ICLR 2023 Papers

Building on recent advances in image generation, we present a fully data-driven approach to rendering markup into images. The approach is based on diffusion models, which parameterize the distribution of data using a sequence of denoising operations on top of a Gaussian noise distribution. We view the diffusion denoising process a sequential decision making process, and show that it exhibits compounding errors similar to exposure bias issues in imitation learning problems. To mitigate these issues, we adapt the scheduled sampling algorithm to diffusion training. We conduct experiments on four markup datasets: formulas (LaTeX), table layouts (HTML), sheet music (LilyPond), and molecular images (SMILES). These experiments each verify the effectiveness of diffusion and the use of scheduled sampling to fix generation issues. These results also show that the markup-to-image task presents a useful controlled compositional setting for diagnosing and analyzing generative image models.

What can neural networks learn about the visual world when provided with only a single image as input? While any image obviously cannot contain the multitudes of all existing objects, scenes and lighting conditions -- within the space of all $256^{3\cdot224\cdot224}$ possible $224$-sized square images, it might still provide a strong prior for natural images. To analyze this ``augmented image prior'' hypothesis, we develop a simple framework for training neural networks from scratch using a single image and augmentations using knowledge distillation from a supervised pretrained teacher. With this, we find the answer to the above question to be: `surprisingly, a lot'. In quantitative terms, we find accuracies of $94\%$/$74\%$ on CIFAR-10/100, $69$\% on ImageNet, and by extending this method to video and audio, $51\%$ on Kinetics-400 and $84$\% on SpeechCommands. In extensive analyses spanning 13 datasets, we disentangle the effect of augmentations, choice of data and network architectures and also provide qualitative evaluations that include lucid ``panda neurons'' in networks that have never even seen one.

Over-parameterized residual networks (ResNets) are amongst the most successful convolutional neural architectures for image processing. Here we study their properties through their Gaussian Process and Neural Tangent kernels. We derive explicit formulas for these kernels, analyze their spectra, and provide bounds on their implied condition numbers. Our results indicate that (1) with ReLU activation, the eigenvalues of these residual kernels decay polynomially at a similar rate compared to the same kernels when skip connections are not used, thus maintaining a similar frequency bias; (2) however, residual kernels are more locally biased. Our analysis further shows that the matrices obtained by these residual kernels yield favorable condition numbers at finite depths than those obtained without the skip connections, enabling therefore faster convergence of training with gradient descent.

Ultra-High-Definition (UHD) photo has gradually become the standard configuration in advanced imaging devices. The new standard unveils many issues in existing approaches for low-light image enhancement (LLIE), especially in dealing with the intricate issue of joint luminance enhancement and noise removal while remaining efficient. Unlike existing methods that address the problem in the spatial domain, we propose a new solution, UHDFour, that embeds Fourier transform into a cascaded network. Our approach is motivated by a few unique characteristics in the Fourier domain: 1) most luminance information concentrates on amplitudes while noise is closely related to phases, and 2) a high-resolution image and its low-resolution version share similar amplitude patterns. Through embedding Fourier into our network, the amplitude and phase of a low-light image are separately processed to avoid amplifying noise when enhancing luminance. Besides, UHDFour is scalable to UHD images by implementing amplitude and phase enhancement under the low-resolution regime and then adjusting the high-resolution scale with few computations. We also contribute the first real UHD LLIE dataset, UHD-LL, that contains 2,150 low-noise/normal-clear 4K image pairs with diverse darkness and noise levels captured in different scenarios. With this dataset, we systematically analyze the performance of existing LLIE methods for processing UHD images and demonstrate the advantage of our solution. We believe our new framework, coupled with the dataset, would push the frontier of LLIE towards UHD. The code and dataset are available at https://li-chongyi.github.io/UHDFour/.

