Stamatios Lefkimmiatis

Denis Makhov, Dmitriy Shopkhoev, Magauiya Zhussip et al.

Post-training compression of Transformer models commonly relies on truncated singular value decomposition (SVD). However, enforcing a single shared subspace can degrade accuracy even at moderate compression. Sparse dictionary learning provides a more flexible union-of-subspaces representation, but existing approaches often suffer from iterative dictionary and coefficient updates. We propose COMPOT (Calibration-Optimized Matrix Procrustes Orthogonalization for Transformers), a training-free compression framework that uses a small calibration dataset to estimate a sparse weight factorization. COMPOT employs orthogonal dictionaries that enable closed-form Procrustes updates for the dictionary and analytical single-step sparse coding for the coefficients, eliminating iterative optimization. To handle heterogeneous layer sensitivity under a global compression budget, COMPOT further introduces a one-shot dynamic allocation strategy that adaptively redistributes layer-wise compression rates. Extensive experiments across diverse architectures and tasks show that COMPOT consistently delivers a superior quality-compression trade-off over strong low-rank and sparse baselines, while remaining fully compatible with post-training quantization for extreme compression. Code is available $\href{https://github.com/mts-ai/COMPOT}{here}$.

Ammar Ali, Baher Mohammad, Denis Makhov et al.

We present ROCKET, a training-free model compression method that achieves state-of-the-art performance in comparison with factorization, structured-sparsification and dynamic compression baselines. Operating under a global compression budget, ROCKET comprises two key innovations: First, it formulates layer-wise compression allocation as a multi-choice knapsack problem, selecting the optimal compression level for each layer to minimize total reconstruction error while adhering to a target model size. Second, it introduces a single-step sparse matrix factorization inspired by dictionary learning: using only a small calibration set, it sparsifies weight coefficients based on activation-weights sensitivity and then updates the dictionary in closed form via least squares bypassing iterative optimization, sparse coding, or backpropagation entirely. ROCKET consistently outperforms existing compression approaches across different model architectures at 20-50\% compression rates. Notably, it retains over 90\% of the original model's performance at 30\% compression without any fine-tuning. Moreover, when applying a light fine-tuning phase, recovery is substantially enhanced: for instance, compressing Qwen3-14B to an 8B-parameter model and healing it with just 30 million tokens yields performance nearly on par with the original Qwen3-8B. The code for ROCKET is at github.com/mts-ai/ROCKET/tree/main.

Stamatios Lefkimmiatis, Iaroslav Koshelev

We introduce a novel optimization algorithm for image recovery under learned sparse and low-rank constraints, which we parameterize as weighted extensions of the $\ell_p^p$-vector and $\mathcal S_p^p$ Schatten-matrix quasi-norms for $0\!<p\!\le1$, respectively. Our proposed algorithm generalizes the Iteratively Reweighted Least Squares (IRLS) method, used for signal recovery under $\ell_1$ and nuclear-norm constrained minimization. Further, we interpret our overall minimization approach as a recurrent network that we then employ to deal with inverse low-level computer vision problems. Thanks to the convergence guarantees that our IRLS strategy offers, we are able to train the derived reconstruction networks using a memory-efficient implicit back-propagation scheme, which does not pose any restrictions on their effective depth. To assess our networks' performance, we compare them against other existing reconstruction methods on several inverse problems, namely image deblurring, super-resolution, demosaicking and sparse recovery. Our reconstruction results are shown to be very competitive and in many cases outperform those of existing unrolled networks, whose number of parameters is orders of magnitude higher than that of our learned models.

