Yu-Chuan Su

Xuhui Jia, Yang Zhao, Kelvin C. K. Chan et al.

This paper proposes a method for generating images of customized objects specified by users. The method is based on a general framework that bypasses the lengthy optimization required by previous approaches, which often employ a per-object optimization paradigm. Our framework adopts an encoder to capture high-level identifiable semantics of objects, producing an object-specific embedding with only a single feed-forward pass. The acquired object embedding is then passed to a text-to-image synthesis model for subsequent generation. To effectively blend a object-aware embedding space into a well developed text-to-image model under the same generation context, we investigate different network designs and training strategies, and propose a simple yet effective regularized joint training scheme with an object identity preservation loss. Additionally, we propose a caption generation scheme that become a critical piece in fostering object specific embedding faithfully reflected into the generation process, while keeping control and editing abilities. Once trained, the network is able to produce diverse content and styles, conditioned on both texts and objects. We demonstrate through experiments that our proposed method is able to synthesize images with compelling output quality, appearance diversity, and object fidelity, without the need of test-time optimization. Systematic studies are also conducted to analyze our models, providing insights for future work.

Yu-Chuan Su, Kelvin C. K. Chan, Yandong Li et al.

Many applications can benefit from personalized image generation models, including image enhancement, video conferences, just to name a few. Existing works achieved personalization by fine-tuning one model for each person. While being successful, this approach incurs additional computation and storage overhead for each new identity. Furthermore, it usually expects tens or hundreds of examples per identity to achieve the best performance. To overcome these challenges, we propose an encoder-based approach for personalization. We learn an identity encoder which can extract an identity representation from a set of reference images of a subject, together with a diffusion generator that can generate new images of the subject conditioned on the identity representation. Once being trained, the model can be used to generate images of arbitrary identities given a few examples even if the model hasn't been trained on the identity. Our approach greatly reduces the overhead for personalized image generation and is more applicable in many potential applications. Empirical results show that our approach consistently outperforms existing fine-tuning based approach in both image generation and reconstruction, and the outputs is preferred by users more than 95% of the time compared with the best performing baseline.

Yang Zhao, Tingbo Hou, Yu-Chuan Su et al.

An authentic face restoration system is becoming increasingly demanding in many computer vision applications, e.g., image enhancement, video communication, and taking portrait. Most of the advanced face restoration models can recover high-quality faces from low-quality ones but usually fail to faithfully generate realistic and high-frequency details that are favored by users. To achieve authentic restoration, we propose $\textbf{IDM}$, an $\textbf{I}$teratively learned face restoration system based on denoising $\textbf{D}$iffusion $\textbf{M}$odels (DDMs). We define the criterion of an authentic face restoration system, and argue that denoising diffusion models are naturally endowed with this property from two aspects: intrinsic iterative refinement and extrinsic iterative enhancement. Intrinsic learning can preserve the content well and gradually refine the high-quality details, while extrinsic enhancement helps clean the data and improve the restoration task one step further. We demonstrate superior performance on blind face restoration tasks. Beyond restoration, we find the authentically cleaned data by the proposed restoration system is also helpful to image generation tasks in terms of training stabilization and sample quality. Without modifying the models, we achieve better quality than state-of-the-art on FFHQ and ImageNet generation using either GANs or diffusion models.

Hsin-Ping Huang, Yu-Chuan Su, Ming-Hsuan Yang

We tackle the long video generation problem, i.e.~generating videos beyond the output length of video generation models. Due to the computation resource constraints, video generation models can only generate video clips that are relatively short compared with the length of real videos. Existing works apply a sliding window approach to generate long videos at inference time, which is often limited to generating recurrent events or homogeneous content. To generate long videos covering diverse content and multiple events, we propose to use additional guidance to control the video generation process. We further present a two-stage approach to the problem, which allows us to utilize existing video generation models to generate high-quality videos within a small time window while modeling the video holistically based on the input guidance. The proposed approach is complementary to existing efforts on video generation, which focus on generating realistic video within a fixed time window. Extensive experiments on challenging real-world videos validate the benefit of the proposed method, which improves over state-of-the-art by up to 9.5% in objective metrics and is preferred by users more than 80% of time.

