearly automatic and shared by all observers. To achieve this we use recent text-to-image models to create synthetic pairs that are perturbed along various dimensions. We observe that popular perceptual metrics fall short of explaining our new data, and we introduce a new metric, DreamSim, tuned to better align with human perception. We analyze how our metric is affected by different visual attributes, and find that it focuses heavily on foreground objects and semantic content while also being sensitive to color and layout. Notably, despite being trained on synthetic data, our metric generalizes to real images, giving strong results on retrieval and reconstruction tasks. Furthermore, our metric outperforms both prior learned metrics and recent large vision models on these tasks.

Lucy Chai, Michael Gharbi, Eli Shechtman et al.

Generative models operate at fixed resolution, even though natural images come in a variety of sizes. As high-resolution details are downsampled away and low-resolution images are discarded altogether, precious supervision is lost. We argue that every pixel matters and create datasets with variable-size images, collected at their native resolutions. To take advantage of varied-size data, we introduce continuous-scale training, a process that samples patches at random scales to train a new generator with variable output resolutions. First, conditioning the generator on a target scale allows us to generate higher resolution images than previously possible, without adding layers to the model. Second, by conditioning on continuous coordinates, we can sample patches that still obey a consistent global layout, which also allows for scalable training at higher resolutions. Controlled FFHQ experiments show that our method can take advantage of multi-resolution training data better than discrete multi-scale approaches, achieving better FID scores and cleaner high-frequency details. We also train on other natural image domains including churches, mountains, and birds, and demonstrate arbitrary scale synthesis with both coherent global layouts and realistic local details, going beyond 2K resolution in our experiments. Our project page is available at: https://chail.github.io/anyres-gan/.

Gaurav Parmar, Krishna Kumar Singh, Richard Zhang et al.

Large-scale text-to-image generative models have shown their remarkable ability to synthesize diverse and high-quality images. However, it is still challenging to directly apply these models for editing real images for two reasons. First, it is hard for users to come up with a perfect text prompt that accurately describes every visual detail in the input image. Second, while existing models can introduce desirable changes in certain regions, they often dramatically alter the input content and introduce unexpected changes in unwanted regions. In this work, we propose pix2pix-zero, an image-to-image translation method that can preserve the content of the original image without manual prompting. We first automatically discover editing directions that reflect desired edits in the text embedding space. To preserve the general content structure after editing, we further propose cross-attention guidance, which aims to retain the cross-attention maps of the input image throughout the diffusion process. In addition, our method does not need additional training for these edits and can directly use the existing pre-trained text-to-image diffusion model. We conduct extensive experiments and show that our method outperforms existing and concurrent works for both real and synthetic image editing.

Yotam Nitzan, Michaël Gharbi, Richard Zhang et al.

Can one inject new concepts into an already trained generative model, while respecting its existing structure and knowledge? We propose a new task - domain expansion - to address this. Given a pretrained generator and novel (but related) domains, we expand the generator to jointly model all domains, old and new, harmoniously. First, we note the generator contains a meaningful, pretrained latent space. Is it possible to minimally perturb this hard-earned representation, while maximally representing the new domains? Interestingly, we find that the latent space offers unused, "dormant" directions, which do not affect the output. This provides an opportunity: By "repurposing" these directions, we can represent new domains without perturbing the original representation. In fact, we find that pretrained generators have the capacity to add several - even hundreds - of new domains! Using our expansion method, one "expanded" model can supersede numerous domain-specific models, without expanding the model size. Additionally, a single expanded generator natively supports smooth transitions between domains, as well as composition of domains. Code and project page available at https://yotamnitzan.github.io/domain-expansion/.

Tianwei Yin, Michaël Gharbi, Richard Zhang et al.

Diffusion models generate high-quality images but require dozens of forward passes. We introduce Distribution Matching Distillation (DMD), a procedure to transform a diffusion model into a one-step image generator with minimal impact on image quality. We enforce the one-step image generator match the diffusion model at distribution level, by minimizing an approximate KL divergence whose gradient can be expressed as the difference between 2 score functions, one of the target distribution and the other of the synthetic distribution being produced by our one-step generator. The score functions are parameterized as two diffusion models trained separately on each distribution. Combined with a simple regression loss matching the large-scale structure of the multi-step diffusion outputs, our method outperforms all published few-step diffusion approaches, reaching 2.62 FID on ImageNet 64x64 and 11.49 FID on zero-shot COCO-30k, comparable to Stable Diffusion but orders of magnitude faster. Utilizing FP16 inference, our model generates images at 20 FPS on modern hardware.

Difan Liu, Sandesh Shetty, Tobias Hinz et al.

We present ASSET, a neural architecture for automatically modifying an input high-resolution image according to a user's edits on its semantic segmentation map. Our architecture is based on a transformer with a novel attention mechanism. Our key idea is to sparsify the transformer's attention matrix at high resolutions, guided by dense attention extracted at lower image resolutions. While previous attention mechanisms are computationally too expensive for handling high-resolution images or are overly constrained within specific image regions hampering long-range interactions, our novel attention mechanism is both computationally efficient and effective. Our sparsified attention mechanism is able to capture long-range interactions and context, leading to synthesizing interesting phenomena in scenes, such as reflections of landscapes onto water or flora consistent with the rest of the landscape, that were not possible to generate reliably with previous convnets and transformer approaches. We present qualitative and quantitative results, along with user studies, demonstrating the effectiveness of our method.