Common image-to-image translation methods rely on joint training over data from both source and target domains. The training process requires concurrent access to both datasets, which hinders data separation and privacy protection; and existing models cannot be easily adapted for translation of new domain pairs. We present Dual Diffusion Implicit Bridges (DDIBs), an image translation method based on diffusion models, that circumvents training on domain pairs. Image translation with DDIBs relies on two diffusion models trained independently on each domain, and is a two-step process: DDIBs first obtain latent encodings for source images with the source diffusion model, and then decode such encodings using the target model to construct target images. Both steps are defined via ordinary differential equations (ODEs), thus the process is cycle consistent only up to discretization errors of the ODE solvers. Theoretically, we interpret DDIBs as concatenation of source to latent, and latent to target Schrodinger Bridges, a form of entropy-regularized optimal transport, to explain the efficacy of the method. Experimentally, we apply DDIBs on synthetic and high-resolution image datasets, to demonstrate their utility in a wide variety of translation tasks and their inherent optimal transport properties.

We show vectorized sketch generation can be identified as a reversal of the stroke deformation process. This relationship was established by means of a diffusion model that learns data distributions over the stroke-point locations and pen states of real human sketches. Given randomly scattered stroke-points, sketch generation becomes a process of deformation-based denoising, where the generator rectifies positions of stroke points at each timestep to converge at a recognizable sketch. A key innovation was to embed recognizability into the reverse time diffusion process. It was observed that the estimated noise during the reversal process is strongly correlated with sketch classification accuracy. An auxiliary recurrent neural network (RNN) was consequently used to quantify recognizability during data sampling. It follows that, based on the recognizability scores, a sampling shortcut function can also be devised that renders better quality sketches with fewer sampling steps. Finally it is shown that the model can be easily extended to a conditional generation framework, where given incomplete and unfaithful sketches, it yields one that is more visually appealing and with higher recognizability.

Shape-based virtual screening is widely used in ligand-based drug design to search chemical libraries for molecules with similar 3D shapes yet novel 2D graph structures compared to known ligands. 3D deep generative models can potentially automate this exploration of shape-conditioned 3D chemical space; however, no existing models can reliably generate geometrically realistic drug-like molecules in conformations with a specific shape. We introduce a new multimodal 3D generative model that enables shape-conditioned 3D molecular design by equivariantly encoding molecular shape and variationally encoding chemical identity. We ensure local geometric and chemical validity of generated molecules by using autoregressive fragment-based generation with heuristic bonding geometries, allowing the model to prioritize the scoring of rotatable bonds to best align the growing conformation to the target shape. We evaluate our 3D generative model in tasks relevant to drug design including shape-conditioned generation of chemically diverse molecular structures and shape-constrained molecular property optimization, demonstrating its utility over virtual screening of enumerated libraries.

With the advantages of fast inference and human-friendly flexible manipulation, image-agnostic style manipulation via text guidance enables new applications that were not previously available. The state-of-the-art text-guided image-agnostic manipulation method embeds the representation of each channel of StyleGAN independently in the Contrastive Language-Image Pre-training (CLIP) space, and provides it in the form of a Dictionary to quickly find out the channel-wise manipulation direction during inference time. However, in this paper we argue that this dictionary which is constructed by controlling single channel individually is limited to accommodate the versatility of text guidance since the collective and interactive relation among multiple channels are not considered. Indeed, we show that it fails to discover a large portion of manipulation directions that can be found by existing methods, which manually manipulates latent space without texts. To alleviate this issue, we propose a novel method that learns a Dictionary, whose entry corresponds to the representation of a single channel, by taking into account the manipulation effect coming from the interaction with multiple other channels. We demonstrate that our strategy resolves the inability of previous methods in finding diverse known directions from unsupervised methods and unknown directions from random text while maintaining the real-time inference speed and disentanglement ability.

Image generation has recently seen tremendous advances, with diffusion models allowing to synthesize convincing images for a large variety of text prompts. In this article, we propose DiffEdit, a method to take advantage of text-conditioned diffusion models for the task of semantic image editing, where the goal is to edit an image based on a text query. Semantic image editing is an extension of image generation, with the additional constraint that the generated image should be as similar as possible to a given input image. Current editing methods based on diffusion models usually require to provide a mask, making the task much easier by treating it as a conditional inpainting task. In contrast, our main contribution is able to automatically generate a mask highlighting regions of the input image that need to be edited, by contrasting predictions of a diffusion model conditioned on different text prompts. Moreover, we rely on latent inference to preserve content in those regions of interest and show excellent synergies with mask-based diffusion. DiffEdit achieves state-of-the-art editing performance on ImageNet. In addition, we evaluate semantic image editing in more challenging settings, using images from the COCO dataset as well as text-based generated images.