Kangning Liu, Yu-Chuan Su, Wei et al.

The one-shot talking-head synthesis task aims to animate a source image to another pose and expression, which is dictated by a driving frame. Recent methods rely on warping the appearance feature extracted from the source, by using motion fields estimated from the sparse keypoints, that are learned in an unsupervised manner. Due to their lightweight formulation, they are suitable for video conferencing with reduced bandwidth. However, based on our study, current methods suffer from two major limitations: 1) unsatisfactory generation quality in the case of large head poses and the existence of observable pose misalignment between the source and the first frame in driving videos. 2) fail to capture fine yet critical face motion details due to the lack of semantic understanding and appropriate face geometry regularization. To address these shortcomings, we propose a novel method that leverages the rich face prior information, the proposed model can generate face videos with improved semantic consistency (improve baseline by $7\%$ in average keypoint distance) and expression-preserving (outperform baseline by $15 \%$ in average emotion embedding distance) under equivalent bandwidth. Additionally, incorporating such prior information provides us with a convenient interface to achieve highly controllable generation in terms of both pose and expression.

Imagen-Team-Google, Jason Baldridge, Jakob Bauer et al.

We introduce Imagen 3, a latent diffusion model that generates high quality images from text prompts. We describe our quality and responsibility evaluations. Imagen 3 is preferred over other state-of-the-art (SOTA) models at the time of evaluation. In addition, we discuss issues around safety and representation, as well as methods we used to minimize the potential harm of our models.

Nanye Ma, Shangyuan Tong, Haolin Jia et al.

Generative models have made significant impacts across various domains, largely due to their ability to scale during training by increasing data, computational resources, and model size, a phenomenon characterized by the scaling laws. Recent research has begun to explore inference-time scaling behavior in Large Language Models (LLMs), revealing how performance can further improve with additional computation during inference. Unlike LLMs, diffusion models inherently possess the flexibility to adjust inference-time computation via the number of denoising steps, although the performance gains typically flatten after a few dozen. In this work, we explore the inference-time scaling behavior of diffusion models beyond increasing denoising steps and investigate how the generation performance can further improve with increased computation. Specifically, we consider a search problem aimed at identifying better noises for the diffusion sampling process. We structure the design space along two axes: the verifiers used to provide feedback, and the algorithms used to find better noise candidates. Through extensive experiments on class-conditioned and text-conditioned image generation benchmarks, our findings reveal that increasing inference-time compute leads to substantial improvements in the quality of samples generated by diffusion models, and with the complicated nature of images, combinations of the components in the framework can be specifically chosen to conform with different application scenario.

Yu-Chuan Su, Soravit Changpinyo, Xiangning Chen et al.

Visual 2.5D perception involves understanding the semantics and geometry of a scene through reasoning about object relationships with respect to the viewer in an environment. However, existing works in visual recognition primarily focus on the semantics. To bridge this gap, we study 2.5D visual relationship detection (2.5VRD), in which the goal is to jointly detect objects and predict their relative depth and occlusion relationships. Unlike general VRD, 2.5VRD is egocentric, using the camera's viewpoint as a common reference for all 2.5D relationships. Unlike depth estimation, 2.5VRD is object-centric and not only focuses on depth. To enable progress on this task, we create a new dataset consisting of 220k human-annotated 2.5D relationships among 512K objects from 11K images. We analyze this dataset and conduct extensive experiments including benchmarking multiple state-of-the-art VRD models on this task. Our results show that existing models largely rely on semantic cues and simple heuristics to solve 2.5VRD, motivating further research on models for 2.5D perception. The new dataset is available at https://github.com/google-research-datasets/2.5vrd.