Gaurav Parmar, Yijun Li, Jingwan Lu et al.

Existing GAN inversion and editing methods work well for aligned objects with a clean background, such as portraits and animal faces, but often struggle for more difficult categories with complex scene layouts and object occlusions, such as cars, animals, and outdoor images. We propose a new method to invert and edit such complex images in the latent space of GANs, such as StyleGAN2. Our key idea is to explore inversion with a collection of layers, spatially adapting the inversion process to the difficulty of the image. We learn to predict the "invertibility" of different image segments and project each segment into a latent layer. Easier regions can be inverted into an earlier layer in the generator's latent space, while more challenging regions can be inverted into a later feature space. Experiments show that our method obtains better inversion results compared to the recent approaches on complex categories, while maintaining downstream editability. Please refer to our project page at https://www.cs.cmu.edu/~SAMInversion.

Sheng-Yu Wang, Aaron Hertzmann, Alexei A Efros et al.

Data attribution for text-to-image models aims to identify the training images that most significantly influenced a generated output. Existing attribution methods involve considerable computational resources for each query, making them impractical for real-world applications. We propose a novel approach for scalable and efficient data attribution. Our key idea is to distill a slow, unlearning-based attribution method to a feature embedding space for efficient retrieval of highly influential training images. During deployment, combined with efficient indexing and search methods, our method successfully finds highly influential images without running expensive attribution algorithms. We show extensive results on both medium-scale models trained on MSCOCO and large-scale Stable Diffusion models trained on LAION, demonstrating that our method can achieve better or competitive performance in a few seconds, faster than existing methods by 2,500x - 400,000x. Our work represents a meaningful step towards the large-scale application of data attribution methods on real-world models such as Stable Diffusion.

Jaskirat Singh, Boyang Zheng, Zongze Wu et al.

Representation Autoencoders (RAE) replace traditional VAE with pretrained vision encoders. In this paper, we systematically investigate several design choices and find three insights which simplify and improve RAE. First, we study a generalized formulation where the representation is defined as sum of the last k encoder layers rather than solely the final layer. This simple change greatly improves reconstruction without encoder finetuning or specialized data (e.g., text, faces). Second, we study the prevalent assumption that RAE (using pretrained representation as encoder) replaces representation alignment (REPA), which distills the same representation to intermediate layers instead. Through large-scale empirical analysis, we uncover a surprising finding: RAE and REPA exhibit complementary working mechanisms, allowing the same representation to be used as both encoder and target for intermediate diffusion layers. Finally, the original RAE struggles with classifier-free guidance (CFG) and requires training a second, weaker diffusion model for AutoGuidance (AG). We show that REPA itself can be viewed as x-prediction in RAE latent space. By simply re-parameterizing the output of the DiT model, it can provide guidance for "free". Overall, RAEv2 leads to more than 10x faster convergence over the original RAE, achieving a state-of-the-art gFID of 1.06 in just 80 epochs on ImageNet-256. On FDr^k, RAEv2 achieves a state-of-the-art 2.17 at just 80 epochs compared to the previous best 3.26 (800 epochs) without any post-training. This motivates EP_FID@k (epochs to reach unguided gFID <= k) as a measure of training efficiency. RAEv2 attains an EP_FID@2 of 35 epochs, versus 177 for the original RAE. We also validate our approach across diverse settings for text-to-image generation and navigation world models, showing consistent improvements. Code is available at https://raev2.github.io.

Zongze Wu, Nicholas Kolkin, Jonathan Brandt et al.

We address the challenges of precise image inversion and disentangled image editing in the context of few-step diffusion models. We introduce an encoder based iterative inversion technique. The inversion network is conditioned on the input image and the reconstructed image from the previous step, allowing for correction of the next reconstruction towards the input image. We demonstrate that disentangled controls can be easily achieved in the few-step diffusion model by conditioning on an (automatically generated) detailed text prompt. To manipulate the inverted image, we freeze the noise maps and modify one attribute in the text prompt (either manually or via instruction based editing driven by an LLM), resulting in the generation of a new image similar to the input image with only one attribute changed. It can further control the editing strength and accept instructive text prompt. Our approach facilitates realistic text-guided image edits in real-time, requiring only 8 number of functional evaluations (NFEs) in inversion (one-time cost) and 4 NFEs per edit. Our method is not only fast, but also significantly outperforms state-of-the-art multi-step diffusion editing techniques.

Yuchen Liu, Zhixin Shu, Yijun Li et al.

3D-controllable portrait synthesis has significantly advanced, thanks to breakthroughs in generative adversarial networks (GANs). However, it is still challenging to manipulate existing face images with precise 3D control. While concatenating GAN inversion and a 3D-aware, noise-to-image GAN is a straight-forward solution, it is inefficient and may lead to noticeable drop in editing quality. To fill this gap, we propose 3D-FM GAN, a novel conditional GAN framework designed specifically for 3D-controllable face manipulation, and does not require any tuning after the end-to-end learning phase. By carefully encoding both the input face image and a physically-based rendering of 3D edits into a StyleGAN's latent spaces, our image generator provides high-quality, identity-preserved, 3D-controllable face manipulation. To effectively learn such novel framework, we develop two essential training strategies and a novel multiplicative co-modulation architecture that improves significantly upon naive schemes. With extensive evaluations, we show that our method outperforms the prior arts on various tasks, with better editability, stronger identity preservation, and higher photo-realism. In addition, we demonstrate a better generalizability of our design on large pose editing and out-of-domain images.