Modern image retrieval methods typically rely on fine-tuning pre-trained encoders to extract image-level descriptors. However, the most widely used models are pre-trained on ImageNet-1K with limited classes. The pre-trained feature representation is therefore not universal enough to generalize well to the diverse open-world classes. In this paper, we first cluster the large-scale \laion{} into one million pseudo classes based on the joint textual and visual features extracted by the CLIP model. Due to the confusion of label granularity, the automatically clustered dataset inevitably contains heavy inter-class conflict. To alleviate such conflict, we randomly select partial inter-class prototypes to construct the margin-based softmax loss. To further enhance the low-dimensional feature representation, we randomly select partial feature dimensions when calculating the similarities between embeddings and class-wise prototypes. The dual random partial selections are with respect to the class dimension and the feature dimension of the prototype matrix, making the classification conflict-robust and the feature embedding compact. Our method significantly outperforms state-of-the-art unsupervised and supervised image retrieval approaches on multiple benchmarks. The code and pre-trained models are released to facilitate future research \url{https://github.com/deepglint/unicom}.

Since the development of self-supervised visual representation learning from contrastive learning to masked image modeling (MIM), there is no significant difference in essence, that is, how to design proper pretext tasks for vision dictionary look-up. MIM recently dominates this line of research with state-of-the-art performance on vision Transformers (ViTs), where the core is to enhance the patch-level visual context capturing of the network via denoising auto-encoding mechanism. Rather than tailoring image tokenizers with extra training stages as in previous works, we unleash the great potential of contrastive learning on de- noising auto-encoding and introduce a pure MIM method, ConMIM, to produce simple intra-image inter-patch contrastive constraints as the sole learning objectives for masked patch prediction. We further strengthen the denoising mechanism with asymmetric designs, including image perturbations and model progress rates, to improve the network pre-training. ConMIM-pretrained models with various scales achieve competitive results on downstream image classification, semantic segmentation, object detection, and instance segmentation tasks, e.g., on ImageNet-1K classification, we achieve 83.9% top-1 accuracy with ViT-Small and 85.3% with ViT-Base without extra data for pre-training. Code will be available at https://github.com/TencentARC/ConMIM.

Diffusion probabilistic models (DPMs) have become a popular approach to conditional generation, due to their promising results and support for cross-modal synthesis. A key desideratum in conditional synthesis is to achieve high correspondence between the conditioning input and generated output. Most existing methods learn such relationships implicitly, by incorporating the prior into the variational lower bound. In this work, we take a different route---we explicitly enhance input-output connections by maximizing their mutual information. To this end, we introduce a Conditional Discrete Contrastive Diffusion (CDCD) loss and design two contrastive diffusion mechanisms to effectively incorporate it into the denoising process, combining the diffusion training and contrastive learning for the first time by connecting it with the conventional variational objectives. We demonstrate the efficacy of our approach in evaluations with diverse multimodal conditional synthesis tasks: dance-to-music generation, text-to-image synthesis, as well as class-conditioned image synthesis. On each, we enhance the input-output correspondence and achieve higher or competitive general synthesis quality. Furthermore, the proposed approach improves the convergence of diffusion models, reducing the number of required diffusion steps by more than 35% on two benchmarks, significantly increasing the inference speed.

Conditional language models are predominantly trained with maximum likelihood estimation (MLE), giving probability mass to sparsely observed target sequences. While MLE trained models assign high probability to plausible sequences given the context, the model probabilities often do not accurately rank-order generated sequences by quality. This has been empirically observed in beam search decoding as output quality degrading with large beam sizes, and decoding strategies benefiting from heuristics such as length normalization and repetition-blocking. In this work, we introduce sequence likelihood calibration (SLiC) where the likelihood of model generated sequences are calibrated to better align with reference sequences in the model’s latent space. With SLiC, decoding heuristics become unnecessary and decoding candidates’ quality significantly improves regardless of the decoding method. Furthermore, SLiC shows no sign of diminishing returns with model scale, and presents alternative ways to improve quality with limited training and inference budgets. With SLiC, we exceed or match SOTA results on a wide range of generation tasks spanning abstractive summarization, question generation, abstractive question answering and data-to-text generation, even with modest-sized models.