Hexiang Hu, Kelvin C. K. Chan, Yu-Chuan Su et al. · deepmind

This paper presents instruct-imagen, a model that tackles heterogeneous image generation tasks and generalizes across unseen tasks. We introduce *multi-modal instruction* for image generation, a task representation articulating a range of generation intents with precision. It uses natural language to amalgamate disparate modalities (e.g., text, edge, style, subject, etc.), such that abundant generation intents can be standardized in a uniform format. We then build instruct-imagen by fine-tuning a pre-trained text-to-image diffusion model with a two-stage framework. First, we adapt the model using the retrieval-augmented training, to enhance model's capabilities to ground its generation on external multimodal context. Subsequently, we fine-tune the adapted model on diverse image generation tasks that requires vision-language understanding (e.g., subject-driven generation, etc.), each paired with a multi-modal instruction encapsulating the task's essence. Human evaluation on various image generation datasets reveals that instruct-imagen matches or surpasses prior task-specific models in-domain and demonstrates promising generalization to unseen and more complex tasks.

Hsin-Ping Huang, Yu-Chuan Su, Deqing Sun et al. · deepmind

Text-to-video generation has shown promising results. However, by taking only natural languages as input, users often face difficulties in providing detailed information to precisely control the model's output. In this work, we propose fine-grained controllable video generation (FACTOR) to achieve detailed control. Specifically, FACTOR aims to control objects' appearances and context, including their location and category, in conjunction with the text prompt. To achieve detailed control, we propose a unified framework to jointly inject control signals into the existing text-to-video model. Our model consists of a joint encoder and adaptive cross-attention layers. By optimizing the encoder and the inserted layer, we adapt the model to generate videos that are aligned with both text prompts and fine-grained control. Compared to existing methods relying on dense control signals such as edge maps, we provide a more intuitive and user-friendly interface to allow object-level fine-grained control. Our method achieves controllability of object appearances without finetuning, which reduces the per-subject optimization efforts for the users. Extensive experiments on standard benchmark datasets and user-provided inputs validate that our model obtains a 70% improvement in controllability metrics over competitive baselines.

Gwanghyun Kim, Alonso Martinez, Yu-Chuan Su et al. · cmu

Training diffusion models for audiovisual sequences allows for a range of generation tasks by learning conditional distributions of various input-output combinations of the two modalities. Nevertheless, this strategy often requires training a separate model for each task which is expensive. Here, we propose a novel training approach to effectively learn arbitrary conditional distributions in the audiovisual space.Our key contribution lies in how we parameterize the diffusion timestep in the forward diffusion process. Instead of the standard fixed diffusion timestep, we propose applying variable diffusion timesteps across the temporal dimension and across modalities of the inputs. This formulation offers flexibility to introduce variable noise levels for various portions of the input, hence the term mixture of noise levels. We propose a transformer-based audiovisual latent diffusion model and show that it can be trained in a task-agnostic fashion using our approach to enable a variety of audiovisual generation tasks at inference time. Experiments demonstrate the versatility of our method in tackling cross-modal and multimodal interpolation tasks in the audiovisual space. Notably, our proposed approach surpasses baselines in generating temporally and perceptually consistent samples conditioned on the input. Project page: avdit2024.github.io

Hsin-Ping Huang, Xinyi Wang, Yonatan Bitton et al.

Recent advances in text-to-image generation have improved the quality of synthesized images, but evaluations mainly focus on aesthetics or alignment with text prompts. Thus, it remains unclear whether these models can accurately represent a wide variety of realistic visual entities. To bridge this gap, we propose KITTEN, a benchmark for Knowledge-InTensive image generaTion on real-world ENtities. Using KITTEN, we conduct a systematic study of the latest text-to-image models and retrieval-augmented models, focusing on their ability to generate real-world visual entities, such as landmarks and animals. Analysis using carefully designed human evaluations, automatic metrics, and MLLM evaluations show that even advanced text-to-image models fail to generate accurate visual details of entities. While retrieval-augmented models improve entity fidelity by incorporating reference images, they tend to over-rely on them and struggle to create novel configurations of the entity in creative text prompts.

Yu-Chuan Su, Raviteja Vemulapalli, Ben Weiss et al.

Image composition plays an important role in the quality of a photo. However, not every camera user possesses the knowledge and expertise required for capturing well-composed photos. While post-capture cropping can improve the composition sometimes, it does not work in many common scenarios in which the photographer needs to adjust the camera view to capture the best shot. To address this issue, we propose a deep learning-based approach that provides suggestions to the photographer on how to adjust the camera view before capturing. By optimizing the composition before a photo is captured, our system helps photographers to capture better photos. As there is no publicly-available dataset for this task, we create a view adjustment dataset by repurposing existing image cropping datasets. Furthermore, we propose a two-stage semi-supervised approach that utilizes both labeled and unlabeled images for training a view adjustment model. Experiment results show that the proposed semi-supervised approach outperforms the corresponding supervised alternatives, and our user study results show that the suggested view adjustment improves image composition 79% of the time.