David C. Epstein, Ishan Jain, Oliver Wang et al.

With advancements in AI-generated images coming on a continuous basis, it is increasingly difficult to distinguish traditionally-sourced images (e.g., photos, artwork) from AI-generated ones. Previous detection methods study the generalization from a single generator to another in isolation. However, in reality, new generators are released on a streaming basis. We study generalization in this setting, training on N models and testing on the next (N+k), following the historical release dates of well-known generation methods. Furthermore, images increasingly consist of both real and generated components, for example through image inpainting. Thus, we extend this approach to pixel prediction, demonstrating strong performance using automatically-generated inpainted data. In addition, for settings where commercial models are not publicly available for automatic data generation, we evaluate if pixel detectors can be trained solely on whole synthetic images.

Alireza Ganjdanesh, Yan Kang, Yuchen Liu et al.

Diffusion probabilistic models can generate high-quality samples. Yet, their sampling process requires numerous denoising steps, making it slow and computationally intensive. We propose to reduce the sampling cost by pruning a pretrained diffusion model into a mixture of efficient experts. First, we study the similarities between pairs of denoising timesteps, observing a natural clustering, even across different datasets. This suggests that rather than having a single model for all time steps, separate models can serve as ``experts'' for their respective time intervals. As such, we separately fine-tune the pretrained model on each interval, with elastic dimensions in depth and width, to obtain experts specialized in their corresponding denoising interval. To optimize the resource usage between experts, we introduce our Expert Routing Agent, which learns to select a set of proper network configurations. By doing so, our method can allocate the computing budget between the experts in an end-to-end manner without requiring manual heuristics. Finally, with a selected configuration, we fine-tune our pruned experts to obtain our mixture of efficient experts. We demonstrate the effectiveness of our method, DiffPruning, across several datasets, LSUN-Church, LSUN-Beds, FFHQ, and ImageNet, on the Latent Diffusion Model architecture.

Jaskirat Singh, Xingjian Leng, Zongze Wu et al.

Representation alignment (REPA) guides generative training by distilling representations from a strong, pretrained vision encoder to intermediate diffusion features. We investigate a fundamental question: what aspect of the target representation matters for generation, its \textit{global} \revision{semantic} information (e.g., measured by ImageNet-1K accuracy) or its spatial structure (i.e. pairwise cosine similarity between patch tokens)? Prevalent wisdom holds that stronger global semantic performance leads to better generation as a target representation. To study this, we first perform a large-scale empirical analysis across 27 different vision encoders and different model scales. The results are surprising; spatial structure, rather than global performance, drives the generation performance of a target representation. To further study this, we introduce two straightforward modifications, which specifically accentuate the transfer of \emph{spatial} information. We replace the standard MLP projection layer in REPA with a simple convolution layer and introduce a spatial normalization layer for the external representation. Surprisingly, our simple method (implemented in $<$4 lines of code), termed iREPA, consistently improves convergence speed of REPA, across a diverse set of vision encoders, model sizes, and training variants (such as REPA, REPA-E, Meanflow, JiT etc). %, etc. Our work motivates revisiting the fundamental working mechanism of representational alignment and how it can be leveraged for improved training of generative models. The code and project page are available at https://end2end-diffusion.github.io/irepa

Xin Yu, Xiaojuan Qi, Zhengqi Li et al.

We introduce the Self-Evaluating Model (Self-E), a novel, from-scratch training approach for text-to-image generation that supports any-step inference. Self-E learns from data similarly to a Flow Matching model, while simultaneously employing a novel self-evaluation mechanism: it evaluates its own generated samples using its current score estimates, effectively serving as a dynamic self-teacher. Unlike traditional diffusion or flow models, it does not rely solely on local supervision, which typically necessitates many inference steps. Unlike distillation-based approaches, it does not require a pretrained teacher. This combination of instantaneous local learning and self-driven global matching bridges the gap between the two paradigms, enabling the training of a high-quality text-to-image model from scratch that excels even at very low step counts. Extensive experiments on large-scale text-to-image benchmarks show that Self-E not only excels in few-step generation, but is also competitive with state-of-the-art Flow Matching models at 50 steps. We further find that its performance improves monotonically as inference steps increase, enabling both ultra-fast few-step generation and high-quality long-trajectory sampling within a single unified model. To our knowledge, Self-E is the first from-scratch, any-step text-to-image model, offering a unified framework for efficient and scalable generation.

Sicheng Mo, Thao Nguyen, Richard Zhang et al.

In this work, we explore an untapped signal in diffusion model inference. While all previous methods generate images independently at inference, we instead ask if samples can be generated collaboratively. We propose Group Diffusion, unlocking the attention mechanism to be shared across images, rather than limited to just the patches within an image. This enables images to be jointly denoised at inference time, learning both intra and inter-image correspondence. We observe a clear scaling effect - larger group sizes yield stronger cross-sample attention and better generation quality. Furthermore, we introduce a qualitative measure to capture this behavior and show that its strength closely correlates with FID. Built on standard diffusion transformers, our GroupDiff achieves up to 32.2% FID improvement on ImageNet-256x256. Our work reveals cross-sample inference as an effective, previously unexplored mechanism for generative modeling.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Julia Chae, Nicholas Kolkin, Jui-Hsien Wang et al.

