Jakob Verbeek

Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich et al. · meta-ai

Lossy image compression aims to represent images in as few bits as possible while maintaining fidelity to the original. Theoretical results indicate that optimizing distortion metrics such as PSNR or MS-SSIM necessarily leads to a discrepancy in the statistics of original images from those of reconstructions, in particular at low bitrates, often manifested by the blurring of the compressed images. Previous work has leveraged adversarial discriminators to improve statistical fidelity. Yet these binary discriminators adopted from generative modeling tasks may not be ideal for image compression. In this paper, we introduce a non-binary discriminator that is conditioned on quantized local image representations obtained via VQ-VAE autoencoders. Our evaluations on the CLIC2020, DIV2K and Kodak datasets show that our discriminator is more effective for jointly optimizing distortion (e.g., PSNR) and statistical fidelity (e.g., FID) than the PatchGAN of the state-of-the-art HiFiC model. On CLIC2020, we obtain the same FID as HiFiC with 30-40\% fewer bits.

Guillaume Couairon, Jakob Verbeek, Holger Schwenk et al.

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.

Marlène Careil, Matthew J. Muckley, Jakob Verbeek et al. · meta-ai

Image codecs are typically optimized to trade-off bitrate \vs distortion metrics. At low bitrates, this leads to compression artefacts which are easily perceptible, even when training with perceptual or adversarial losses. To improve image quality and remove dependency on the bitrate, we propose to decode with iterative diffusion models. We condition the decoding process on a vector-quantized image representation, as well as a global image description to provide additional context. We dub our model PerCo for 'perceptual compression', and compare it to state-of-the-art codecs at rates from 0.1 down to 0.003 bits per pixel. The latter rate is more than an order of magnitude smaller than those considered in most prior work, compressing a 512x768 Kodak image with less than 153 bytes. Despite this ultra-low bitrate, our approach maintains the ability to reconstruct realistic images. We find that our model leads to reconstructions with state-of-the-art visual quality as measured by FID and KID. As predicted by rate-distortion-perception theory, visual quality is less dependent on the bitrate than previous methods.

Alaaeldin El-Nouby, Matthew J. Muckley, Karen Ullrich et al. · meta-ai

Recent neural compression methods have been based on the popular hyperprior framework. It relies on Scalar Quantization and offers a very strong compression performance. This contrasts from recent advances in image generation and representation learning, where Vector Quantization is more commonly employed. In this work, we attempt to bring these lines of research closer by revisiting vector quantization for image compression. We build upon the VQ-VAE framework and introduce several modifications. First, we replace the vanilla vector quantizer by a product quantizer. This intermediate solution between vector and scalar quantization allows for a much wider set of rate-distortion points: It implicitly defines high-quality quantizers that would otherwise require intractably large codebooks. Second, inspired by the success of Masked Image Modeling (MIM) in the context of self-supervised learning and generative image models, we propose a novel conditional entropy model which improves entropy coding by modelling the co-dependencies of the quantized latent codes. The resulting PQ-MIM model is surprisingly effective: its compression performance on par with recent hyperprior methods. It also outperforms HiFiC in terms of FID and KID metrics when optimized with perceptual losses (e.g. adversarial). Finally, since PQ-MIM is compatible with image generation frameworks, we show qualitatively that it can operate under a hybrid mode between compression and generation, with no further training or finetuning. As a result, we explore the extreme compression regime where an image is compressed into 200 bytes, i.e., less than a tweet.

Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby et al.

After their initial success in natural language processing, transformer architectures have rapidly gained traction in computer vision, providing state-of-the-art results for tasks such as image classification, detection, segmentation, and video analysis. We offer three insights based on simple and easy to implement variants of vision transformers. (1) The residual layers of vision transformers, which are usually processed sequentially, can to some extent be processed efficiently in parallel without noticeably affecting the accuracy. (2) Fine-tuning the weights of the attention layers is sufficient to adapt vision transformers to a higher resolution and to other classification tasks. This saves compute, reduces the peak memory consumption at fine-tuning time, and allows sharing the majority of weights across tasks. (3) Adding MLP-based patch pre-processing layers improves Bert-like self-supervised training based on patch masking. We evaluate the impact of these design choices using the ImageNet-1k dataset, and confirm our findings on the ImageNet-v2 test set. Transfer performance is measured across six smaller datasets.

Guillaume Couairon, Marlène Careil, Matthieu Cord et al.

Large-scale text-to-image diffusion models have significantly improved the state of the art in generative image modelling and allow for an intuitive and powerful user interface to drive the image generation process. Expressing spatial constraints, e.g. to position specific objects in particular locations, is cumbersome using text; and current text-based image generation models are not able to accurately follow such instructions. In this paper we consider image generation from text associated with segments on the image canvas, which combines an intuitive natural language interface with precise spatial control over the generated content. We propose ZestGuide, a zero-shot segmentation guidance approach that can be plugged into pre-trained text-to-image diffusion models, and does not require any additional training. It leverages implicit segmentation maps that can be extracted from cross-attention layers, and uses them to align the generation with input masks. Our experimental results combine high image quality with accurate alignment of generated content with input segmentations, and improve over prior work both quantitatively and qualitatively, including methods that require training on images with corresponding segmentations. Compared to Paint with Words, the previous state-of-the art in image generation with zero-shot segmentation conditioning, we improve by 5 to 10 mIoU points on the COCO dataset with similar FID scores.

Guillaume Couairon, Asya Grechka, Jakob Verbeek et al.

Deep generative models, like GANs, have considerably improved the state of the art in image synthesis, and are able to generate near photo-realistic images in structured domains such as human faces. Based on this success, recent work on image editing proceeds by projecting images to the GAN latent space and manipulating the latent vector. However, these approaches are limited in that only images from a narrow domain can be transformed, and with only a limited number of editing operations. We propose FlexIT, a novel method which can take any input image and a user-defined text instruction for editing. Our method achieves flexible and natural editing, pushing the limits of semantic image translation. First, FlexIT combines the input image and text into a single target point in the CLIP multimodal embedding space. Via the latent space of an auto-encoder, we iteratively transform the input image toward the target point, ensuring coherence and quality with a variety of novel regularization terms. We propose an evaluation protocol for semantic image translation, and thoroughly evaluate our method on ImageNet. Code will be made publicly available.