Iaroslav Koshelev, Stamatios Lefkimmiatis

In this work we present a novel optimization strategy for image reconstruction tasks under analysis-based image regularization, which promotes sparse and/or low-rank solutions in some learned transform domain. We parameterize such regularizers using potential functions that correspond to weighted extensions of the $\ell_p^p$-vector and $\mathcal{S}_p^p$ Schatten-matrix quasi-norms with $0 < p \le 1$. Our proposed minimization strategy extends the Iteratively Reweighted Least Squares (IRLS) method, typically used for synthesis-based $\ell_p$ and $\mathcal{S}_p$ norm and analysis-based $\ell_1$ and nuclear norm regularization. We prove that under mild conditions our minimization algorithm converges linearly to a stationary point, and we provide an upper bound for its convergence rate. Further, to select the parameters of the regularizers that deliver the best results for the problem at hand, we propose to learn them from training data by formulating the supervised learning process as a stochastic bilevel optimization problem. We show that thanks to the convergence guarantees of our proposed minimization strategy, such optimization can be successfully performed with a memory-efficient implicit back-propagation scheme. We implement our learned IRLS variants as recurrent networks and assess their performance on the challenging image reconstruction tasks of non-blind deblurring, super-resolution and demosaicking. The comparisons against other existing learned reconstruction approaches demonstrate that our overall method is very competitive and in many cases outperforms existing unrolled networks, whose number of parameters is orders of magnitude higher than in our case.

Dmitriy Shopkhoev, Ammar Ali, Magauiya Zhussip et al.

We introduce ReplaceMe, a generalized training-free depth pruning method that effectively replaces transformer blocks with a linear operation, while maintaining high performance for low compression ratios. In contrast to conventional pruning approaches that require additional training or fine-tuning, our approach requires only a small calibration dataset that is used to estimate a linear transformation, which approximates the pruned blocks. The estimated linear mapping can be seamlessly merged with the remaining transformer blocks, eliminating the need for any additional network parameters. Our experiments show that ReplaceMe consistently outperforms other training-free approaches and remains highly competitive with state-of-the-art pruning methods that involve extensive retraining/fine-tuning and architectural modifications. Applied to several large language models (LLMs), ReplaceMe achieves up to 25% pruning while retaining approximately 90% of the original model's performance on open benchmarks - without any training or healing steps, resulting in minimal computational overhead (see Fig.1). We provide an open-source library implementing ReplaceMe alongside several state-of-the-art depth pruning techniques, available at https://github.com/mts-ai/ReplaceMe.

Filippos Kokkinos, Stamatios Lefkimmiatis

Modern digital cameras rely on the sequential execution of separate image processing steps to produce realistic images. The first two steps are usually related to denoising and demosaicking where the former aims to reduce noise from the sensor and the latter converts a series of light intensity readings to color images. Modern approaches try to jointly solve these problems, i.e. joint denoising-demosaicking which is an inherently ill-posed problem given that two-thirds of the intensity information is missing and the rest are perturbed by noise. While there are several machine learning systems that have been recently introduced to solve this problem, the majority of them relies on generic network architectures which do not explicitly take into account the physical image model. In this work we propose a novel algorithm which is inspired by powerful classical image regularization methods, large-scale optimization, and deep learning techniques. Consequently, our derived iterative optimization algorithm, which involves a trainable denoising network, has a transparent and clear interpretation compared to other black-box data driven approaches. Our extensive experimentation line demonstrates that our proposed method outperforms any previous approaches for both noisy and noise-free data across many different datasets. This improvement in reconstruction quality is attributed to the rigorous derivation of an iterative solution and the principled way we design our denoising network architecture, which as a result requires fewer trainable parameters than the current state-of-the-art solution and furthermore can be efficiently trained by using a significantly smaller number of training data than existing deep demosaicking networks. Code and results can be found at https://github.com/cig-skoltech/deep_demosaick

Magauiya Zhussip, Dmitriy Shopkhoev, Ammar Ali et al.