Humans have remarkable selective sensitivity to identities -- easily distinguishing between highly similar identities, even across significantly different contexts such as diverse viewpoints or lighting. Vision models have struggled to match this capability, and progress toward identity-focused tasks such as personalized image generation is slowed by a lack of identity-focused evaluation metrics. To help facilitate progress, we propose ID-Sim, a feed-forward metric designed to faithfully reflect human selective sensitivity. To build ID-Sim, we curate a high-quality training set of images spanning diverse real-world domains, augmented with generative synthetic data that provides controlled, fine-grained identity and contextual variations. We evaluate our metric on a new unified evaluation benchmark for assessing consistency with human annotations across identity-focused recognition, retrieval, and generative tasks.

Rohit Gandikota, Zongze Wu, Richard Zhang et al.

We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at https://sliderspace.baulab.info

Tianwei Yin, Michaël Gharbi, Taesung Park et al.

Recent approaches have shown promises distilling diffusion models into efficient one-step generators. Among them, Distribution Matching Distillation (DMD) produces one-step generators that match their teacher in distribution, without enforcing a one-to-one correspondence with the sampling trajectories of their teachers. However, to ensure stable training, DMD requires an additional regression loss computed using a large set of noise-image pairs generated by the teacher with many steps of a deterministic sampler. This is costly for large-scale text-to-image synthesis and limits the student's quality, tying it too closely to the teacher's original sampling paths. We introduce DMD2, a set of techniques that lift this limitation and improve DMD training. First, we eliminate the regression loss and the need for expensive dataset construction. We show that the resulting instability is due to the fake critic not estimating the distribution of generated samples accurately and propose a two time-scale update rule as a remedy. Second, we integrate a GAN loss into the distillation procedure, discriminating between generated samples and real images. This lets us train the student model on real data, mitigating the imperfect real score estimation from the teacher model, and enhancing quality. Lastly, we modify the training procedure to enable multi-step sampling. We identify and address the training-inference input mismatch problem in this setting, by simulating inference-time generator samples during training time. Taken together, our improvements set new benchmarks in one-step image generation, with FID scores of 1.28 on ImageNet-64x64 and 8.35 on zero-shot COCO 2014, surpassing the original teacher despite a 500X reduction in inference cost. Further, we show our approach can generate megapixel images by distilling SDXL, demonstrating exceptional visual quality among few-step methods.

Joonghyuk Shin, Zhengqi Li, Richard Zhang et al.

Current motion-conditioned video generation methods suffer from prohibitive latency (minutes per video) and non-causal processing that prevents real-time interaction. We present MotionStream, enabling sub-second latency with up to 29 FPS streaming generation on a single GPU. Our approach begins by augmenting a text-to-video model with motion control, which generates high-quality videos that adhere to the global text prompt and local motion guidance, but does not perform inference on the fly. As such, we distill this bidirectional teacher into a causal student through Self Forcing with Distribution Matching Distillation, enabling real-time streaming inference. Several key challenges arise when generating videos of long, potentially infinite time-horizons: (1) bridging the domain gap from training on finite length and extrapolating to infinite horizons, (2) sustaining high quality by preventing error accumulation, and (3) maintaining fast inference, without incurring growth in computational cost due to increasing context windows. A key to our approach is introducing carefully designed sliding-window causal attention, combined with attention sinks. By incorporating self-rollout with attention sinks and KV cache rolling during training, we properly simulate inference-time extrapolations with a fixed context window, enabling constant-speed generation of arbitrarily long videos. Our models achieve state-of-the-art results in motion following and video quality while being two orders of magnitude faster, uniquely enabling infinite-length streaming. With MotionStream, users can paint trajectories, control cameras, or transfer motion, and see results unfold in real-time, delivering a truly interactive experience.

Shivam Duggal, Xingjian Bai, Zongze Wu et al.

Latent diffusion models (LDMs) enable high-fidelity synthesis by operating in learned latent spaces. However, training state-of-the-art LDMs requires complex staging: a tokenizer must be trained first, before the diffusion model can be trained in the frozen latent space. We propose UNITE - an autoencoder architecture for unified tokenization and latent diffusion. UNITE consists of a Generative Encoder that serves as both image tokenizer and latent generator via weight sharing. Our key insight is that tokenization and generation can be viewed as the same latent inference problem under different conditioning regimes: tokenization infers latents from fully observed images, whereas generation infers them from noise together with text or class conditioning. Motivated by this, we introduce a single-stage training procedure that jointly optimizes both tasks via two forward passes through the same Generative Encoder. The shared parameters enable gradients to jointly shape the latent space, encouraging a "common latent language". Across image and molecule modalities, UNITE achieves near state of the art performance without adversarial losses or pretrained encoders (e.g., DINO), reaching FID 2.12 and 1.73 for Base and Large models on ImageNet 256 x 256. We further analyze the Generative Encoder through the lenses of representation alignment and compression. These results show that single stage joint training of tokenization & generation from scratch is feasible.

Ji Lin, Richard Zhang, Frieder Ganz et al.