Xiaochuang Han, Youssef Emad, Melissa Hall et al. · meta-ai

Video generation models are rapidly advancing, but can still struggle with complex video outputs that require significant semantic branching or repeated high-level reasoning about what should happen next. In this paper, we introduce a new class of omni video-text models that integrate ideas from recent LM reasoning advances to address this challenge. More specifically, we present TV2TV, a unified generative modeling framework which decomposes video generation into an interleaved text and video generation process. TV2TV jointly learns language modeling (next-token prediction) and video flow matching (next-frame prediction) using a Mixture-of-Transformers (MoT) architecture. At inference time, TV2TV decides when to alternate between generating text and video frames, allowing the model to "think in words" about subsequent content before ``acting in pixels'' to produce frames. This design offloads much of the responsibility for deciding what should happen next to the language modeling tower, enabling improved visual quality and prompt alignment of generated videos. It also enables fine-grained controllability, allowing users to modify the video generation trajectory through text interventions at any point in the process. In controlled experiments on video game data, TV2TV demonstrates substantial improvements in both visual quality and controllability. TV2TV also scales to natural videos, as we show by augmenting sports videos with interleaved natural language action descriptions using vision-language models (VLMs). Training TV2TV on this corpus yields strong visual quality and prompt alignment, showcasing the model's ability to reason about and generate complex real-world action sequences. Together, these results highlight TV2TV as a promising step toward video generation with open-ended textual reasoning and control.

Hugo Touvron, Matthieu Cord, Maxime Oquab et al.

We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochastic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, ``submodels'', with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed cosub, uses a single set of weights, and does not involve a pre-trained external model or temporal averaging. Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our approach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains 87.4% top-1 acc. @448 on ImageNet-val.

Tariq Berrada, Camille Couprie, Karteek Alahari et al.

Although instance segmentation methods have improved considerably, the dominant paradigm is to rely on fully-annotated training images, which are tedious to obtain. To alleviate this reliance, and boost results, semi-supervised approaches leverage unlabeled data as an additional training signal that limits overfitting to the labeled samples. In this context, we present novel design choices to significantly improve teacher-student distillation models. In particular, we (i) improve the distillation approach by introducing a novel "guided burn-in" stage, and (ii) evaluate different instance segmentation architectures, as well as backbone networks and pre-training strategies. Contrary to previous work which uses only supervised data for the burn-in period of the student model, we also use guidance of the teacher model to exploit unlabeled data in the burn-in period. Our improved distillation approach leads to substantial improvements over previous state-of-the-art results. For example, on the Cityscapes dataset we improve mask-AP from 23.7 to 33.9 when using labels for 10\% of images, and on the COCO dataset we improve mask-AP from 18.3 to 34.1 when using labels for only 1\% of the training data.

Pietro Astolfi, Arantxa Casanova, Jakob Verbeek et al.

Data augmentation has become a crucial component to train state-of-the-art visual representation models. However, handcrafting combinations of transformations that lead to improved performances is a laborious task, which can result in visually unrealistic samples. To overcome these limitations, recent works have explored the use of generative models as learnable data augmentation tools, showing promising results in narrow application domains, e.g., few-shot learning and low-data medical imaging. In this paper, we introduce a data augmentation module, called DA_IC-GAN, which leverages instance-conditioned GAN generations and can be used off-the-shelf in conjunction with most state-of-the-art training recipes. We showcase the benefits of DA_IC-GAN by plugging it out-of-the-box into the supervised training of ResNets and DeiT models on the ImageNet dataset, and achieving accuracy boosts up to between 1%p and 2%p with the highest capacity models. Moreover, the learnt representations are shown to be more robust than the baselines when transferred to a handful of out-of-distribution datasets, and exhibit increased invariance to variations of instance and viewpoints. We additionally couple DA_IC-GAN with a self-supervised training recipe and show that we can also achieve an improvement of 1%p in accuracy in some settings. With this work, we strengthen the evidence on the potential of learnable data augmentations to improve visual representation learning, paving the road towards non-handcrafted augmentations in model training.

Shengbang Tong, David Fan, John Nguyen et al.

The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision, to train on diverse data including text, video, image-text pairs, and even action-conditioned video. Our experiments yield four key insights: (i) Representation Autoencoder (RAE) provides an optimal unified visual representation by excelling at both visual understanding and generation; (ii) visual and language data are complementary and yield synergy for downstream capabilities; (iii) unified multimodal pretraining leads naturally to world modeling, with capabilities emerging from general training; and (iv) Mixture-of-Experts (MoE) enables efficient and effective multimodal scaling while naturally inducing modality specialization. Through IsoFLOP analysis, we compute scaling laws for both modalities and uncover a scaling asymmetry: vision is significantly more data-hungry than language. We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language while accommodating the data-intensive nature of vision, paving the way for truly unified multimodal models.

Arantxa Casanova, Marlène Careil, Adriana Romero-Soriano et al.

Recent advances in conditional generative image models have enabled impressive results. On the one hand, text-based conditional models have achieved remarkable generation quality, by leveraging large-scale datasets of image-text pairs. To enable fine-grained controllability, however, text-based models require long prompts, whose details may be ignored by the model. On the other hand, layout-based conditional models have also witnessed significant advances. These models rely on bounding boxes or segmentation maps for precise spatial conditioning in combination with coarse semantic labels. The semantic labels, however, cannot be used to express detailed appearance characteristics. In this paper, we approach fine-grained scene controllability through image collages which allow a rich visual description of the desired scene as well as the appearance and location of the objects therein, without the need of class nor attribute labels. We introduce "mixing and matching scenes" (M&Ms), an approach that consists of an adversarially trained generative image model which is conditioned on appearance features and spatial positions of the different elements in a collage, and integrates these into a coherent image. We train our model on the OpenImages (OI) dataset and evaluate it on collages derived from OI and MS-COCO datasets. Our experiments on the OI dataset show that M&Ms outperforms baselines in terms of fine-grained scene controllability while being very competitive in terms of image quality and sample diversity. On the MS-COCO dataset, we highlight the generalization ability of our model by outperforming DALL-E in terms of the zero-shot FID metric, despite using two magnitudes fewer parameters and data. Collage based generative models have the potential to advance content creation in an efficient and effective way as they are intuitive to use and yield high quality generations.