Large language models (LLMs) have revolutionized AI applications, yet their high computational and memory demands hinder their widespread deployment. Existing compression techniques focus on intra-block optimizations (e.g. low-rank approximation, attention head pruning), while the repetitive layered structure of transformers implies significant inter-block redundancy - a dimension largely unexplored beyond key-value (KV) caching. Inspired by dictionary learning in CNNs, we propose a framework for structured weight sharing across transformer layers. Our approach decomposes attention projection matrices into shared dictionary atoms, reducing the attention module's parameters by 66.7% while achieving on-par performance. Unlike complex methods requiring distillation or architectural changes, MASA (Matrix Atom Sharing in Attention) operates as a drop-in replacement - trained with standard optimizers - and represents each layer's weights as linear combinations of shared matrix atoms. Experiments across scales (100M-700M parameters) show that MASA achieves better benchmark accuracy and perplexity than grouped-query attention (GQA), low-rank baselines and recently proposed Repeat-all-over/Sequential sharing at comparable parameter budgets. Ablation studies confirm robustness to the dictionary size and the efficacy of shared representations in capturing cross-layer statistical regularities. Extending to Vision Transformers (ViT), MASA matches performance metrics on image classification and detection tasks with 66.7% fewer attention parameters. By combining dictionary learning strategies with transformer efficiency, MASA offers a scalable blueprint for parameter-efficient models without sacrificing performance. Finally, we investigate the possibility of employing MASA on pretrained LLMs to reduce their number of parameters without experiencing any significant drop in their performance.

Baher Mohammad, Magauiya Zhussip, Stamatios Lefkimmiatis

We introduce MAVE (Mamba with Cross-Attention for Voice Editing and Synthesis), a novel autoregressive architecture for text-conditioned voice editing and high-fidelity text-to-speech (TTS) synthesis, built on a cross-attentive Mamba backbone. MAVE achieves state-of-the-art performance in speech editing and very competitive results in zero-shot TTS, while not being explicitly trained on the latter task, outperforming leading autoregressive and diffusion models on diverse, real-world audio. By integrating Mamba for efficient audio sequence modeling with cross-attention for precise text-acoustic alignment, MAVE enables context-aware voice editing with exceptional naturalness and speaker consistency. In pairwise human evaluations on a random 40-sample subset of the RealEdit benchmark (400 judgments), 57.2% of listeners rated MAVE - edited speech as perceptually equal to the original, while 24.8% prefered the original and 18.0% MAVE - demonstrating that in the majority of cases edits are indistinguishable from the source. MAVE compares favorably with VoiceCraft and FluentSpeech both on pairwise comparisons and standalone mean opinion score (MOS) evaluations. For zero-shot TTS, MAVE exceeds VoiceCraft in both speaker similarity and naturalness, without requiring multiple inference runs or post-processing. Remarkably, these quality gains come with a significantly lower memory cost and approximately the same latency: MAVE requires ~6x less memory than VoiceCraft during inference on utterances from the RealEdit database (mean duration: 6.21s, A100, FP16, batch size 1). Our results demonstrate that MAVE establishes a new standard for flexible, high-fidelity voice editing and synthesis through the synergistic integration of structured state-space modeling and cross-modal attention.

Dmitriy Shopkhoev, Denis Makhov, Magauiya Zhussip et al.

Post-training compression of large language models (LLMs) largely relies on low-rank weight approximation, which represents each column of a weight matrix in a shared low-dimensional subspace. While this is a computationally efficient strategy, the imposed structural constraint is rigid and can lead to a noticeable model accuracy drop. In this work, we propose CoSpaDi (Compression via Sparse Dictionary Learning), a novel training-free compression framework that replaces low-rank decomposition with a more flexible structured sparse factorization in which each weight matrix is represented with a dense dictionary and a column-sparse coefficient matrix. This formulation enables a union-of-subspaces representation: different columns of the original weight matrix are approximated in distinct subspaces spanned by adaptively selected dictionary atoms, offering greater expressiveness than a single invariant basis. Crucially, CoSpaDi leverages a small calibration dataset to optimize the factorization such that the output activations of compressed projection layers closely match those of the original ones, thereby minimizing functional reconstruction error rather than mere weight approximation. This data-aware strategy preserves better model fidelity without any fine-tuning under reasonable compression ratios. Moreover, the resulting structured sparsity allows efficient sparse-dense matrix multiplication and is compatible with post-training quantization for further memory and latency gains. We evaluate CoSpaDi across multiple Llama and Qwen models under per-layer and per-group settings at 20-50\% compression ratios, demonstrating consistent superiority over state-of-the-art data-aware low-rank methods both in accuracy and perplexity. Our results establish structured sparse dictionary learning as a powerful alternative to conventional low-rank approaches for efficient LLM deployment.

Savva Ignatyev, Daniil Selikhanovych, Oleg Voynov et al.