Generative adversarial networks (GANs) have enabled photorealistic image synthesis and editing. However, due to the high computational cost of large-scale generators (e.g., StyleGAN2), it usually takes seconds to see the results of a single edit on edge devices, prohibiting interactive user experience. In this paper, we take inspirations from modern rendering software and propose Anycost GAN for interactive natural image editing. We train the Anycost GAN to support elastic resolutions and channels for faster image generation at versatile speeds. Running subsets of the full generator produce outputs that are perceptually similar to the full generator, making them a good proxy for preview. By using sampling-based multi-resolution training, adaptive-channel training, and a generator-conditioned discriminator, the anycost generator can be evaluated at various configurations while achieving better image quality compared to separately trained models. Furthermore, we develop new encoder training and latent code optimization techniques to encourage consistency between the different sub-generators during image projection. Anycost GAN can be executed at various cost budgets (up to 10x computation reduction) and adapt to a wide range of hardware and latency requirements. When deployed on desktop CPUs and edge devices, our model can provide perceptually similar previews at 6-12x speedup, enabling interactive image editing. The code and demo are publicly available: https://github.com/mit-han-lab/anycost-gan.

Tianwei Yin, Qiang Zhang, Richard Zhang et al.

Current video diffusion models achieve impressive generation quality but struggle in interactive applications due to bidirectional attention dependencies. The generation of a single frame requires the model to process the entire sequence, including the future. We address this limitation by adapting a pretrained bidirectional diffusion transformer to an autoregressive transformer that generates frames on-the-fly. To further reduce latency, we extend distribution matching distillation (DMD) to videos, distilling 50-step diffusion model into a 4-step generator. To enable stable and high-quality distillation, we introduce a student initialization scheme based on teacher's ODE trajectories, as well as an asymmetric distillation strategy that supervises a causal student model with a bidirectional teacher. This approach effectively mitigates error accumulation in autoregressive generation, allowing long-duration video synthesis despite training on short clips. Our model achieves a total score of 84.27 on the VBench-Long benchmark, surpassing all previous video generation models. It enables fast streaming generation of high-quality videos at 9.4 FPS on a single GPU thanks to KV caching. Our approach also enables streaming video-to-video translation, image-to-video, and dynamic prompting in a zero-shot manner.

Yiran Xu, Taesung Park, Richard Zhang et al.

Video super-resolution (VSR) approaches have shown impressive temporal consistency in upsampled videos. However, these approaches tend to generate blurrier results than their image counterparts as they are limited in their generative capability. This raises a fundamental question: can we extend the success of a generative image upsampler to the VSR task while preserving the temporal consistency? We introduce VideoGigaGAN, a new generative VSR model that can produce videos with high-frequency details and temporal consistency. VideoGigaGAN builds upon a large-scale image upsampler -- GigaGAN. Simply inflating GigaGAN to a video model by adding temporal modules produces severe temporal flickering. We identify several key issues and propose techniques that significantly improve the temporal consistency of upsampled videos. Our experiments show that, unlike previous VSR methods, VideoGigaGAN generates temporally consistent videos with more fine-grained appearance details. We validate the effectiveness of VideoGigaGAN by comparing it with state-of-the-art VSR models on public datasets and showcasing video results with $8\times$ super-resolution.

Yotam Nitzan, Zongze Wu, Richard Zhang et al.

We introduce a novel diffusion transformer, LazyDiffusion, that generates partial image updates efficiently. Our approach targets interactive image editing applications in which, starting from a blank canvas or an image, a user specifies a sequence of localized image modifications using binary masks and text prompts. Our generator operates in two phases. First, a context encoder processes the current canvas and user mask to produce a compact global context tailored to the region to generate. Second, conditioned on this context, a diffusion-based transformer decoder synthesizes the masked pixels in a "lazy" fashion, i.e., it only generates the masked region. This contrasts with previous works that either regenerate the full canvas, wasting time and computation, or confine processing to a tight rectangular crop around the mask, ignoring the global image context altogether. Our decoder's runtime scales with the mask size, which is typically small, while our encoder introduces negligible overhead. We demonstrate that our approach is competitive with state-of-the-art inpainting methods in terms of quality and fidelity while providing a 10x speedup for typical user interactions, where the editing mask represents 10% of the image.

Joanna Materzynska, Josef Sivic, Eli Shechtman et al.

We introduce an approach for augmenting text-to-video generation models with customized motions, extending their capabilities beyond the motions depicted in the original training data. By leveraging a few video samples demonstrating specific movements as input, our method learns and generalizes the input motion patterns for diverse, text-specified scenarios. Our contributions are threefold. First, to achieve our results, we finetune an existing text-to-video model to learn a novel mapping between the depicted motion in the input examples to a new unique token. To avoid overfitting to the new custom motion, we introduce an approach for regularization over videos. Second, by leveraging the motion priors in a pretrained model, our method can produce novel videos featuring multiple people doing the custom motion, and can invoke the motion in combination with other motions. Furthermore, our approach extends to the multimodal customization of motion and appearance of individualized subjects, enabling the generation of videos featuring unique characters and distinct motions. Third, to validate our method, we introduce an approach for quantitatively evaluating the learned custom motion and perform a systematic ablation study. We show that our method significantly outperforms prior appearance-based customization approaches when extended to the motion customization task.

Jiteng Mu, Michaël Gharbi, Richard Zhang et al.

Diffusion models have made significant advances in text-guided synthesis tasks. However, editing user-provided images remains challenging, as the high dimensional noise input space of diffusion models is not naturally suited for image inversion or spatial editing. In this work, we propose an image representation that promotes spatial editing of input images using a diffusion model. Concretely, we learn to encode an input into "image elements" that can faithfully reconstruct an input image. These elements can be intuitively edited by a user, and are decoded by a diffusion model into realistic images. We show the effectiveness of our representation on various image editing tasks, such as object resizing, rearrangement, dragging, de-occlusion, removal, variation, and image composition. Project page: https://jitengmu.github.io/Editable_Image_Elements/

Cusuh Ham, Matthew Fisher, James Hays et al.

