Patrick Pérez

Alexandre Défossez, Laurent Mazaré, Manu Orsini et al.

We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning -- such as emotion or non-speech sounds -- is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this "Inner Monologue" method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at https://github.com/kyutai-labs/moshi.

Soshi Shimada, Vladislav Golyanik, Zhi Li et al.

Marker-less monocular 3D human motion capture (MoCap) with scene interactions is a challenging research topic relevant for extended reality, robotics and virtual avatar generation. Due to the inherent depth ambiguity of monocular settings, 3D motions captured with existing methods often contain severe artefacts such as incorrect body-scene inter-penetrations, jitter and body floating. To tackle these issues, we propose HULC, a new approach for 3D human MoCap which is aware of the scene geometry. HULC estimates 3D poses and dense body-environment surface contacts for improved 3D localisations, as well as the absolute scale of the subject. Furthermore, we introduce a 3D pose trajectory optimisation based on a novel pose manifold sampling that resolves erroneous body-environment inter-penetrations. Although the proposed method requires less structured inputs compared to existing scene-aware monocular MoCap algorithms, it produces more physically-plausible poses: HULC significantly and consistently outperforms the existing approaches in various experiments and on different metrics. Project page: https://vcai.mpi-inf.mpg.de/projects/HULC/.

Eslam Abdelrahman, Liangbing Zhao, Vincent Tao Hu et al.

Diffusion models break down the challenging task of generating data from high-dimensional distributions into a series of easier denoising steps. Inspired by this paradigm, we propose a novel approach that extends the diffusion framework into modality space, decomposing the complex task of RGB image generation into simpler, interpretable stages. Our method, termed ToddlerDiffusion, cascades modality-specific models, each responsible for generating an intermediate representation, such as contours, palettes, and detailed textures, ultimately culminating in a high-quality RGB image. Instead of relying on the naive LDM concatenation conditioning mechanism to connect the different stages together, we employ Schrödinger Bridge to determine the optimal transport between different modalities. Although employing a cascaded pipeline introduces more stages, which could lead to a more complex architecture, each stage is meticulously formulated for efficiency and accuracy, surpassing Stable-Diffusion (LDM) performance. Modality composition not only enhances overall performance but enables emerging proprieties such as consistent editing, interaction capabilities, high-level interpretability, and faster convergence and sampling rate. Extensive experiments on diverse datasets, including LSUN-Churches, ImageNet, CelebHQ, and LAION-Art, demonstrate the efficacy of our approach, consistently outperforming state-of-the-art methods. For instance, ToddlerDiffusion achieves notable efficiency, matching LDM performance on LSUN-Churches while operating 2$\times$ faster with a 3$\times$ smaller architecture. The project website is available at: https://toddlerdiffusion.github.io/website/

Neil Zeghidour, Eugene Kharitonov, Manu Orsini et al.

We introduce Delayed Streams Modeling (DSM), a flexible formulation for streaming, multimodal sequence-to-sequence learning. Sequence-to-sequence generation is often cast in an offline manner, where the model consumes the complete input sequence before generating the first output timestep. Alternatively, streaming sequence-to-sequence rely on learning a policy for choosing when to advance on the input stream, or write to the output stream. DSM instead models already time-aligned streams with a decoder-only language model. By moving the alignment to a pre-processing step,and introducing appropriate delays between streams, DSM provides streaming inference of arbitrary output sequences, from any input combination, making it applicable to many sequence-to-sequence problems. In particular, given text and audio streams, automatic speech recognition (ASR) corresponds to the text stream being delayed, while the opposite gives a text-to-speech (TTS) model. We perform extensive experiments for these two major sequence-to-sequence tasks, showing that DSM provides state-of-the-art performance and latency while supporting arbitrary long sequences, being even competitive with offline baselines. Code, samples and demos are available at https://github.com/kyutai-labs/delayed-streams-modeling

Moritz Böhle, Amélie Royer, Juliette Marrie et al.

