Text-to-image (T2I) systems enable rapid generation of high-fidelity imagery but are misaligned with how visual ideas develop. T2I systems generate outputs that make implicit visual decisions on behalf of the user, often introduce fine-grained details that can anchor users prematurely and limit their ability to keep options open early on, and cause unintended changes during editing that are difficult to correct and reduce users' sense of control. To address these concerns, we present Creo, a multi-stage T2I system that scaffolds image generation by progressing from rough sketches to high-resolution outputs, exposing intermediary abstractions where users can make incremental changes. Sketch-like abstractions invite user editing and allow users to keep design options open when ideas are still forming due to their provisional nature. Each stage in Creo can be modified with manual changes and AI-assisted operations, enabling fine-grained, step-wise control through a locking mechanism that preserves prior decisions so subsequent edits affect only specified regions or attributes. Users remain in the loop, making and verifying decisions across stages, while the system applies diffs instead of regenerating full images, reducing drift as fidelity increases. A comparative study with a one-shot baseline shows that participants felt stronger ownership over Creo outputs, as they were able to trace their decisions in building up the image. Furthermore, embedding-based analysis indicates that Creo outputs are less homogeneous than one-shot results. These findings suggest that multi-stage generation, combined with intermediate control and decision locking, is a key design principle for improving controllability, user agency, creativity, and output diversity in generative systems.

Digital content creation is experiencing a profound change with the advent of deep generative models. For texturing, conditional image generators now allow the synthesis of realistic RGB images of a 3D scene that align with the geometry of that scene. For appearance modeling, SVBRDF prediction networks recover material parameters from RGB images. Combining these technologies allows us to quickly generate SVBRDF maps for multiple views of a 3D scene, which can be merged to form a SVBRDF texture atlas of that scene. In this paper, we analyze the challenges and opportunities for SVBRDF prediction in the context of such a fast appearance modeling pipeline. On the one hand, single-view SVBRDF predictions might suffer from multiview incoherence and yield inconsistent texture atlases. On the other hand, generated RGB images, and the different modalities on which they are conditioned, can provide additional information for SVBRDF estimation compared to photographs. We compare neural architectures and conditions to identify designs that achieve high accuracy and coherence. We find that, surprisingly, a standard UNet is competitive with more complex designs. Project page: http://repo-sam.inria.fr/nerphys/svbrdf-evaluation

Gaussian Splatting has become the method of choice for 3D reconstruction and real-time rendering of captured real scenes. However, fine appearance details need to be represented as a large number of small Gaussian primitives, which can be wasteful when geometry and appearance exhibit different frequency characteristics. Inspired by the long tradition of texture mapping, we propose to use texture to represent detailed appearance where possible. Our main focus is to incorporate per-primitive texture maps that adapt to the scene in a principled manner during Gaussian Splatting optimization. We do this by proposing a new appearance representation for 2D Gaussian primitives with textures where the size of a texel is bounded by the image sampling frequency and adapted to the content of the input images. We achieve this by adaptively upscaling or downscaling the texture resolution during optimization. In addition, our approach enables control of the number of primitives during optimization based on texture resolution. We show that our approach performs favorably in image quality and total number of parameters used compared to alternative solutions for textured Gaussian primitives. Project page: https://repo-sam.inria.fr/nerphys/gs-texturing/

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. We release our code and pretrained models.

Measuring alignment between language and vision is a fundamental challenge, especially as multimodal data becomes increasingly detailed and complex. Existing methods often rely on collecting human or AI preferences, which can be costly and time-intensive. We propose an alternative approach that leverages cycle consistency as a supervisory signal. Given an image and generated text, we map the text back to image space using a text-to-image model and compute the similarity between the original image and its reconstruction. Analogously, for text-to-image generation, we measure the textual similarity between an input caption and its reconstruction through the cycle. We use the cycle consistency score to rank candidates and construct a preference dataset of 866K comparison pairs. The reward model trained on our dataset, CycleReward, outperforms state-of-the-art alignment metrics on detailed captioning, with superior inference-time scalability when used as a verifier for Best-of-N sampling, while maintaining speed and differentiability. Furthermore, performing DPO and Diffusion DPO using our dataset enhances performance across a wide range of vision-language tasks and text-to-image generation. Our dataset, model, and code are publicly released at https://cyclereward.github.io.

We leverage repetitive elements in 3D scenes to improve novel view synthesis. Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have greatly improved novel view synthesis but renderings of unseen and occluded parts remain low-quality if the training views are not exhaustive enough. Our key observation is that our environment is often full of repetitive elements. We propose to leverage those repetitions to improve the reconstruction of low-quality parts of the scene due to poor coverage and occlusions. We propose a method that segments each repeated instance in a 3DGS reconstruction, registers them together, and allows information to be shared among instances. Our method improves the geometry while also accounting for appearance variations across instances. We demonstrate our method on a variety of synthetic and real scenes with typical repetitive elements, leading to a substantial improvement in the quality of novel view synthesis.