We present personalized residuals and localized attention-guided sampling for efficient concept-driven generation using text-to-image diffusion models. Our method first represents concepts by freezing the weights of a pretrained text-conditioned diffusion model and learning low-rank residuals for a small subset of the model's layers. The residual-based approach then directly enables application of our proposed sampling technique, which applies the learned residuals only in areas where the concept is localized via cross-attention and applies the original diffusion weights in all other regions. Localized sampling therefore combines the learned identity of the concept with the existing generative prior of the underlying diffusion model. We show that personalized residuals effectively capture the identity of a concept in ~3 minutes on a single GPU without the use of regularization images and with fewer parameters than previous models, and localized sampling allows using the original model as strong prior for large parts of the image.

Nupur Kumari, Grace Su, Richard Zhang et al.

Model customization introduces new concepts to existing text-to-image models, enabling the generation of these new concepts/objects in novel contexts. However, such methods lack accurate camera view control with respect to the new object, and users must resort to prompt engineering (e.g., adding ``top-view'') to achieve coarse view control. In this work, we introduce a new task -- enabling explicit control of the object viewpoint in the customization of text-to-image diffusion models. This allows us to modify the custom object's properties and generate it in various background scenes via text prompts, all while incorporating the object viewpoint as an additional control. This new task presents significant challenges, as one must harmoniously merge a 3D representation from the multi-view images with the 2D pre-trained model. To bridge this gap, we propose to condition the diffusion process on the 3D object features rendered from the target viewpoint. During training, we fine-tune the 3D feature prediction modules to reconstruct the object's appearance and geometry, while reducing overfitting to the input multi-view images. Our method outperforms existing image editing and model customization baselines in preserving the custom object's identity while following the target object viewpoint and the text prompt.

Xiaojuan Wang, Taesung Park, Yang Zhou et al.

A jump cut offers an abrupt, sometimes unwanted change in the viewing experience. We present a novel framework for smoothing these jump cuts, in the context of talking head videos. We leverage the appearance of the subject from the other source frames in the video, fusing it with a mid-level representation driven by DensePose keypoints and face landmarks. To achieve motion, we interpolate the keypoints and landmarks between the end frames around the cut. We then use an image translation network from the keypoints and source frames, to synthesize pixels. Because keypoints can contain errors, we propose a cross-modal attention scheme to select and pick the most appropriate source amongst multiple options for each key point. By leveraging this mid-level representation, our method can achieve stronger results than a strong video interpolation baseline. We demonstrate our method on various jump cuts in the talking head videos, such as cutting filler words, pauses, and even random cuts. Our experiments show that we can achieve seamless transitions, even in the challenging cases where the talking head rotates or moves drastically in the jump cut.

Netanel Y. Tamir, Shir Amir, Ranel Itzhaky et al.

With rapid advancements in virtual reality (VR) headsets, effectively measuring stereoscopic quality of experience (SQoE) has become essential for delivering immersive and comfortable 3D experiences. However, most existing stereo metrics focus on isolated aspects of the viewing experience such as visual discomfort or image quality, and have traditionally faced data limitations. To address these gaps, we present SCOPE (Stereoscopic COntent Preference Evaluation), a new dataset comprised of real and synthetic stereoscopic images featuring a wide range of common perceptual distortions and artifacts. The dataset is labeled with preference annotations collected on a VR headset, with our findings indicating a notable degree of consistency in user preferences across different headsets. Additionally, we present iSQoE, a new model for stereo quality of experience assessment trained on our dataset. We show that iSQoE aligns better with human preferences than existing methods when comparing mono-to-stereo conversion methods.

Nupur Kumari, Sheng-Yu Wang, Nanxuan Zhao et al.

Recent image editing models have achieved impressive results while following natural language editing instructions, but they rely on supervised fine-tuning with large datasets of input-target pairs. This is a critical bottleneck, as such naturally occurring pairs are hard to curate at scale. Current workarounds use synthetic training pairs that leverage the zero-shot capabilities of existing models. However, this can propagate and magnify the artifacts of the pretrained model into the final trained model. In this work, we present a new training paradigm that eliminates the need for paired data entirely. Our approach directly optimizes a few-step diffusion model by unrolling it during training and leveraging feedback from vision-language models (VLMs). For each input and editing instruction, the VLM evaluates if an edit follows the instruction and preserves unchanged content, providing direct gradients for end-to-end optimization. To ensure visual fidelity, we incorporate distribution matching loss (DMD), which constrains generated images to remain within the image manifold learned by pretrained models. We evaluate our method on standard benchmarks and include an extensive ablation study. Without any paired data, our method performs on par with various image editing diffusion models trained on extensive supervised paired data, under the few-step setting. Given the same VLM as the reward model, we also outperform RL-based techniques like Flow-GRPO.

Grace Su, Sheng-Yu Wang, Aaron Hertzmann et al.