Marlène Careil, Jakob Verbeek, Stéphane Lathuilière

Semantic image synthesis aims to generate photo realistic images given a semantic segmentation map. Despite much recent progress, training them still requires large datasets of images annotated with per-pixel label maps that are extremely tedious to obtain. To alleviate the high annotation cost, we propose a transfer method that leverages a model trained on a large source dataset to improve the learning ability on small target datasets via estimated pairwise relations between source and target classes. The class affinity matrix is introduced as a first layer to the source model to make it compatible with the target label maps, and the source model is then further finetuned for the target domain. To estimate the class affinities we consider different approaches to leverage prior knowledge: semantic segmentation on the source domain, textual label embeddings, and self-supervised vision features. We apply our approach to GAN-based and diffusion-based architectures for semantic synthesis. Our experiments show that the different ways to estimate class affinity can be effectively combined, and that our approach significantly improves over existing state-of-the-art transfer approaches for generative image models.

João Maria Janeiro, Stanislav Frolov, Alaaeldin El-Nouby et al.

Reducing the data footprint of visual content via image compression is essential to reduce storage requirements, but also to reduce the bandwidth and latency requirements for transmission. In particular, the use of compressed images allows for faster transfer of data, and faster response times for visual recognition in edge devices that rely on cloud-based services. In this paper, we first analyze the impact of image compression using traditional codecs, as well as recent state-of-the-art neural compression approaches, on three visual recognition tasks: image classification, object detection, and semantic segmentation. We consider a wide range of compression levels, ranging from 0.1 to 2 bits-per-pixel (bpp). We find that for all three tasks, the recognition ability is significantly impacted when using strong compression. For example, for segmentation mIoU is reduced from 44.5 to 30.5 mIoU when compressing to 0.1 bpp using the best compression model we evaluated. Second, we test to what extent this performance drop can be ascribed to a loss of relevant information in the compressed image, or to a lack of generalization of visual recognition models to images with compression artefacts. We find that to a large extent the performance loss is due to the latter: by finetuning the recognition models on compressed training images, most of the performance loss is recovered. For example, bringing segmentation accuracy back up to 42 mIoU, i.e. recovering 82% of the original drop in accuracy.

Ekaterina Iakovleva, Karteek Alahari, Jakob Verbeek

Deep convolutional networks are ubiquitous in computer vision, due to their excellent performance across different tasks for various domains. Models are, however, often trained in isolation for each task, failing to exploit relatedness between tasks and domains to learn more compact models that generalise better in low-data regimes. Multi-domain learning aims to handle related tasks, such as image classification across multiple domains, simultaneously. Previous work on this problem explored the use of a pre-trained and fixed domain-agnostic base network, in combination with smaller learnable domain-specific adaptation modules. In this paper, we introduce Modulation Adapters, which update the convolutional filter weights of the model in a multiplicative manner for each task. Parameterising these adaptation weights in a factored manner allows us to scale the number of per-task parameters in a flexible manner, and to strike different parameter-accuracy trade-offs. We evaluate our approach on the Visual Decathlon challenge, composed of ten image classification tasks across different domains, and on the ImageNet-to-Sketch benchmark, which consists of six image classification tasks. Our approach yields excellent results, with accuracies that are comparable to or better than those of existing state-of-the-art approaches.

Tariq Berrada Ifriqi, John Nguyen, Karteek Alahari et al.

We present Flowception, a novel non-autoregressive and variable-length video generation framework. Flowception learns a probability path that interleaves discrete frame insertions with continuous frame denoising. Compared to autoregressive methods, Flowception alleviates error accumulation/drift as the frame insertion mechanism during sampling serves as an efficient compression mechanism to handle long-term context. Compared to full-sequence flows, our method reduces FLOPs for training three-fold, while also being more amenable to local attention variants, and allowing to learn the length of videos jointly with their content. Quantitative experimental results show improved FVD and VBench metrics over autoregressive and full-sequence baselines, which is further validated with qualitative results. Finally, by learning to insert and denoise frames in a sequence, Flowception seamlessly integrates different tasks such as image-to-video generation and video interpolation.

Marlène Careil, Stéphane Lathuilière, Camille Couprie et al.

Generative image models have been extensively studied in recent years. In the unconditional setting, they model the marginal distribution from unlabelled images. To allow for more control, image synthesis can be conditioned on semantic segmentation maps that instruct the generator the position of objects in the image. While these two tasks are intimately related, they are generally studied in isolation. We propose OCO-GAN, for Optionally COnditioned GAN, which addresses both tasks in a unified manner, with a shared image synthesis network that can be conditioned either on semantic maps or directly on latents. Trained adversarially in an end-to-end approach with a shared discriminator, we are able to leverage the synergy between both tasks. We experiment with Cityscapes, COCO-Stuff, ADE20K datasets in a limited data, semi-supervised and full data regime and obtain excellent performance, improving over existing hybrid models that can generate both with and without conditioning in all settings. Moreover, our results are competitive or better than state-of-the art specialised unconditional and conditional models.

Melissa Hall, Oscar Mañas, Reyhane Askari-Hemmat et al.

