On Inductive Biases That Enable Generalization of Diffusion Transformers

Recent work studying the generalization of diffusion models with UNet-based denoisers reveals inductive biases that can be expressed via geometry-adaptive harmonic bases. However, in practice, more recent denoising networks are often based on transformers, e.g., the diffusion transformer (DiT). This raises the question: do transformer-based denoising networks exhibit inductive biases that can also be expressed via geometry-adaptive harmonic bases? To our surprise, we find that this is not the case. This discrepancy motivates our search for the inductive bias that can lead to good generalization in DiT models. Investigating the pivotal attention modules of a DiT, we find that locality of attention maps are closely associated with generalization. To verify this finding, we modify the generalization of a DiT by restricting its attention windows. We inject local attention windows to a DiT and observe an improvement in generalization. Furthermore, we empirically find that both the placement and the effective attention size of these local attention windows are crucial factors. Experimental results on the CelebA, ImageNet, and LSUN datasets show that strengthening the inductive bias of a DiT can improve both generalization and generation quality when less training data is available. Source code will be released publicly upon paper publication. Project page: dit-generalization.github.io/.

Pixel-Aligned Multi-View Generation with Depth Guided Decoder

The task of image-to-multi-view generation refers to generating novel views of an instance from a single image. Recent methods achieve this by extending text-to-image latent diffusion models to multi-view version, which contains an VAE image encoder and a U-Net diffusion model. Specifically, these generation methods usually fix VAE and finetune the U-Net only. However, the significant downscaling of the latent vectors computed from the input images and independent decoding leads to notable pixel-level misalignment across multiple views. To address this, we propose a novel method for pixel-level image-to-multi-view generation. Unlike prior work, we incorporate attention layers across multi-view images in the VAE decoder of a latent video diffusion model. Specifically, we introduce a depth-truncated epipolar attention, enabling the model to focus on spatially adjacent regions while remaining memory efficient. Applying depth-truncated attn is challenging during inference as the ground-truth depth is usually difficult to obtain and pre-trained depth estimation models is hard to provide accurate depth. Thus, to enhance the generalization to inaccurate depth when ground truth depth is missing, we perturb depth inputs during training. During inference, we employ a rapid multi-view to 3D reconstruction approach, NeuS, to obtain coarse depth for the depth-truncated epipolar attention. Our model enables better pixel alignment across multi-view images. Moreover, we demonstrate the efficacy of our approach in improving downstream multi-view to 3D reconstruction tasks.

NeRFDeformer: NeRF Transformation from a Single View via 3D Scene Flows

We present a method for automatically modifying a NeRF representation based on a single observation of a non-rigid transformed version of the original scene. Our method defines the transformation as a 3D flow, specifically as a weighted linear blending of rigid transformations of 3D anchor points that are defined on the surface of the scene. In order to identify anchor points, we introduce a novel correspondence algorithm that first matches RGB-based pairs, then leverages multi-view information and 3D reprojection to robustly filter false positives in two steps. We also introduce a new dataset for exploring the problem of modifying a NeRF scene through a single observation. Our dataset ( https://github.com/nerfdeformer/nerfdeformer ) contains 113 synthetic scenes leveraging 47 3D assets. We show that our proposed method outperforms NeRF editing methods as well as diffusion-based methods, and we also explore different methods for filtering correspondences.

Virtual Pets: Animatable Animal Generation in 3D Scenes

Toward unlocking the potential of generative models in immersive 4D experiences, we introduce Virtual Pet, a novel pipeline to model realistic and diverse motions for target animal species within a 3D environment. To circumvent the limited availability of 3D motion data aligned with environmental geometry, we leverage monocular internet videos and extract deformable NeRF representations for the foreground and static NeRF representations for the background. For this, we develop a reconstruction strategy, encompassing species-level shared template learning and per-video fine-tuning. Utilizing the reconstructed data, we then train a conditional 3D motion model to learn the trajectory and articulation of foreground animals in the context of 3D backgrounds. We showcase the efficacy of our pipeline with comprehensive qualitative and quantitative evaluations using cat videos. We also demonstrate versatility across unseen cats and indoor environments, producing temporally coherent 4D outputs for enriched virtual experiences.