A common and controversial use of text-to-image models is to generate pictures by explicitly naming artists, such as "in the style of Greg Rutkowski". We introduce a benchmark for prompted-artist recognition: predicting which artist names were invoked in the prompt from the image alone. The dataset contains 1.95M images covering 110 artists and spans four generalization settings: held-out artists, increasing prompt complexity, multiple-artist prompts, and different text-to-image models. We evaluate feature similarity baselines, contrastive style descriptors, data attribution methods, supervised classifiers, and few-shot prototypical networks. Generalization patterns vary: supervised and few-shot models excel on seen artists and complex prompts, whereas style descriptors transfer better when the artist's style is pronounced; multi-artist prompts remain the most challenging. Our benchmark reveals substantial headroom and provides a public testbed to advance the responsible moderation of text-to-image models. We release the dataset and benchmark to foster further research: https://graceduansu.github.io/IdentifyingPromptedArtists/

Chun-Hsiao Yeh, Yilin Wang, Nanxuan Zhao et al.

Recent diffusion-based image editing methods have significantly advanced text-guided tasks but often struggle to interpret complex, indirect instructions. Moreover, current models frequently suffer from poor identity preservation, unintended edits, or rely heavily on manual masks. To address these challenges, we introduce X-Planner, a Multimodal Large Language Model (MLLM)-based planning system that effectively bridges user intent with editing model capabilities. X-Planner employs chain-of-thought reasoning to systematically decompose complex instructions into simpler, clear sub-instructions. For each sub-instruction, X-Planner automatically generates precise edit types and segmentation masks, eliminating manual intervention and ensuring localized, identity-preserving edits. Additionally, we propose a novel automated pipeline for generating large-scale data to train X-Planner which achieves state-of-the-art results on both existing benchmarks and our newly introduced complex editing benchmark.

Sheng-Yu Wang, Aaron Hertzmann, Alexei A. Efros et al.

The goal of data attribution for text-to-image models is to identify the training images that most influence the generation of a new image. Influence is defined such that, for a given output, if a model is retrained from scratch without the most influential images, the model would fail to reproduce the same output. Unfortunately, directly searching for these influential images is computationally infeasible, since it would require repeatedly retraining models from scratch. In our work, we propose an efficient data attribution method by simulating unlearning the synthesized image. We achieve this by increasing the training loss on the output image, without catastrophic forgetting of other, unrelated concepts. We then identify training images with significant loss deviations after the unlearning process and label these as influential. We evaluate our method with a computationally intensive but "gold-standard" retraining from scratch and demonstrate our method's advantages over previous methods.

Yinbo Chen, Oliver Wang, Richard Zhang et al.

Diffusion models have shown an impressive ability to model complex data distributions, with several key advantages over GANs, such as stable training, better coverage of the training distribution's modes, and the ability to solve inverse problems without extra training. However, most diffusion models learn the distribution of fixed-resolution images. We propose to learn the distribution of continuous images by training diffusion models on image neural fields, which can be rendered at any resolution, and show its advantages over fixed-resolution models. To achieve this, a key challenge is to obtain a latent space that represents photorealistic image neural fields. We propose a simple and effective method, inspired by several recent techniques but with key changes to make the image neural fields photorealistic. Our method can be used to convert existing latent diffusion autoencoders into image neural field autoencoders. We show that image neural field diffusion models can be trained using mixed-resolution image datasets, outperform fixed-resolution diffusion models followed by super-resolution models, and can solve inverse problems with conditions applied at different scales efficiently.

Minguk Kang, Richard Zhang, Connelly Barnes et al.

We propose a method to distill a complex multistep diffusion model into a single-step conditional GAN student model, dramatically accelerating inference, while preserving image quality. Our approach interprets diffusion distillation as a paired image-to-image translation task, using noise-to-image pairs of the diffusion model's ODE trajectory. For efficient regression loss computation, we propose E-LatentLPIPS, a perceptual loss operating directly in diffusion model's latent space, utilizing an ensemble of augmentations. Furthermore, we adapt a diffusion model to construct a multi-scale discriminator with a text alignment loss to build an effective conditional GAN-based formulation. E-LatentLPIPS converges more efficiently than many existing distillation methods, even accounting for dataset construction costs. We demonstrate that our one-step generator outperforms cutting-edge one-step diffusion distillation models -- DMD, SDXL-Turbo, and SDXL-Lightning -- on the zero-shot COCO benchmark.

Nupur Kumari, Richard Zhang, Eli Shechtman et al.

The advent of large-scale training has produced a cornucopia of powerful visual recognition models. However, generative models, such as GANs, have traditionally been trained from scratch in an unsupervised manner. Can the collective "knowledge" from a large bank of pretrained vision models be leveraged to improve GAN training? If so, with so many models to choose from, which one(s) should be selected, and in what manner are they most effective? We find that pretrained computer vision models can significantly improve performance when used in an ensemble of discriminators. Notably, the particular subset of selected models greatly affects performance. We propose an effective selection mechanism, by probing the linear separability between real and fake samples in pretrained model embeddings, choosing the most accurate model, and progressively adding it to the discriminator ensemble. Interestingly, our method can improve GAN training in both limited data and large-scale settings. Given only 10k training samples, our FID on LSUN Cat matches the StyleGAN2 trained on 1.6M images. On the full dataset, our method improves FID by 1.5x to 2x on cat, church, and horse categories of LSUN.

William Peebles, Jun-Yan Zhu, Richard Zhang et al.