As the use of text-to-image generative models increases, so does the adoption of automatic benchmarking methods used in their evaluation. However, while metrics and datasets abound, there are few unified benchmarking libraries that provide a framework for performing evaluations across many datasets and metrics. Furthermore, the rapid introduction of increasingly robust benchmarking methods requires that evaluation libraries remain flexible to new datasets and metrics. Finally, there remains a gap in synthesizing evaluations in order to deliver actionable takeaways about model performance. To enable unified, flexible, and actionable evaluations, we introduce EvalGIM (pronounced ''EvalGym''), a library for evaluating generative image models. EvalGIM contains broad support for datasets and metrics used to measure quality, diversity, and consistency of text-to-image generative models. In addition, EvalGIM is designed with flexibility for user customization as a top priority and contains a structure that allows plug-and-play additions of new datasets and metrics. To enable actionable evaluation insights, we introduce ''Evaluation Exercises'' that highlight takeaways for specific evaluation questions. The Evaluation Exercises contain easy-to-use and reproducible implementations of two state-of-the-art evaluation methods of text-to-image generative models: consistency-diversity-realism Pareto Fronts and disaggregated measurements of performance disparities across groups. EvalGIM also contains Evaluation Exercises that introduce two new analysis methods for text-to-image generative models: robustness analyses of model rankings and balanced evaluations across different prompt styles. We encourage text-to-image model exploration with EvalGIM and invite contributions at https://github.com/facebookresearch/EvalGIM/.

Enrico Fini, Pietro Astolfi, Adriana Romero-Soriano et al.

Contrastive learning has emerged as an efficient framework to learn multimodal representations. CLIP, a seminal work in this area, achieved impressive results by training on paired image-text data using the contrastive loss. Recent work claims improvements over CLIP using additional non-contrastive losses inspired from self-supervised learning. However, it is sometimes hard to disentangle the contribution of these additional losses from other implementation details, e.g., data augmentation or regularization techniques, used to train the model. To shed light on this matter, in this paper, we first propose, implement and evaluate several baselines obtained by combining contrastive learning with recent advances in self-supervised learning. In particular, we use the loss functions that were proven successful for visual self-supervised learning to align image and text modalities. We find that these baselines outperform a basic implementation of CLIP. However, when a stronger training recipe is employed, the advantage disappears. Indeed, we find that a simple CLIP baseline can also be improved substantially, up to a 25% relative improvement on downstream zero-shot tasks, by using well-known training techniques that are popular in other subfields. Moreover, we discover that it is enough to apply image and text augmentations to make up for most of the improvement attained by prior works. With our improved training recipe for CLIP, we obtain state-of-the-art performance on four standard datasets, and consistently outperform prior work (up to +4% on the largest dataset), while being substantially simpler. The code is available at https://github.com/facebookresearch/clip-rocket

Arantxa Casanova, Marlène Careil, Jakob Verbeek et al.

Generative Adversarial Networks (GANs) can generate near photo realistic images in narrow domains such as human faces. Yet, modeling complex distributions of datasets such as ImageNet and COCO-Stuff remains challenging in unconditional settings. In this paper, we take inspiration from kernel density estimation techniques and introduce a non-parametric approach to modeling distributions of complex datasets. We partition the data manifold into a mixture of overlapping neighborhoods described by a datapoint and its nearest neighbors, and introduce a model, called instance-conditioned GAN (IC-GAN), which learns the distribution around each datapoint. Experimental results on ImageNet and COCO-Stuff show that IC-GAN significantly improves over unconditional models and unsupervised data partitioning baselines. Moreover, we show that IC-GAN can effortlessly transfer to datasets not seen during training by simply changing the conditioning instances, and still generate realistic images. Finally, we extend IC-GAN to the class-conditional case and show semantically controllable generation and competitive quantitative results on ImageNet; while improving over BigGAN on ImageNet-LT. Code and trained models to reproduce the reported results are available at https://github.com/facebookresearch/ic_gan.

Théophane Vallaeys, Mustafa Shukor, Matthieu Cord et al.

The abilities of large language models (LLMs) have recently progressed to unprecedented levels, paving the way to novel applications in a wide variety of areas. In computer vision, LLMs can be used to prime vision-language tasks such image captioning and visual question answering when coupled with pre-trained vision backbones. While different approaches have been explored to interface LLMs with ``perceptual backbones'' that process, e.g., visual or audio data, they are often explored for different tasks, different datasets, and using different perceptual backbones and language models, hindering direct comparison of the interfacing mechanisms. To remedy this lack of comparability between methods, we present an extensive experimental evaluation of different interfacing mechanisms, across multiple tasks (including image, video, and audio captioning as well as visual question answering), datasets and backbones, paying special attention to low-data settings. We find improved performance using existing mechanisms over state-of-the-art results, and identify a new interfacing mechanism that yields (near) optimal results across different tasks, while obtaining a 4x reduction in training time.

Tariq Berrada, Jakob Verbeek, Camille Couprie et al.

Semantic image synthesis, i.e., generating images from user-provided semantic label maps, is an important conditional image generation task as it allows to control both the content as well as the spatial layout of generated images. Although diffusion models have pushed the state of the art in generative image modeling, the iterative nature of their inference process makes them computationally demanding. Other approaches such as GANs are more efficient as they only need a single feed-forward pass for generation, but the image quality tends to suffer on large and diverse datasets. In this work, we propose a new class of GAN discriminators for semantic image synthesis that generates highly realistic images by exploiting feature backbone networks pre-trained for tasks such as image classification. We also introduce a new generator architecture with better context modeling and using cross-attention to inject noise into latent variables, leading to more diverse generated images. Our model, which we dub DP-SIMS, achieves state-of-the-art results in terms of image quality and consistency with the input label maps on ADE-20K, COCO-Stuff, and Cityscapes, surpassing recent diffusion models while requiring two orders of magnitude less compute for inference.

Tariq Berrada, Pietro Astolfi, Melissa Hall et al.

Latent diffusion models (LDMs) power state-of-the-art high-resolution generative image models. LDMs learn the data distribution in the latent space of an autoencoder (AE) and produce images by mapping the generated latents into RGB image space using the AE decoder. While this approach allows for efficient model training and sampling, it induces a disconnect between the training of the diffusion model and the decoder, resulting in a loss of detail in the generated images. To remediate this disconnect, we propose to leverage the internal features of the decoder to define a latent perceptual loss (LPL). This loss encourages the models to create sharper and more realistic images. Our loss can be seamlessly integrated with common autoencoders used in latent diffusion models, and can be applied to different generative modeling paradigms such as DDPM with epsilon and velocity prediction, as well as flow matching. Extensive experiments with models trained on three datasets at 256 and 512 resolution show improved quantitative -- with boosts between 6% and 20% in FID -- and qualitative results when using our perceptual loss.