Vision-language models (VLMs) are commonly trained by inserting image tokens from a pretrained vision encoder into the textual stream of a language model. This allows text and image information to fully attend to one another within the model, but becomes extremely costly for high-resolution images, long conversations, or streaming videos, both in memory and compute. VLMs leveraging cross-attention are an efficient alternative to token insertion but exhibit a clear performance gap, in particular on tasks involving fine-grained visual details. We find that a key to improving such models is to also enable local text-to-text interaction in the dedicated cross-attention layers. Building on this, we propose CASA, Cross-Attention via Self-Attention, a simple and efficient paradigm which substantially reduces the gap with full token insertion on common image understanding benchmarks, while enjoying the same scalability as cross-attention models when applied to long-context multimodal tasks such as streaming video captioning. For samples and code, please see our project page at https://kyutai.org/casa .

Tom Labiausse, Laurent Mazaré, Edouard Grave et al.

We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart, where one waits for the end of the source utterance to start translating, adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples as well as models and inference code.

Amélie Royer, Moritz Böhle, Gabriel de Marmiesse et al.

The recent successes of Vision-Language models raise the question of how to equivalently imbue a pretrained speech model with vision understanding, an important milestone towards building a multimodal speech model able to freely converse about images. Building such a conversational Vision-Speech model brings its unique challenges: (i) paired image-speech datasets are much scarcer than their image-text counterparts, (ii) ensuring real-time latency at inference is crucial thus bringing compute and memory constraints, and (iii) the model should preserve prosodic features (e.g., speaker tone) which cannot be inferred from text alone. In this work, we introduce MoshiVis, augmenting a recent dialogue speech LLM, Moshi, with visual inputs through lightweight adaptation modules. An additional dynamic gating mechanism enables the model to more easily switch between the visual inputs and unrelated conversation topics. To reduce training costs, we design a simple one-stage, parameter-efficient fine-tuning pipeline in which we leverage a mixture of image-text (i.e., "speechless") and image-speech samples. We evaluate the model on downstream visual understanding tasks with both audio and text prompts, and report qualitative samples of interactions with MoshiVis. Our inference code will be made available, as well as the image-speech data used for audio evaluation.

Himalaya Jain, Tuan-Hung Vu, Patrick Pérez et al.

With the rapid advances in generative adversarial networks (GANs), the visual quality of synthesised scenes keeps improving, including for complex urban scenes with applications to automated driving. We address in this work a continual scene generation setup in which GANs are trained on a stream of distinct domains; ideally, the learned models should eventually be able to generate new scenes in all seen domains. This setup reflects the real-life scenario where data are continuously acquired in different places at different times. In such a continual setup, we aim for learning with zero forgetting, \IE, with no degradation in synthesis quality over earlier domains due to catastrophic forgetting. To this end, we introduce a novel framework that not only (i) enables seamless knowledge transfer in continual training but also (ii) guarantees zero forgetting with a small overhead cost. While being more memory efficient, thanks to continual learning, our model obtains better synthesis quality as compared against the brute-force solution that trains one full model for each domain. Especially, under extreme low-data regimes, our approach outperforms the brute-force one by a large margin.

Guillaume Le Moing, Tuan-Hung Vu, Himalaya Jain et al.

Despite the recent progress of generative adversarial networks (GANs) at synthesizing photo-realistic images, producing complex urban scenes remains a challenging problem. Previous works break down scene generation into two consecutive phases: unconditional semantic layout synthesis and image synthesis conditioned on layouts. In this work, we propose to condition layout generation as well for higher semantic control: given a vector of class proportions, we generate layouts with matching composition. To this end, we introduce a conditional framework with novel architecture designs and learning objectives, which effectively accommodates class proportions to guide the scene generation process. The proposed architecture also allows partial layout editing with interesting applications. Thanks to the semantic control, we can produce layouts close to the real distribution, helping enhance the whole scene generation process. On different metrics and urban scene benchmarks, our models outperform existing baselines. Moreover, we demonstrate the merit of our approach for data augmentation: semantic segmenters trained on real layout-image pairs along with additional ones generated by our approach outperform models only trained on real pairs.

Ayush Tewari, Mohamed Elgharib, Mallikarjun B R. et al.