Neural scene representations, both continuous and discrete, have recently emerged as a powerful new paradigm for 3D scene understanding. Recent efforts have tackled unsupervised discovery of object-centric neural scene representations. However, the high cost of ray-marching, exacerbated by the fact that each object representation has to be ray-marched separately, leads to insufficiently sampled radiance fields and thus, noisy renderings, poor framerates, and high memory and time complexity during training and rendering. Here, we propose to represent objects in an object-centric, compositional scene representation as light fields. We propose a novel light field compositor module that enables reconstructing the global light field from a set of object-centric light fields. Dubbed Compositional Object Light Fields (COLF), our method enables unsupervised learning of object-centric neural scene representations, state-of-the-art reconstruction and novel view synthesis performance on standard datasets, and rendering and training speeds at orders of magnitude faster than existing 3D approaches.

We propose a method to control material attributes of objects like roughness, metallic, albedo, and transparency in real images. Our method capitalizes on the generative prior of text-to-image models known for photorealism, employing a scalar value and instructions to alter low-level material properties. Addressing the lack of datasets with controlled material attributes, we generated an object-centric synthetic dataset with physically-based materials. Fine-tuning a modified pre-trained text-to-image model on this synthetic dataset enables us to edit material properties in real-world images while preserving all other attributes. We show the potential application of our model to material edited NeRFs.

Recent approaches have shown promises distilling expensive diffusion models into efficient one-step generators. Amongst them, Distribution Matching Distillation (DMD) produces one-step generators that match their teacher in distribution, i.e., the distillation process does not enforce a one-to-one correspondence with the sampling trajectories of their teachers. However, to ensure stable training in practice, 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 not only computationally expensive for large-scale text-to-image synthesis, but it also 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 with sufficient accuracy 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, thus mitigating the imperfect "real" score estimation from the teacher model, and thereby enhancing quality. Third, we introduce a new training procedure that enables multi-step sampling in the student, and addresses the training--inference input mismatch of previous work, by simulating inference-time generator samples during training. Taken together, our improvements set new benchmarks in one-step image generation, with FID scores of 1.28 on ImageNet-64×64 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, and surpassing the teacher. We release our code and pretrained models.

Diffusion models excel at text-to-image generation, especially in subject-driven generation for personalized images. However, existing methods are inefficient due to the subject-specific fine-tuning, which is computationally intensive and hampers efficient deployment. Moreover, existing methods struggle with multi-subject generation as they often blend features among subjects. We present FastComposer which enables efficient, personalized, multi-subject text-to-image generation without fine-tuning. FastComposer uses subject embeddings extracted by an image encoder to augment the generic text conditioning in diffusion models, enabling personalized image generation based on subject images and textual instructions with only forward passes. To address the identity blending problem in the multi-subject generation, FastComposer proposes cross-attention localization supervision during training, enforcing the attention of reference subjects localized to the correct regions in the target images. Naively conditioning on subject embeddings results in subject overfitting. FastComposer proposes delayed subject conditioning in the denoising step to maintain both identity and editability in subject-driven image generation. FastComposer generates images of multiple unseen individuals with different styles, actions, and contexts. It achieves 300$\times$-2500$\times$ speedup compared to fine-tuning-based methods and requires zero extra storage for new subjects. FastComposer paves the way for efficient, personalized, and high-quality multi-subject image creation. Code, model, and dataset are available at https://github.com/mit-han-lab/fastcomposer.

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.

Embedded sensors in smart devices pose privacy risks, often unintentionally leaking user information. We investigate how combining an ambient light sensor with a device display can capture an image of touch interaction without a camera. By displaying a known video sequence, we use the light sensor to capture reflected light intensity variations partially blocked by the touching hand, formulating an inverse problem similar to single-pixel imaging. Because of the sensors’ heavy quantization and low sensitivity, we propose an inversion algorithm involving an ℓ p -norm dequantizer and a deep denoiser as natural image priors, to reconstruct images from the screen’s perspective. We demonstrate touch interactions and eavesdropping hand gestures on an off-the-shelf Android tablet. Despite limitations in resolution and speed, we aim to raise awareness of potential security/privacy threats induced by the combination of passive and active components in smart devices and promote the development of ways to mitigate them.

We present a method to map 2D image observations of a scene to a persistent 3D scene representation, enabling novel view synthesis and disentangled representation of the movable and immovable components of the scene. Motivated by the bird’s-eye-view (BEV) representation commonly used in vision and robotics, we propose conditional neural groundplans, ground-aligned 2D feature grids, as persistent and memory-efficient scene representations. Our method is trained self-supervised from unlabeled multi-view observations using differentiable rendering, and learns to complete geometry and appearance of occluded regions. In addition, we show that we can leverage multi-view videos at training time to learn to separately reconstruct static and movable components of the scene from a single image at test time. The ability to separately reconstruct movable objects enables a variety of downstream tasks using simple heuristics, such as extraction of object-centric 3D representations, novel view synthesis, instance-level segmentation, 3D bounding box prediction, and scene editing. This highlights the value of neural groundplans as a backbone for efficient 3D scene understanding models.