Reyhane Askari-Hemmat, Mohammad Pezeshki, Elvis Dohmatob et al.

Inspired by the principle of deliberate practice in human learning, we propose Deliberate Practice for Synthetic Data Generation (DP), a novel framework that improves sample efficiency through dynamic synthetic data generation. Prior work has shown that scaling synthetic data is inherently challenging, as naively adding new data leads to diminishing returns. To address this, pruning has been identified as a key mechanism for improving scaling, enabling models to focus on the most informative synthetic samples. Rather than generating a large dataset and pruning it afterward, DP efficiently approximates the direct generation of informative samples. We theoretically show how training on challenging, informative examples improves scaling laws and empirically validate that DP achieves better scaling performance with significantly fewer training samples and iterations. On ImageNet-100, DP generates 3.4x fewer samples and requires six times fewer iterations, while on ImageNet-1k, it generates 8x fewer samples with a 30 percent reduction in iterations, all while achieving superior performance compared to prior work.

Tariq Berrada Ifriqi, Pietro Astolfi, Melissa Hall et al.

Large-scale training of latent diffusion models (LDMs) has enabled unprecedented quality in image generation. However, the key components of the best performing LDM training recipes are oftentimes not available to the research community, preventing apple-to-apple comparisons and hindering the validation of progress in the field. In this work, we perform an in-depth study of LDM training recipes focusing on the performance of models and their training efficiency. To ensure apple-to-apple comparisons, we re-implement five previously published models with their corresponding recipes. Through our study, we explore the effects of (i)~the mechanisms used to condition the generative model on semantic information (e.g., text prompt) and control metadata (e.g., crop size, random flip flag, etc.) on the model performance, and (ii)~the transfer of the representations learned on smaller and lower-resolution datasets to larger ones on the training efficiency and model performance. We then propose a novel conditioning mechanism that disentangles semantic and control metadata conditionings and sets a new state-of-the-art in class-conditional generation on the ImageNet-1k dataset -- with FID improvements of 7% on 256 and 8% on 512 resolutions -- as well as text-to-image generation on the CC12M dataset -- with FID improvements of 8% on 256 and 23% on 512 resolution.

Théophane Vallaeys, Matthew Muckley, Jakob Verbeek et al.

Vector quantization is a fundamental technique for compression and large-scale nearest neighbor search. For high-accuracy operating points, multi-codebook quantization associates data vectors with one element from each of multiple codebooks. An example is residual quantization (RQ), which iteratively quantizes the residual error of previous steps. Dependencies between the different parts of the code are, however, ignored in RQ, which leads to suboptimal rate-distortion performance. QINCo recently addressed this inefficiency by using a neural network to determine the quantization codebook in RQ based on the vector reconstruction from previous steps. In this paper we introduce QINCo2 which extends and improves QINCo with (i) improved vector encoding using codeword pre-selection and beam-search, (ii) a fast approximate decoder leveraging codeword pairs to establish accurate short-lists for search, and (iii) an optimized training procedure and network architecture. We conduct experiments on four datasets to evaluate QINCo2 for vector compression and billion-scale nearest neighbor search. We obtain outstanding results in both settings, improving the state-of-the-art reconstruction MSE by 34% for 16-byte vector compression on BigANN, and search accuracy by 24% with 8-byte encodings on Deep1M.

François Porcher, Camille Couprie, Marc Szafraniec et al.

Despite the availability of large datasets for tasks like image classification and image-text alignment, labeled data for more complex recognition tasks, such as detection and segmentation, is less abundant. In particular, for instance segmentation annotations are time-consuming to produce, and the distribution of instances is often highly skewed across classes. While semi-supervised teacher-student distillation methods show promise in leveraging vast amounts of unlabeled data, they suffer from miscalibration, resulting in overconfidence in frequently represented classes and underconfidence in rarer ones. Additionally, these methods encounter difficulties in efficiently learning from a limited set of examples. We introduce a dual-strategy to enhance the teacher model's training process, substantially improving the performance on few-shot learning. Secondly, we propose a calibration correction mechanism that that enables the student model to correct the teacher's calibration errors. Using our approach, we observed marked improvements over a state-of-the-art supervised baseline performance on the LVIS dataset, with an increase of 2.8% in average precision (AP) and 10.3% gain in AP for rare classes.

Théophane Vallaeys, Jakob Verbeek, Matthieu Cord

Tokenizers are a key component of state-of-the-art generative image models, extracting the most important features from the signal while reducing data dimension and redundancy. Most current tokenizers are based on KL-regularized variational autoencoders (KL-VAE), trained with reconstruction, perceptual and adversarial losses. Diffusion decoders have been proposed as a more principled alternative to model the distribution over images conditioned on the latent. However, matching the performance of KL-VAE still requires adversarial losses, as well as a higher decoding time due to iterative sampling. To address these limitations, we introduce a new pixel diffusion decoder architecture for improved scaling and training stability, benefiting from transformer components and GAN-free training. We use distillation to replicate the performance of the diffusion decoder in an efficient single-step decoder. This makes SSDD the first diffusion decoder optimized for single-step reconstruction trained without adversarial losses, reaching higher reconstruction quality and faster sampling than KL-VAE. In particular, SSDD improves reconstruction FID from $0.87$ to $0.50$ with $1.4\times$ higher throughput and preserve generation quality of DiTs with $3.8\times$ faster sampling. As such, SSDD can be used as a drop-in replacement for KL-VAE, and for building higher-quality and faster generative models.

Nicola Dall'Asen, Xiaofeng Zhang, Reyhane Askari Hemmat et al.