Editing of portrait images is a very popular and important research topic with a large variety of applications. For ease of use, control should be provided via a semantically meaningful parameterization that is akin to computer animation controls. The vast majority of existing techniques do not provide such intuitive and fine-grained control, or only enable coarse editing of a single isolated control parameter. Very recently, high-quality semantically controlled editing has been demonstrated, however only on synthetically created StyleGAN images. We present the first approach for embedding real portrait images in the latent space of StyleGAN, which allows for intuitive editing of the head pose, facial expression, and scene illumination in the image. Semantic editing in parameter space is achieved based on StyleRig, a pretrained neural network that maps the control space of a 3D morphable face model to the latent space of the GAN. We design a novel hierarchical non-linear optimization problem to obtain the embedding. An identity preservation energy term allows spatially coherent edits while maintaining facial integrity. Our approach runs at interactive frame rates and thus allows the user to explore the space of possible edits. We evaluate our approach on a wide set of portrait photos, compare it to the current state of the art, and validate the effectiveness of its components in an ablation study.

Xu Yao, Gilles Puy, Patrick Pérez

We address the problem of style transfer between two photos and propose a new way to preserve photorealism. Using the single pair of photos available as input, we train a pair of deep convolution networks (convnets), each of which transfers the style of one photo to the other. To enforce photorealism, we introduce a content preserving mechanism by combining a cycle-consistency loss with a self-consistency loss. Experimental results show that this method does not suffer from typical artifacts observed in methods working in the same settings. We then further analyze some properties of these trained convnets. First, we notice that they can be used to stylize other unseen images with same known style. Second, we show that retraining only a small subset of the network parameters can be sufficient to adapt these convnets to new styles.

Gilles Puy, Patrick Pérez

We propose a new flexible deep convolutional neural network (convnet) to perform fast visual style transfer. In contrast to existing convnets that address the same task, our architecture derives directly from the structure of the gradient descent originally used to solve the style transfer problem [Gatys et al., 2016]. Like existing convnets, ours approximately solves the original problem much faster than the gradient descent. However, our network is uniquely flexible by design: it can be manipulated at runtime to enforce new constraints on the final solution. In particular, we show how to modify it to obtain a photorealistic result with no retraining. We study the modifications made by [Luan et al., 2017] to the original cost function of [Gatys et al., 2016] to achieve photorealistic style transfer. These modifications affect directly the gradient descent and can be reported on-the-fly in our network. These modifications are possible as the proposed architecture stems from unrolling the gradient descent.

Hyeongwoo Kim, Pablo Garrido, Ayush Tewari et al.

We present a novel approach that enables photo-realistic re-animation of portrait videos using only an input video. In contrast to existing approaches that are restricted to manipulations of facial expressions only, we are the first to transfer the full 3D head position, head rotation, face expression, eye gaze, and eye blinking from a source actor to a portrait video of a target actor. The core of our approach is a generative neural network with a novel space-time architecture. The network takes as input synthetic renderings of a parametric face model, based on which it predicts photo-realistic video frames for a given target actor. The realism in this rendering-to-video transfer is achieved by careful adversarial training, and as a result, we can create modified target videos that mimic the behavior of the synthetically-created input. In order to enable source-to-target video re-animation, we render a synthetic target video with the reconstructed head animation parameters from a source video, and feed it into the trained network -- thus taking full control of the target. With the ability to freely recombine source and target parameters, we are able to demonstrate a large variety of video rewrite applications without explicitly modeling hair, body or background. For instance, we can reenact the full head using interactive user-controlled editing, and realize high-fidelity visual dubbing. To demonstrate the high quality of our output, we conduct an extensive series of experiments and evaluations, where for instance a user study shows that our video edits are hard to detect.

Sanjeel Parekh, Slim Essid, Alexey Ozerov et al.