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

We introduce SLANG.D, an extension to the Slang shading language that incorporates first-class automatic differentiation support. The new shading language allows us to transform a Direct3D-based path tracer to be fully differentiable with minor modifications to existing code. SLANG.D enables a shared ecosystem between machine learning frameworks and pre-existing graphics hardware API-based rendering systems, promoting the interchange of components and ideas across these two domains. Our contributions include a differentiable type system designed to ensure type safety and semantic clarity in codebases that blend differentiable and non-differentiable code, language primitives that automatically generate both forward and reverse gradient propagation methods, and a compiler architecture that generates efficient derivative propagation shader code for graphics pipelines. Our compiler supports differentiating code that involves arbitrary control-flow, dynamic dispatch, generics and higher-order differentiation, while providing developers flexible control of checkpointing and gradient aggregation strategies for best performance. Our system allows us to differentiate an existing real-time path tracer, Falcor, with minimal change to its shader code. We show that the compiler-generated derivative kernels perform as efficiently as handwritten ones. In several benchmarks, the SLANG.D code achieves significant speedup when compared to prior automatic differentiation systems.

We study the problem of extracting biometric information of individuals by looking at shadows of objects cast on diffuse surfaces. We show that the biometric information leakage from shadows can be sufficient for reliable identity inference under representative scenarios via a maximum likelihood analysis. We then develop a learning-based method that demonstrates this phenomenon in real settings, exploiting the subtle cues in the shadows that are the source of the leakage without requiring any labeled real data. In particular, our approach relies on building synthetic scenes composed of 3D face models obtained from a single photograph of each identity. We transfer what we learn from the synthetic data to the real data using domain adaptation in a completely unsupervised way. Our model is able to generalize well to the real domain and is robust to several variations in the scenes. We report high classification accuracies in an identity classification task that takes place in a scene with unknown geometry and occluding objects.

Separating an image into meaningful underlying components is a crucial first step for both editing and understanding images. We present a method capable of selecting the regions of a photograph exhibiting the same material as an artist-chosen area. Our proposed approach is robust to shading, specular highlights, and cast shadows, enabling selection in real images. As we do not rely on semantic segmentation (different woods or metal should not be selected together), we formulate the problem as a similarity-based grouping problem based on a user-provided image location. In particular, we propose to leverage the unsupervised DINO [Caron et al. 2021] features coupled with a proposed Cross-Similarity Feature Weighting module and an MLP head to extract material similarities in an image. We train our model on a new synthetic image dataset, that we release. We show that our method generalizes well to real-world images. We carefully analyze our model's behavior on varying material properties and lighting. Additionally, we evaluate it against a hand-annotated benchmark of 50 real photographs. We further demonstrate our model on a set of applications, including material editing, in-video selection, and retrieval of object photographs with similar materials.

Object recognition and viewpoint estimation lie at the heart of visual understanding. Recent works suggest that convolutional neural networks (CNNs) fail to generalize to out-of-distribution (OOD) category-viewpoint combinations, ie. combinations not seen during training. In this paper, we investigate when and how such OOD generalization may be possible by evaluating CNNs trained to classify both object category and 3D viewpoint on OOD combinations, and identifying the neural mechanisms that facilitate such OOD generalization. We show that increasing the number of in-distribution combinations (ie. data diversity) substantially improves generalization to OOD combinations, even with the same amount of training data. We compare learning category and viewpoint in separate and shared network architectures, and observe starkly different trends on in-distribution and OOD combinations, ie. while shared networks are helpful in-distribution, separate networks significantly outperform shared ones at OOD combinations. Finally, we demonstrate that such OOD generalization is facilitated by the neural mechanism of specialization, ie. the emergence of two types of neurons -- neurons selective to category and invariant to viewpoint, and vice versa.

Video conferencing systems suffer from poor user experience when network conditions deteriorate because current video codecs simply cannot operate at extremely low bitrates. Recently, several neural alternatives have been proposed that reconstruct talking head videos at very low bitrates using sparse representations of each frame such as facial landmark information. However, these approaches produce poor reconstructions in scenarios with major movement or occlusions over the course of a call, and do not scale to higher resolutions. We design Gemino, a new neural compression system for video conferencing based on a novel high-frequency-conditional super-resolution pipeline. Gemino upsamples a very low-resolution version of each target frame while enhancing high-frequency details (e.g., skin texture, hair, etc.) based on information extracted from a single high-resolution reference image. We use a multi-scale architecture that runs different components of the model at different resolutions, allowing it to scale to resolutions comparable to 720p, and we personalize the model to learn specific details of each person, achieving much better fidelity at low bitrates. We implement Gemino atop aiortc, an open-source Python implementation of WebRTC, and show that it operates on 1024x1024 videos in real-time on a Titan X GPU, and achieves 2.2-5x lower bitrate than traditional video codecs for the same perceptual quality.