Conditional image generative models hold considerable promise to produce infinite amounts of synthetic training data. Yet, recent progress in generation quality has come at the expense of generation diversity, limiting the utility of these models as a source of synthetic training data. Although guidance-based approaches have been introduced to improve the utility of generated data by focusing on quality or diversity, the (implicit or explicit) utility functions oftentimes disregard the potential distribution shift between synthetic and real data. In this work, we introduce Chamfer Guidance: a training-free guidance approach which leverages a handful of real exemplar images to characterize the quality and diversity of synthetic data. We show that by leveraging the proposed Chamfer Guidance, we can boost the diversity of the generations w.r.t. a dataset of real images while maintaining or improving the generation quality on ImageNet-1k and standard geo-diversity benchmarks. Our approach achieves state-of-the-art few-shot performance with as little as 2 exemplar real images, obtaining 96.4% in terms of precision, and 86.4% in terms of distributional coverage, which increase to 97.5% and 92.7%, respectively, when using 32 real images. We showcase the benefits of the Chamfer Guidance generation by training downstream image classifiers on synthetic data, achieving accuracy boost of up to 15% for in-distribution over the baselines, and up to 16% in out-of-distribution. Furthermore, our approach does not require using the unconditional model, and thus obtains a 31% reduction in FLOPs w.r.t. classifier-free-guidance-based approaches at sampling time.

Tariq Berrada Ifriqi, Adriana Romero-Soriano, Michal Drozdzal et al.

Guidance techniques are commonly used in diffusion and flow models to improve image quality and consistency for conditional generative tasks such as class-conditional and text-to-image generation. In particular, classifier-free guidance (CFG) -- the most widely adopted guidance technique -- contrasts conditional and unconditional predictions to improve the generated images. This results, however, in trade-offs across quality, diversity and consistency, improving some at the expense of others. While recent work has shown that it is possible to disentangle these factors to some extent, such methods come with an overhead of requiring an additional (weaker) model, or require more forward passes per sampling step. In this paper, we propose Entropy Rectifying Guidance (ERG), a simple and effective guidance mechanism based on inference-time changes in the attention mechanism of state-of-the-art diffusion transformer architectures, which allows for simultaneous improvements over image quality, diversity and prompt consistency. ERG is more general than CFG and similar guidance techniques, as it extends to unconditional sampling. ERG results in significant improvements in various generation tasks such as text-to-image, class-conditional and unconditional image generation. We also show that ERG can be seamlessly combined with other recent guidance methods such as CADS and APG, further boosting generation performance.

Krunoslav Lehman Pavasovic, Jakob Verbeek, Giulio Biroli et al.

Classifier-Free Guidance (CFG) is a widely adopted technique in diffusion and flow-based generative models, enabling high-quality conditional generation. A key theoretical challenge is characterizing the distribution induced by CFG, particularly in high-dimensional settings relevant to real-world data. Previous works have shown that CFG modifies the target distribution, steering it towards a distribution sharper than the target one, more shifted towards the boundary of the class. In this work, we provide a high-dimensional analysis of CFG, showing that these distortions vanish as the data dimension grows. We present a blessing-of-dimensionality result demonstrating that in sufficiently high and infinite dimensions, CFG accurately reproduces the target distribution. Using our high-dimensional theory, we show that there is a large family of guidances enjoying this property, in particular non-linear CFG generalizations. We study a simple non-linear power-law version, for which we demonstrate improved robustness, sample fidelity and diversity. Our findings are validated with experiments on class-conditional and text-to-image generation using state-of-the-art diffusion and flow-matching models.

Xiangyi Chen, Théophane Vallaeys, Maha Elbayad et al.

Recent advances in Vision-Language Models (VLMs) have enabled unified understanding across text and images, yet equipping these models with robust image generation capabilities remains challenging. Existing approaches often rely on reconstruction-oriented autoencoders or complex bridging mechanisms, leading to misalignment between understanding and generation representations, or architectural complexity. In this work, we propose VUGEN, a novel framework that explicitly leverages VLM's pretrained visual understanding priors for efficient and high-quality image generation. Our approach first transforms the high-dimensional latent space of the VLM's native vision encoder into a lower-dimensional, tractable distribution that maximally preserves visual information. The VLM is then trained to sample within this reduced latent space, ensuring alignment with its visual understanding capabilities. Finally, a dedicated pixel decoder maps these generated latents back to the image space. We find that a VAE-free pixel diffusion decoder to be on par or better than commonly used complex latent diffusion decoders that internally rely on VAE latents. Extensive experiments demonstrate that VUGEN achieves superior image generation performance, improving DPG Bench from 71.17 to 74.32 and FID from 11.86 to 9.06 on COCO, while fully preserving the VLM's original understanding capabilities.

Pietro Astolfi, Marlene Careil, Melissa Hall et al.

Building world models that accurately and comprehensively represent the real world is the utmost aspiration for conditional image generative models as it would enable their use as world simulators. For these models to be successful world models, they should not only excel at image quality and prompt-image consistency but also ensure high representation diversity. However, current research in generative models mostly focuses on creative applications that are predominantly concerned with human preferences of image quality and aesthetics. We note that generative models have inference time mechanisms - or knobs - that allow the control of generation consistency, quality, and diversity. In this paper, we use state-of-the-art text-to-image and image-and-text-to-image models and their knobs to draw consistency-diversity-realism Pareto fronts that provide a holistic view on consistency-diversity-realism multi-objective. Our experiments suggest that realism and consistency can both be improved simultaneously; however there exists a clear tradeoff between realism/consistency and diversity. By looking at Pareto optimal points, we note that earlier models are better at representation diversity and worse in consistency/realism, and more recent models excel in consistency/realism while decreasing significantly the representation diversity. By computing Pareto fronts on a geodiverse dataset, we find that the first version of latent diffusion models tends to perform better than more recent models in all axes of evaluation, and there exist pronounced consistency-diversity-realism disparities between geographical regions. Overall, our analysis clearly shows that there is no best model and the choice of model should be determined by the downstream application. With this analysis, we invite the research community to consider Pareto fronts as an analytical tool to measure progress towards world models.

Iris A. M. Huijben, Matthijs Douze, Matthew Muckley et al.

Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.

Alaaeldin El-Nouby, Hugo Touvron, Mathilde Caron et al.