Audio-visual representation learning is an important task from the perspective of designing machines with the ability to understand complex events. To this end, we propose a novel multimodal framework that instantiates multiple instance learning. We show that the learnt representations are useful for classifying events and localizing their characteristic audio-visual elements. The system is trained using only video-level event labels without any timing information. An important feature of our method is its capacity to learn from unsynchronized audio-visual events. We achieve state-of-the-art results on a large-scale dataset of weakly-labeled audio event videos. Visualizations of localized visual regions and audio segments substantiate our system's efficacy, especially when dealing with noisy situations where modality-specific cues appear asynchronously.

Huy V. Vo, Ngoc Q. K. Duong, Patrick Perez

Scene-agnostic visual inpainting remains very challenging despite progress in patch-based methods. Recently, Pathak et al. 2016 have introduced convolutional "context encoders" (CEs) for unsupervised feature learning through image completion tasks. With the additional help of adversarial training, CEs turned out to be a promising tool to complete complex structures in real inpainting problems. In the present paper we propose to push further this key ability by relying on perceptual reconstruction losses at training time. We show on a wide variety of visual scenes the merit of the approach for structural inpainting, and confirm it through a user study. Combined with the optimization-based refinement of Yang et al. 2016 with neural patches, our context encoder opens up new opportunities for prior-free visual inpainting.

Himalaya Jain, Joaquin Zepeda, Patrick Pérez et al.

To work at scale, a complete image indexing system comprises two components: An inverted file index to restrict the actual search to only a subset that should contain most of the items relevant to the query; An approximate distance computation mechanism to rapidly scan these lists. While supervised deep learning has recently enabled improvements to the latter, the former continues to be based on unsupervised clustering in the literature. In this work, we propose a first system that learns both components within a unifying neural framework of structured binary encoding.

Ayush Tewari, Michael Zollhöfer, Pablo Garrido et al.

The reconstruction of dense 3D models of face geometry and appearance from a single image is highly challenging and ill-posed. To constrain the problem, many approaches rely on strong priors, such as parametric face models learned from limited 3D scan data. However, prior models restrict generalization of the true diversity in facial geometry, skin reflectance and illumination. To alleviate this problem, we present the first approach that jointly learns 1) a regressor for face shape, expression, reflectance and illumination on the basis of 2) a concurrently learned parametric face model. Our multi-level face model combines the advantage of 3D Morphable Models for regularization with the out-of-space generalization of a learned corrective space. We train end-to-end on in-the-wild images without dense annotations by fusing a convolutional encoder with a differentiable expert-designed renderer and a self-supervised training loss, both defined at multiple detail levels. Our approach compares favorably to the state-of-the-art in terms of reconstruction quality, better generalizes to real world faces, and runs at over 250 Hz.

Erwan Le Merrer, Patrick Perez, Gilles Trédan

The state of the art performance of deep learning models comes at a high cost for companies and institutions, due to the tedious data collection and the heavy processing requirements. Recently, [35, 22] proposed to watermark convolutional neural networks for image classification, by embedding information into their weights. While this is a clear progress towards model protection, this technique solely allows for extracting the watermark from a network that one accesses locally and entirely. Instead, we aim at allowing the extraction of the watermark from a neural network (or any other machine learning model) that is operated remotely, and available through a service API. To this end, we propose to mark the model's action itself, tweaking slightly its decision frontiers so that a set of specific queries convey the desired information. In the present paper, we formally introduce the problem and propose a novel zero-bit watermarking algorithm that makes use of adversarial model examples. While limiting the loss of performance of the protected model, this algorithm allows subsequent extraction of the watermark using only few queries. We experimented the approach on three neural networks designed for image classification, in the context of MNIST digit recognition task.

Eric Grinstein, Ngoc Duong, Alexey Ozerov et al.

'Style transfer' among images has recently emerged as a very active research topic, fuelled by the power of convolution neural networks (CNNs), and has become fast a very popular technology in social media. This paper investigates the analogous problem in the audio domain: How to transfer the style of a reference audio signal to a target audio content? We propose a flexible framework for the task, which uses a sound texture model to extract statistics characterizing the reference audio style, followed by an optimization-based audio texture synthesis to modify the target content. In contrast to mainstream optimization-based visual transfer method, the proposed process is initialized by the target content instead of random noise and the optimized loss is only about texture, not structure. These differences proved key for audio style transfer in our experiments. In order to extract features of interest, we investigate different architectures, whether pre-trained on other tasks, as done in image style transfer, or engineered based on the human auditory system. Experimental results on different types of audio signal confirm the potential of the proposed approach.