Putting the Object Back into Video Object Segmentation

We present Cutie, a video object segmentation (VOS) network with object-level memory reading, which puts the object representation from memory back into the video object segmentation result. Recent works on VOS employ bottom-up pixel-level memory reading which struggles due to matching noise, especially in the presence of distractors, resulting in lower performance in more challenging data. In contrast, Cutie performs top-down object-level memory reading by adapting a small set of object queries for restructuring and interacting with the bottom-up pixel features iteratively with a query-based object transformer (qt, hence Cutie). The object queries act as a high-level summary of the target object, while high-resolution feature maps are retained for accurate segmentation. Together with foreground-background masked attention, Cutie cleanly separates the semantics of the foreground object from the background. On the challenging MOSE dataset, Cutie improves by 8.7 J&F over XMem with a similar running time and improves by 4.2 J&F over DeAOT while running three times as fast. Code is available at: https://hkchengrex.github.io/Cutie

Tracking Anything with Decoupled Video Segmentation

Training data for video segmentation are expensive to annotate. This impedes extensions of end-to-end algorithms to new video segmentation tasks, especially in large-vocabulary settings. To 'track anything' without training on video data for every individual task, we develop a decoupled video segmentation approach (DEVA), composed of task-specific image-level segmentation and class/task-agnostic bi-directional temporal propagation. Due to this design, we only need an image-level model for the target task (which is cheaper to train) and a universal temporal propagation model which is trained once and generalizes across tasks. To effectively combine these two modules, we use bi-directional propagation for (semi-)online fusion of segmentation hypotheses from different frames to generate a coherent segmentation. We show that this decoupled formulation compares favorably to end-to-end approaches in several data-scarce tasks including large-vocabulary video panoptic segmentation, open-world video segmentation, referring video segmentation, and unsupervised video object segmentation. Code is available at: https://hkchengrex.github.io/Tracking-Anything-with-DEVA

SDFusion: Multimodal 3D Shape Completion, Reconstruction, and Generation

In this work, we present a novel framework built to simplify 3D asset generation for amateur users. To enable interactive generation, our method supports a variety of input modalities that can be easily provided by a human, including images, text, partially observed shapes and combinations of these, further allowing to adjust the strength of each input. At the core of our approach is an encoder-decoder, compressing 3D shapes into a compact latent representation, upon which a diffusion model is learned. To enable a variety of multi-modal inputs, we employ task-specific encoders with dropout followed by a cross-attention mechanism. Due to its flexibility, our model naturally supports a variety of tasks, outperforming prior works on shape completion, image-based 3D reconstruction, and text-to-3D. Most interestingly, our model can combine all these tasks into one swiss-army-knife tool, enabling the user to perform shape generation using incomplete shapes, images, and textual descriptions at the same time, providing the relative weights for each input and facilitating interactivity. Despite our approach being shape-only, we further show an efficient method to texture the generated shape using large-scale text-to-image models.

On the Importance of Gradient Norm in PAC-Bayesian Bounds

Generalization bounds which assess the difference between the true risk and the empirical risk, have been studied extensively. However, to obtain bounds, current techniques use strict assumptions such as a uniformly bounded or a Lipschitz loss function. To avoid these assumptions, in this paper, we follow an alternative approach: we relax uniform bounds assumptions by using on-average bounded loss and on-average bounded gradient norm assumptions. Following this relaxation, we propose a new generalization bound that exploits the contractivity of the log-Sobolev inequalities. These inequalities add an additional loss-gradient norm term to the generalization bound, which is intuitively a surrogate of the model complexity. We apply the proposed bound on Bayesian deep nets and empirically analyze the effect of this new loss-gradient norm term on different neural architectures.

Joint Forecasting of Panoptic Segmentations with Difference Attention

Forecasting of a representation is important for safe and effective autonomy. For this, panoptic segmentations have been studied as a compelling representation in recent work. However, recent state-of-the-art on panoptic segmentation forecasting suffers from two issues: first, individual object instances are treated independently of each other; second, individual object instance forecasts are merged in a heuristic manner. To address both issues, we study a new panoptic segmentation forecasting model that jointly forecasts all object instances in a scene using a transformer model based on 'difference attention.' It further refines the predictions by taking depth estimates into account. We evaluate the proposed model on the Cityscapes and AIODrive datasets. We find difference attention to be particularly suitable for forecasting because the difference of quantities like locations enables a model to explicitly reason about velocities and acceleration. Because of this, we attain state-of-the-art on panoptic segmentation forecasting metrics.

MuMuQA: Multimedia Multi-Hop News Question Answering via Cross-Media Knowledge Extraction and Grounding

Recently, there has been an increasing interest in building question answering (QA) models that reason across multiple modalities, such as text and images. However, QA using images is often limited to just picking the answer from a pre-defined set of options. In addition, images in the real world, especially in news, have objects that are co-referential to the text, with complementary information from both modalities. In this paper, we present a new QA evaluation benchmark with 1,384 questions over news articles that require cross-media grounding of objects in images onto text. Specifically, the task involves multi-hop questions that require reasoning over image-caption pairs to identify the grounded visual object being referred to and then predicting a span from the news body text to answer the question. In addition, we introduce a novel multimedia data augmentation framework, based on cross-media knowledge extraction and synthetic question-answer generation, to automatically augment data that can provide weak supervision for this task. We evaluate both pipeline-based and end-to-end pretraining-based multimedia QA models on our benchmark, and show that they achieve promising performance, while considerably lagging behind human performance hence leaving large room for future work on this challenging new task.