Following their success in natural language processing, transformers have recently shown much promise for computer vision. The self-attention operation underlying transformers yields global interactions between all tokens ,i.e. words or image patches, and enables flexible modelling of image data beyond the local interactions of convolutions. This flexibility, however, comes with a quadratic complexity in time and memory, hindering application to long sequences and high-resolution images. We propose a "transposed" version of self-attention that operates across feature channels rather than tokens, where the interactions are based on the cross-covariance matrix between keys and queries. The resulting cross-covariance attention (XCA) has linear complexity in the number of tokens, and allows efficient processing of high-resolution images. Our cross-covariance image transformer (XCiT) is built upon XCA. It combines the accuracy of conventional transformers with the scalability of convolutional architectures. We validate the effectiveness and generality of XCiT by reporting excellent results on multiple vision benchmarks, including image classification and self-supervised feature learning on ImageNet-1k, object detection and instance segmentation on COCO, and semantic segmentation on ADE20k.

Hugo Touvron, Piotr Bojanowski, Mathilde Caron et al.

We present ResMLP, an architecture built entirely upon multi-layer perceptrons for image classification. It is a simple residual network that alternates (i) a linear layer in which image patches interact, independently and identically across channels, and (ii) a two-layer feed-forward network in which channels interact independently per patch. When trained with a modern training strategy using heavy data-augmentation and optionally distillation, it attains surprisingly good accuracy/complexity trade-offs on ImageNet. We also train ResMLP models in a self-supervised setup, to further remove priors from employing a labelled dataset. Finally, by adapting our model to machine translation we achieve surprisingly good results. We share pre-trained models and our code based on the Timm library.

Nitika Verma, Adnane Boukhayma, Jakob Verbeek et al.

Convolutional networks have been extremely successful for regular data structures such as 2D images and 3D voxel grids. The transposition to meshes is, however, not straight-forward due to their irregular structure. We explore how the dual, face-based representation of triangular meshes can be leveraged as a data structure for graph convolutional networks. In the dual mesh, each node (face) has a fixed number of neighbors, which makes the networks less susceptible to overfitting on the mesh topology, and also al-lows the use of input features that are naturally defined over faces, such as surface normals and face areas. We evaluate the dual approach on the shape correspondence task on theFaust human shape dataset and variants of it with differ-ent mesh topologies. Our experiments show that results of graph convolutional networks improve when defined over the dual rather than primal mesh. Moreover, our models that explicitly leverage the neighborhood regularity of dual meshes allow improving results further while being more robust to changes in the mesh topology.

Ekaterina Iakovleva, Jakob Verbeek, Karteek Alahari

We propose a novel amortized variational inference scheme for an empirical Bayes meta-learning model, where model parameters are treated as latent variables. We learn the prior distribution over model parameters conditioned on limited training data using a variational autoencoder approach. Our framework proposes sharing the same amortized inference network between the conditional prior and variational posterior distributions over the model parameters. While the posterior leverages both the labeled support and query data, the conditional prior is based only on the labeled support data. We show that in earlier work, relying on Monte-Carlo approximation, the conditional prior collapses to a Dirac delta function. In contrast, our variational approach prevents this collapse and preserves uncertainty over the model parameters. We evaluate our approach on the miniImageNet, CIFAR-FS and FC100 datasets, and present results demonstrating its advantages over previous work.

Roman Klokov, Edmond Boyer, Jakob Verbeek

Generative models have proven effective at modeling 3D shapes and their statistical variations. In this paper we investigate their application to point clouds, a 3D shape representation widely used in computer vision for which, however, only few generative models have yet been proposed. We introduce a latent variable model that builds on normalizing flows with affine coupling layers to generate 3D point clouds of an arbitrary size given a latent shape representation. To evaluate its benefits for shape modeling we apply this model for generation, autoencoding, and single-view shape reconstruction tasks. We improve over recent GAN-based models in terms of most metrics that assess generation and autoencoding. Compared to recent work based on continuous flows, our model offers a significant speedup in both training and inference times for similar or better performance. For single-view shape reconstruction we also obtain results on par with state-of-the-art voxel, point cloud, and mesh-based methods.

Maha Elbayad, Michael Ustaszewski, Emmanuelle Esperança-Rodier et al.

We conduct in this work an evaluation study comparing offline and online neural machine translation architectures. Two sequence-to-sequence models: convolutional Pervasive Attention (Elbayad et al. 2018) and attention-based Transformer (Vaswani et al. 2017) are considered. We investigate, for both architectures, the impact of online decoding constraints on the translation quality through a carefully designed human evaluation on English-German and German-English language pairs, the latter being particularly sensitive to latency constraints. The evaluation results allow us to identify the strengths and shortcomings of each model when we shift to the online setup.

Maha Elbayad, Laurent Besacier, Jakob Verbeek

Simultaneous machine translation consists in starting output generation before the entire input sequence is available. Wait-k decoders offer a simple but efficient approach for this problem. They first read k source tokens, after which they alternate between producing a target token and reading another source token. We investigate the behavior of wait-k decoding in low resource settings for spoken corpora using IWSLT datasets. We improve training of these models using unidirectional encoders, and training across multiple values of k. Experiments with Transformer and 2D-convolutional architectures show that our wait-k models generalize well across a wide range of latency levels. We also show that the 2D-convolution architecture is competitive with Transformers for simultaneous translation of spoken language.

Adria Ruiz, Jakob Verbeek

Inference in deep neural networks can be computationally expensive, and networks capable of anytime inference are important in mscenarios where the amount of compute or quantity of input data varies over time. In such networks the inference process can interrupted to provide a result faster, or continued to obtain a more accurate result. We propose Hierarchical Neural Ensembles (HNE), a novel framework to embed an ensemble of multiple networks in a hierarchical tree structure, sharing intermediate layers. In HNE we control the complexity of inference on-the-fly by evaluating more or less models in the ensemble. Our second contribution is a novel hierarchical distillation method to boost the prediction accuracy of small ensembles. This approach leverages the nested structure of our ensembles, to optimally allocate accuracy and diversity across the individual models. Our experiments show that, compared to previous anytime inference models, HNE provides state-of-the-art accuracy-computate trade-offs on the CIFAR-10/100 and ImageNet datasets.

Xiaotian Li, Shuzhe Wang, Yi Zhao et al.