The success of large-scale contrastive vision-language pretraining (CLIP) has benefited both visual recognition and multimodal content understanding. The concise design brings CLIP the advantage in inference efficiency against other vision-language models with heavier cross-attention fusion layers, making it a popular choice for a wide spectrum of downstream tasks. However, CLIP does not explicitly capture the hierarchical nature of high-level and fine-grained semantics conveyed in images and texts, which is arguably critical to vision-language understanding and reasoning. To this end, we equip both the visual and language branches in CLIP with hierarchy-aware attentions, namely Hierarchy-aware CLIP (HiCLIP), to progressively discover semantic hierarchies layer-by-layer from both images and texts in an unsupervised manner. As a result, such hierarchical aggregation significantly improves the cross-modal alignment. To demonstrate the advantages of HiCLIP, we conduct qualitative analysis on its unsupervised hierarchy induction during inference, as well as extensive quantitative experiments on both visual recognition and vision-language downstream tasks.

The classification of gigapixel-sized whole slide images (WSIs) with slide-level labels can be formulated as a multiple-instance-learning (MIL) problem. State-of-the-art models often consist of two decoupled parts: local feature embedding with a pre-trained model followed by a global feature aggregation network for classification. We leverage the properties of the apparent similarity in high-resolution WSIs, which essentially exhibit \textit{low-rank} structures in the data manifold, to develop a novel MIL with a boost in both feature embedding and feature aggregation. We extend the contrastive learning with a pathology-specific Low-Rank Constraint (LRC) for feature embedding to pull together samples (i.e., patches) belonging to the same pathological tissue in the low-rank subspace and simultaneously push apart those from different latent subspaces. At the feature aggregation stage, we introduce an iterative low-rank attention MIL (ILRA-MIL) model to aggregate features with low-rank learnable latent vectors to model global interactions among all instances. We highlight the importance of instance correlation modeling but refrain from directly using the transformer encoder considering the $O(n^2)$ complexity. ILRA-MIL with LRC pre-trained features achieves strong empirical results across various benchmarks, including (i) 96.49\% AUC on the CAMELYON16 for binary metastasis classification, (ii) 97.63\% AUC on the TCGA-NSCLC for lung cancer subtyping, and (iii) 0.6562 kappa on the large-scale PANDA dataset for prostate cancer classification. The code is available at https://github.com/jinxixiang/low_rank_wsi.

Data-efficiency has always been an essential issue in pixel-based reinforcement learning (RL). As the agent not only learns decision-making but also meaningful representations from images. The line of reinforcement learning with data augmentation shows significant improvements in sample-efficiency. However, it is challenging to guarantee the optimality invariant transformation, that is, the augmented data are readily recognized as a completely different state by the agent. In the end, we propose a contrastive invariant transformation (CoIT), a simple yet promising learnable data augmentation combined with standard model-free algorithms to improve sample-efficiency. Concretely, the differentiable CoIT leverages original samples with augmented samples and hastens the state encoder for a contrastive invariant embedding. We evaluate our approach on DeepMind Control Suite and Atari100K. Empirical results verify advances using CoIT, enabling it to outperform the new state-of-the-art on various tasks. Source code is available at https://github.com/mooricAnna/CoIT.

We present a method that enables synthesizing novel views and novel poses of arbitrary human performers from sparse multi-view images. A key ingredient of our method is a hybrid appearance blending module that combines the advantages of the implicit body NeRF representation and image-based rendering. Existing generalizable human NeRF methods that are conditioned on the body model have shown robustness against the geometric variation of arbitrary human performers. Yet they often exhibit blurry results when generalized onto unseen identities. Meanwhile, image-based rendering shows high-quality results when sufficient observations are available, whereas it suffers artifacts in sparse-view settings. We propose Neural Image-based Avatars (NIA) that exploits the best of those two methods: to maintain robustness under new articulations and self-occlusions while directly leveraging the available (sparse) source view colors to preserve appearance details of new subject identities. Our hybrid design outperforms recent methods on both in-domain identity generalization as well as challenging cross-dataset generalization settings. Also, in terms of the pose generalization, our method outperforms even the per-subject optimized animatable NeRF methods.