We present a novel method for 3D surface reconstruction from multiple images where only a part of the object of interest is captured. Our approach builds on two recent developments: surface reconstruction using neural radiance fields for the reconstruction of the visible parts of the surface, and guidance of pre-trained 2D diffusion models in the form of Score Distillation Sampling (SDS) to complete the shape in unobserved regions in a plausible manner. We introduce three components. First, we suggest employing normal maps as a pure geometric representation for SDS instead of color renderings which are entangled with the appearance information. Second, we introduce the freezing of the SDS noise during training which results in more coherent gradients and better convergence. Third, we propose Multi-View SDS as a way to condition the generation of the non-observable part of the surface without fine-tuning or making changes to the underlying 2D Stable Diffusion model. We evaluate our approach on the BlendedMVS dataset demonstrating significant qualitative and quantitative improvements over competing methods.

Iaroslav Koshelev, Daniil Selikhanovych, Stamatios Lefkimmiatis

In this work, we study the problem of non-blind image deconvolution and propose a novel recurrent network architecture that leads to very competitive restoration results of high image quality. Motivated by the computational efficiency and robustness of existing large scale linear solvers, we manage to express the solution to this problem as the solution of a series of adaptive non-negative least-squares problems. This gives rise to our proposed Recurrent Least Squares Deconvolution Network (RLSDN) architecture, which consists of an implicit layer that imposes a linear constraint between its input and output. By design, our network manages to serve two important purposes simultaneously. The first is that it implicitly models an effective image prior that can adequately characterize the set of natural images, while the second is that it recovers the corresponding maximum a posteriori (MAP) estimate. Experiments on publicly available datasets, comparing recent state-of-the-art methods, show that our proposed RLSDN approach achieves the best reported performance both for grayscale and color images for all tested scenarios. Furthermore, we introduce a novel training strategy that can be adopted by any network architecture that involves the solution of linear systems as part of its pipeline. Our strategy eliminates completely the need to unroll the iterations required by the linear solver and, thus, it reduces significantly the memory footprint during training. Consequently, this enables the training of deeper network architectures which can further improve the reconstruction results.

Valeriya Pronina, Filippos Kokkinos, Dmitry V. Dylov et al.

Microscopy is a powerful visualization tool in biology, enabling the study of cells, tissues, and the fundamental biological processes; yet, the observed images typically suffer from blur and background noise. In this work, we propose a unifying framework of algorithms for Gaussian image deblurring and denoising. These algorithms are based on deep learning techniques for the design of learnable regularizers integrated into the Wiener-Kolmogorov filter. Our extensive experimentation line showcases that the proposed approach achieves a superior quality of image reconstruction and surpasses the solutions that rely either on deep learning or on optimization schemes alone. Augmented with the variance stabilizing transformation, the proposed reconstruction pipeline can also be successfully applied to the problem of Poisson image deblurring, surpassing the state-of-the-art methods. Moreover, several variants of the proposed framework demonstrate competitive performance at low computational complexity, which is of high importance for real-time imaging applications.

Filippos Kokkinos, Stamatios Lefkimmiatis

Modern inexpensive imaging sensors suffer from inherent hardware constraints which often result in captured images of poor quality. Among the most common ways to deal with such limitations is to rely on burst photography, which nowadays acts as the backbone of all modern smartphone imaging applications. In this work, we focus on the fact that every frame of a burst sequence can be accurately described by a forward (physical) model. This in turn allows us to restore a single image of higher quality from a sequence of low quality images as the solution of an optimization problem. Inspired by an extension of the gradient descent method that can handle non-smooth functions, namely the proximal gradient descent, and modern deep learning techniques, we propose a convolutional iterative network with a transparent architecture. Our network, uses a burst of low quality image frames and is able to produce an output of higher image quality recovering fine details which are not distinguishable in any of the original burst frames. We focus both on the burst photography pipeline as a whole, i.e. burst demosaicking and denoising, as well as on the traditional Gaussian denoising task. The developed method demonstrates consistent state-of-the art performance across the two tasks and as opposed to other recent deep learning approaches does not have any inherent restrictions either to the number of frames or their ordering. Code can be found at https://fkokkinos.github.io/deep_burst/