Visual localization is critical to many applications in computer vision and robotics. To address single-image RGB localization, state-of-the-art feature-based methods match local descriptors between a query image and a pre-built 3D model. Recently, deep neural networks have been exploited to regress the mapping between raw pixels and 3D coordinates in the scene, and thus the matching is implicitly performed by the forward pass through the network. However, in a large and ambiguous environment, learning such a regression task directly can be difficult for a single network. In this work, we present a new hierarchical scene coordinate network to predict pixel scene coordinates in a coarse-to-fine manner from a single RGB image. The network consists of a series of output layers, each of them conditioned on the previous ones. The final output layer predicts the 3D coordinates and the others produce progressively finer discrete location labels. The proposed method outperforms the baseline regression-only network and allows us to train compact models which scale robustly to large environments. It sets a new state-of-the-art for single-image RGB localization performance on the 7-Scenes, 12-Scenes, Cambridge Landmarks datasets, and three combined scenes. Moreover, for large-scale outdoor localization on the Aachen Day-Night dataset, we present a hybrid approach which outperforms existing scene coordinate regression methods, and reduces significantly the performance gap w.r.t. explicit feature matching methods.

Roman Klokov, Jakob Verbeek, Edmond Boyer

We study end-to-end learning strategies for 3D shape inference from images, in particular from a single image. Several approaches in this direction have been investigated that explore different shape representations and suitable learning architectures. We focus instead on the underlying probabilistic mechanisms involved and contribute a more principled probabilistic inference-based reconstruction framework, which we coin Probabilistic Reconstruction Networks. This framework expresses image conditioned 3D shape inference through a family of latent variable models, and naturally decouples the choice of shape representations from the inference itself. Moreover, it suggests different options for the image conditioning and allows training in two regimes, using either Monte Carlo or variational approximation of the marginal likelihood. Using our Probabilistic Reconstruction Networks we obtain single image 3D reconstruction results that set a new state of the art on the ShapeNet dataset in terms of the intersection over union and earth mover's distance evaluation metrics. Interestingly, we obtain these results using a basic voxel grid representation, improving over recent work based on finer point cloud or mesh based representations.

Adria Ruiz, Jakob Verbeek

Despite the outstanding performance of convolutional neural networks (CNNs) for many vision tasks, the required computational cost during inference is problematic when resources are limited. In this context, we propose Convolutional Neural Mixture Models (CNMMs), a probabilistic model embedding a large number of CNNs that can be jointly trained and evaluated in an efficient manner. Within the proposed framework, we present different mechanisms to prune subsets of CNNs from the mixture, allowing to easily adapt the computational cost required for inference. Image classification and semantic segmentation experiments show that our method achieve excellent accuracy-compute trade-offs. Moreover, unlike most of previous approaches, a single CNMM provides a large range of operating points along this trade-off, without any re-training.

Adria Ruiz, Oriol Martinez, Xavier Binefa et al.

Learning disentangled representations from visual data, where different high-level generative factors are independently encoded, is of importance for many computer vision tasks. Solving this problem, however, typically requires to explicitly label all the factors of interest in training images. To alleviate the annotation cost, we introduce a learning setting which we refer to as "reference-based disentangling". Given a pool of unlabeled images, the goal is to learn a representation where a set of target factors are disentangled from others. The only supervision comes from an auxiliary "reference set" containing images where the factors of interest are constant. In order to address this problem, we propose reference-based variational autoencoders, a novel deep generative model designed to exploit the weak-supervision provided by the reference set. By addressing tasks such as feature learning, conditional image generation or attribute transfer, we validate the ability of the proposed model to learn disentangled representations from this minimal form of supervision.

Thomas Lucas, Konstantin Shmelkov, Karteek Alahari et al.

Unsupervised learning of generative models has seen tremendous progress over recent years, in particular due to generative adversarial networks (GANs), variational autoencoders, and flow-based models. GANs have dramatically improved sample quality, but suffer from two drawbacks: (i) they mode-drop, i.e., do not cover the full support of the train data, and (ii) they do not allow for likelihood evaluations on held-out data. In contrast, likelihood-based training encourages models to cover the full support of the train data, but yields poorer samples. These mutual shortcomings can in principle be addressed by training generative latent variable models in a hybrid adversarial-likelihood manner. However, we show that commonly made parametric assumptions create a conflict between them, making successful hybrid models non trivial. As a solution, we propose to use deep invertible transformations in the latent variable decoder. This approach allows for likelihood computations in image space, is more efficient than fully invertible models, and can take full advantage of adversarial training. We show that our model significantly improves over existing hybrid models: offering GAN-like samples, IS and FID scores that are competitive with fully adversarial models, and improved likelihood scores.

Mariia Vladimirova, Jakob Verbeek, Pablo Mesejo et al.

We investigate deep Bayesian neural networks with Gaussian weight priors and a class of ReLU-like nonlinearities. Bayesian neural networks with Gaussian priors are well known to induce an L2, "weight decay", regularization. Our results characterize a more intricate regularization effect at the level of the unit activations. Our main result establishes that the induced prior distribution on the units before and after activation becomes increasingly heavy-tailed with the depth of the layer. We show that first layer units are Gaussian, second layer units are sub-exponential, and units in deeper layers are characterized by sub-Weibull distributions. Our results provide new theoretical insight on deep Bayesian neural networks, which we corroborate with simulation experiments.

Xiaotian Li, Juha Ylioinas, Jakob Verbeek et al.

Image-based camera relocalization is an important problem in computer vision and robotics. Recent works utilize convolutional neural networks (CNNs) to regress for pixels in a query image their corresponding 3D world coordinates in the scene. The final pose is then solved via a RANSAC-based optimization scheme using the predicted coordinates. Usually, the CNN is trained with ground truth scene coordinates, but it has also been shown that the network can discover 3D scene geometry automatically by minimizing single-view reprojection loss. However, due to the deficiencies of the reprojection loss, the network needs to be carefully initialized. In this paper, we present a new angle-based reprojection loss, which resolves the issues of the original reprojection loss. With this new loss function, the network can be trained without careful initialization, and the system achieves more accurate results. The new loss also enables us to utilize available multi-view constraints, which further improve performance.