Yu-Chuan Su, Kristen Grauman

Ideally, 360° imagery could inherit the deep convolutional neural networks (CNNs) already trained with great success on perspective projection images. However, existing methods to transfer CNNs from perspective to spherical images introduce significant computational costs and/or degradations in accuracy. In this work, we present the Kernel Transformer Network (KTN). KTNs efficiently transfer convolution kernels from perspective images to the equirectangular projection of 360° images. Given a source CNN for perspective images as input, the KTN produces a function parameterized by a polar angle and kernel as output. Given a novel 360° image, that function in turn can compute convolutions for arbitrary layers and kernels as would the source CNN on the corresponding tangent plane projections. Distinct from all existing methods, KTNs allow model transfer: the same model can be applied to different source CNNs with the same base architecture. This enables application to multiple recognition tasks without re-training the KTN. Validating our approach with multiple source CNNs and datasets, we show that KTNs improve the state of the art for spherical convolution. KTNs successfully preserve the source CNN's accuracy, while offering transferability, scalability to typical image resolutions, and, in many cases, a substantially lower memory footprint.

Yu-Chuan Su, Kristen Grauman

Standard video encoders developed for conventional narrow field-of-view video are widely applied to 360° video as well, with reasonable results. However, while this approach commits arbitrarily to a projection of the spherical frames, we observe that some orientations of a 360° video, once projected, are more compressible than others. We introduce an approach to predict the sphere rotation that will yield the maximal compression rate. Given video clips in their original encoding, a convolutional neural network learns the association between a clip's visual content and its compressibility at different rotations of a cubemap projection. Given a novel video, our learning-based approach efficiently infers the most compressible direction in one shot, without repeated rendering and compression of the source video. We validate our idea on thousands of video clips and multiple popular video codecs. The results show that this untapped dimension of 360° compression has substantial potential--"good" rotations are typically 8-10% more compressible than bad ones, and our learning approach can predict them reliably 82% of the time.

Yu-Chuan Su, Kristen Grauman

While 360° cameras offer tremendous new possibilities in vision, graphics, and augmented reality, the spherical images they produce make core feature extraction non-trivial. Convolutional neural networks (CNNs) trained on images from perspective cameras yield "flat" filters, yet 360° images cannot be projected to a single plane without significant distortion. A naive solution that repeatedly projects the viewing sphere to all tangent planes is accurate, but much too computationally intensive for real problems. We propose to learn a spherical convolutional network that translates a planar CNN to process 360° imagery directly in its equirectangular projection. Our approach learns to reproduce the flat filter outputs on 360° data, sensitive to the varying distortion effects across the viewing sphere. The key benefits are 1) efficient feature extraction for 360° images and video, and 2) the ability to leverage powerful pre-trained networks researchers have carefully honed (together with massive labeled image training sets) for perspective images. We validate our approach compared to several alternative methods in terms of both raw CNN output accuracy as well as applying a state-of-the-art "flat" object detector to 360° data. Our method yields the most accurate results while saving orders of magnitude in computation versus the existing exact reprojection solution.

Yu-Chuan Su, Kristen Grauman

360$^{\circ}$ video requires human viewers to actively control "where" to look while watching the video. Although it provides a more immersive experience of the visual content, it also introduces additional burden for viewers; awkward interfaces to navigate the video lead to suboptimal viewing experiences. Virtual cinematography is an appealing direction to remedy these problems, but conventional methods are limited to virtual environments or rely on hand-crafted heuristics. We propose a new algorithm for virtual cinematography that automatically controls a virtual camera within a 360$^{\circ}$ video. Compared to the state of the art, our algorithm allows more general camera control, avoids redundant outputs, and extracts its output videos substantially more efficiently. Experimental results on over 7 hours of real "in the wild" video show that our generalized camera control is crucial for viewing 360$^{\circ}$ video, while the proposed efficient algorithm is essential for making the generalized control computationally tractable.