Himalaya Jain, Joaquin Zepeda, Patrick Pérez et al.

For large-scale visual search, highly compressed yet meaningful representations of images are essential. Structured vector quantizers based on product quantization and its variants are usually employed to achieve such compression while minimizing the loss of accuracy. Yet, unlike binary hashing schemes, these unsupervised methods have not yet benefited from the supervision, end-to-end learning and novel architectures ushered in by the deep learning revolution. We hence propose herein a novel method to make deep convolutional neural networks produce supervised, compact, structured binary codes for visual search. Our method makes use of a novel block-softmax non-linearity and of batch-based entropy losses that together induce structure in the learned encodings. We show that our method outperforms state-of-the-art compact representations based on deep hashing or structured quantization in single and cross-domain category retrieval, instance retrieval and classification. We make our code and models publicly available online.

Himalaya Jain, Patrick Pérez, Rémi Gribonval et al.

This paper tackles the task of storing a large collection of vectors, such as visual descriptors, and of searching in it. To this end, we propose to approximate database vectors by constrained sparse coding, where possible atom weights are restricted to belong to a finite subset. This formulation encompasses, as particular cases, previous state-of-the-art methods such as product or residual quantization. As opposed to traditional sparse coding methods, quantized sparse coding includes memory usage as a design constraint, thereby allowing us to index a large collection such as the BIGANN billion-sized benchmark. Our experiments, carried out on standard benchmarks, show that our formulation leads to competitive solutions when considering different trade-offs between learning/coding time, index size and search quality.

Nicolas Keriven, Anthony Bourrier, Rémi Gribonval et al.

Learning parameters from voluminous data can be prohibitive in terms of memory and computational requirements. We propose a "compressive learning" framework where we estimate model parameters from a sketch of the training data. This sketch is a collection of generalized moments of the underlying probability distribution of the data. It can be computed in a single pass on the training set, and is easily computable on streams or distributed datasets. The proposed framework shares similarities with compressive sensing, which aims at drastically reducing the dimension of high-dimensional signals while preserving the ability to reconstruct them. To perform the estimation task, we derive an iterative algorithm analogous to sparse reconstruction algorithms in the context of linear inverse problems. We exemplify our framework with the compressive estimation of a Gaussian Mixture Model (GMM), providing heuristics on the choice of the sketching procedure and theoretical guarantees of reconstruction. We experimentally show on synthetic data that the proposed algorithm yields results comparable to the classical Expectation-Maximization (EM) technique while requiring significantly less memory and fewer computations when the number of database elements is large. We further demonstrate the potential of the approach on real large-scale data (over 10 8 training samples) for the task of model-based speaker verification. Finally, we draw some connections between the proposed framework and approximate Hilbert space embedding of probability distributions using random features. We show that the proposed sketching operator can be seen as an innovative method to design translation-invariant kernels adapted to the analysis of GMMs. We also use this theoretical framework to derive information preservation guarantees, in the spirit of infinite-dimensional compressive sensing.

Pablo Garrido, Levi Valgaerts, Ole Rehmsen et al.

We propose an image-based, facial reenactment system that replaces the face of an actor in an existing target video with the face of a user from a source video, while preserving the original target performance. Our system is fully automatic and does not require a database of source expressions. Instead, it is able to produce convincing reenactment results from a short source video captured with an off-the-shelf camera, such as a webcam, where the user performs arbitrary facial gestures. Our reenactment pipeline is conceived as part image retrieval and part face transfer: The image retrieval is based on temporal clustering of target frames and a novel image matching metric that combines appearance and motion to select candidate frames from the source video, while the face transfer uses a 2D warping strategy that preserves the user's identity. Our system excels in simplicity as it does not rely on a 3D face model, it is robust under head motion and does not require the source and target performance to be similar. We show convincing reenactment results for videos that we recorded ourselves and for low-quality footage taken from the Internet.