We propose GAN-Supervised Learning, a framework for learning discriminative models and their GAN-generated training data jointly end-to-end. We apply our framework to the dense visual alignment problem. Inspired by the classic Congealing method, our GANgealing algorithm trains a Spatial Transformer to map random samples from a GAN trained on unaligned data to a common, jointly-learned target mode. We show results on eight datasets, all of which demonstrate our method successfully aligns complex data and discovers dense correspondences. GANgealing significantly outperforms past self-supervised correspondence algorithms and performs on-par with (and sometimes exceeds) state-of-the-art supervised correspondence algorithms on several datasets -- without making use of any correspondence supervision or data augmentation and despite being trained exclusively on GAN-generated data. For precise correspondence, we improve upon state-of-the-art supervised methods by as much as $3\times$. We show applications of our method for augmented reality, image editing and automated pre-processing of image datasets for downstream GAN training.

Alex Andonian, Taesung Park, Bryan Russell et al.

Training supervised image synthesis models requires a critic to compare two images: the ground truth to the result. Yet, this basic functionality remains an open problem. A popular line of approaches uses the L1 (mean absolute error) loss, either in the pixel or the feature space of pretrained deep networks. However, we observe that these losses tend to produce overly blurry and grey images, and other techniques such as GANs need to be employed to fight these artifacts. In this work, we introduce an information theory based approach to measuring similarity between two images. We argue that a good reconstruction should have high mutual information with the ground truth. This view enables learning a lightweight critic to "calibrate" a feature space in a contrastive manner, such that reconstructions of corresponding spatial patches are brought together, while other patches are repulsed. We show that our formulation immediately boosts the perceptual realism of output images when used as a drop-in replacement for the L1 loss, with or without an additional GAN loss.

Steven Liu, Xiuming Zhang, Zhoutong Zhang et al.

A neural radiance field (NeRF) is a scene model supporting high-quality view synthesis, optimized per scene. In this paper, we explore enabling user editing of a category-level NeRF - also known as a conditional radiance field - trained on a shape category. Specifically, we introduce a method for propagating coarse 2D user scribbles to the 3D space, to modify the color or shape of a local region. First, we propose a conditional radiance field that incorporates new modular network components, including a shape branch that is shared across object instances. Observing multiple instances of the same category, our model learns underlying part semantics without any supervision, thereby allowing the propagation of coarse 2D user scribbles to the entire 3D region (e.g., chair seat). Next, we propose a hybrid network update strategy that targets specific network components, which balances efficiency and accuracy. During user interaction, we formulate an optimization problem that both satisfies the user's constraints and preserves the original object structure. We demonstrate our approach on various editing tasks over three shape datasets and show that it outperforms prior neural editing approaches. Finally, we edit the appearance and shape of a real photograph and show that the edit propagates to extrapolated novel views.

Lucy Chai, Jun-Yan Zhu, Eli Shechtman et al.

Recent generative models can synthesize "views" of artificial images that mimic real-world variations, such as changes in color or pose, simply by learning from unlabeled image collections. Here, we investigate whether such views can be applied to real images to benefit downstream analysis tasks such as image classification. Using a pretrained generator, we first find the latent code corresponding to a given real input image. Applying perturbations to the code creates natural variations of the image, which can then be ensembled together at test-time. We use StyleGAN2 as the source of generative augmentations and investigate this setup on classification tasks involving facial attributes, cat faces, and cars. Critically, we find that several design decisions are required towards making this process work; the perturbation procedure, weighting between the augmentations and original image, and training the classifier on synthesized images can all impact the result. Currently, we find that while test-time ensembling with GAN-based augmentations can offer some small improvements, the remaining bottlenecks are the efficiency and accuracy of the GAN reconstructions, coupled with classifier sensitivities to artifacts in GAN-generated images.

Gaurav Parmar, Richard Zhang, Jun-Yan Zhu

Metrics for evaluating generative models aim to measure the discrepancy between real and generated images. The often-used Frechet Inception Distance (FID) metric, for example, extracts "high-level" features using a deep network from the two sets. However, we find that the differences in "low-level" preprocessing, specifically image resizing and compression, can induce large variations and have unforeseen consequences. For instance, when resizing an image, e.g., with a bilinear or bicubic kernel, signal processing principles mandate adjusting prefilter width depending on the downsampling factor, to antialias to the appropriate bandwidth. However, commonly-used implementations use a fixed-width prefilter, resulting in aliasing artifacts. Such aliasing leads to corruptions in the feature extraction downstream. Next, lossy compression, such as JPEG, is commonly used to reduce the file size of an image. Although designed to minimally degrade the perceptual quality of an image, the operation also produces variations downstream. Furthermore, we show that if compression is used on real training images, FID can actually improve if the generated images are also subsequently compressed. This paper shows that choices in low-level image processing have been an underappreciated aspect of generative modeling. We identify and characterize variations in generative modeling development pipelines, provide recommendations based on signal processing principles, and release a reference implementation to facilitate future comparisons.

Utkarsh Ojha, Yijun Li, Jingwan Lu et al.

Training generative models, such as GANs, on a target domain containing limited examples (e.g., 10) can easily result in overfitting. In this work, we seek to utilize a large source domain for pretraining and transfer the diversity information from source to target. We propose to preserve the relative similarities and differences between instances in the source via a novel cross-domain distance consistency loss. To further reduce overfitting, we present an anchor-based strategy to encourage different levels of realism over different regions in the latent space. With extensive results in both photorealistic and non-photorealistic domains, we demonstrate qualitatively and quantitatively that our few-shot model automatically discovers correspondences between source and target domains and generates more diverse and realistic images than previous methods.