Yu-Chuan Su, Dinesh Jayaraman, Kristen Grauman

We introduce the novel task of Pano2Vid $-$ automatic cinematography in panoramic 360$^{\circ}$ videos. Given a 360$^{\circ}$ video, the goal is to direct an imaginary camera to virtually capture natural-looking normal field-of-view (NFOV) video. By selecting "where to look" within the panorama at each time step, Pano2Vid aims to free both the videographer and the end viewer from the task of determining what to watch. Towards this goal, we first compile a dataset of 360$^{\circ}$ videos downloaded from the web, together with human-edited NFOV camera trajectories to facilitate evaluation. Next, we propose AutoCam, a data-driven approach to solve the Pano2Vid task. AutoCam leverages NFOV web video to discriminatively identify space-time "glimpses" of interest at each time instant, and then uses dynamic programming to select optimal human-like camera trajectories. Through experimental evaluation on multiple newly defined Pano2Vid performance measures against several baselines, we show that our method successfully produces informative videos that could conceivably have been captured by human videographers.

Yu-Chuan Su, Kristen Grauman

In a wearable camera video, we see what the camera wearer sees. While this makes it easy to know roughly what he chose to look at, it does not immediately reveal when he was engaged with the environment. Specifically, at what moments did his focus linger, as he paused to gather more information about something he saw? Knowing this answer would benefit various applications in video summarization and augmented reality, yet prior work focuses solely on the "what" question (estimating saliency, gaze) without considering the "when" (engagement). We propose a learning-based approach that uses long-term egomotion cues to detect engagement, specifically in browsing scenarios where one frequently takes in new visual information (e.g., shopping, touring). We introduce a large, richly annotated dataset for ego-engagement that is the first of its kind. Our approach outperforms a wide array of existing methods. We show engagement can be detected well independent of both scene appearance and the camera wearer's identity.

Yu-Chuan Su, Kristen Grauman

Current approaches for activity recognition often ignore constraints on computational resources: 1) they rely on extensive feature computation to obtain rich descriptors on all frames, and 2) they assume batch-mode access to the entire test video at once. We propose a new active approach to activity recognition that prioritizes "what to compute when" in order to make timely predictions. The main idea is to learn a policy that dynamically schedules the sequence of features to compute on selected frames of a given test video. In contrast to traditional static feature selection, our approach continually re-prioritizes computation based on the accumulated history of observations and accounts for the transience of those observations in ongoing video. We develop variants to handle both the batch and streaming settings. On two challenging datasets, our method provides significantly better accuracy than alternative techniques for a wide range of computational budgets.

Yu-Chuan Su, Tzu-Hsuan Chiu, Chun-Yen Yeh et al.

Unconstrained video recognition and Deep Convolution Network (DCN) are two active topics in computer vision recently. In this work, we apply DCNs as frame-based recognizers for video recognition. Our preliminary studies, however, show that video corpora with complete ground truth are usually not large and diverse enough to learn a robust model. The networks trained directly on the video data set suffer from significant overfitting and have poor recognition rate on the test set. The same lack-of-training-sample problem limits the usage of deep models on a wide range of computer vision problems where obtaining training data are difficult. To overcome the problem, we perform transfer learning from images to videos to utilize the knowledge in the weakly labeled image corpus for video recognition. The image corpus help to learn important visual patterns for natural images, while these patterns are ignored by models trained only on the video corpus. Therefore, the resultant networks have better generalizability and better recognition rate. We show that by means of transfer learning from image to video, we can learn a frame-based recognizer with only 4k videos. Because the image corpus is weakly labeled, the entire learning process requires only 4k annotated instances, which is far less than the million scale image data sets required by previous works. The same approach may be applied to other visual recognition tasks where only scarce training data is available, and it improves the applicability of DCNs in various computer vision problems. Our experiments also reveal the correlation between meta-parameters and the performance of DCNs, given the properties of the target problem and data. These results lead to a heuristic for meta-parameter selection for future researches, which does not rely on the time consuming meta-parameter search.