Filippos Kokkinos, Stamatios Lefkimmiatis

Demosaicking and denoising are among the most crucial steps of modern digital camera pipelines and their joint treatment is a highly ill-posed inverse problem where at-least two-thirds of the information are missing and the rest are corrupted by noise. This poses a great challenge in obtaining meaningful reconstructions and a special care for the efficient treatment of the problem is required. While there are several machine learning approaches that have been recently introduced to deal with joint image demosaicking-denoising, in this work we propose a novel deep learning architecture which is inspired by powerful classical image regularization methods and large-scale convex optimization techniques. Consequently, our derived network is more transparent and has a clear interpretation compared to alternative competitive deep learning approaches. Our extensive experiments demonstrate that our network outperforms any previous approaches on both noisy and noise-free data. This improvement in reconstruction quality is attributed to the principled way we design our network architecture, which also requires fewer trainable parameters than the current state-of-the-art deep network solution. Finally, we show that our network has the ability to generalize well even when it is trained on small datasets, while keeping the overall number of trainable parameters low.

Stamatios Lefkimmiatis

We design a novel network architecture for learning discriminative image models that are employed to efficiently tackle the problem of grayscale and color image denoising. Based on the proposed architecture, we introduce two different variants. The first network involves convolutional layers as a core component, while the second one relies instead on non-local filtering layers and thus it is able to exploit the inherent non-local self-similarity property of natural images. As opposed to most of the existing deep network approaches, which require the training of a specific model for each considered noise level, the proposed models are able to handle a wide range of noise levels using a single set of learned parameters, while they are very robust when the noise degrading the latent image does not match the statistics of the noise used during training. The latter argument is supported by results that we report on publicly available images corrupted by unknown noise and which we compare against solutions obtained by competing methods. At the same time the introduced networks achieve excellent results under additive white Gaussian noise (AWGN), which are comparable to those of the current state-of-the-art network, while they depend on a more shallow architecture with the number of trained parameters being one order of magnitude smaller. These properties make the proposed networks ideal candidates to serve as sub-solvers on restoration methods that deal with general inverse imaging problems such as deblurring, demosaicking, superresolution, etc.

Stamatios Lefkimmiatis

We propose a novel deep network architecture for grayscale and color image denoising that is based on a non-local image model. Our motivation for the overall design of the proposed network stems from variational methods that exploit the inherent non-local self-similarity property of natural images. We build on this concept and introduce deep networks that perform non-local processing and at the same time they significantly benefit from discriminative learning. Experiments on the Berkeley segmentation dataset, comparing several state-of-the-art methods, show that the proposed non-local models achieve the best reported denoising performance both for grayscale and color images for all the tested noise levels. It is also worth noting that this increase in performance comes at no extra cost on the capacity of the network compared to existing alternative deep network architectures. In addition, we highlight a direct link of the proposed non-local models to convolutional neural networks. This connection is of significant importance since it allows our models to take full advantage of the latest advances on GPU computing in deep learning and makes them amenable to efficient implementations through their inherent parallelism.

Stamatios Lefkimmiatis, John Paul Ward, Michael Unser

We introduce a novel family of invariant, convex, and non-quadratic functionals that we employ to derive regularized solutions of ill-posed linear inverse imaging problems. The proposed regularizers involve the Schatten norms of the Hessian matrix, computed at every pixel of the image. They can be viewed as second-order extensions of the popular total-variation (TV) semi-norm since they satisfy the same invariance properties. Meanwhile, by taking advantage of second-order derivatives, they avoid the staircase effect, a common artifact of TV-based reconstructions, and perform well for a wide range of applications. To solve the corresponding optimization problems, we propose an algorithm that is based on a primal-dual formulation. A fundamental ingredient of this algorithm is the projection of matrices onto Schatten norm balls of arbitrary radius. This operation is performed efficiently based on a direct link we provide between vector projections onto $\ell_q$ norm balls and matrix projections onto Schatten norm balls. Finally, we demonstrate the effectiveness of the proposed methods through experimental results on several inverse imaging problems with real and simulated data.