arXiv subjects

Kamil Deja

Publications and source records attributed to Kamil Deja.

At least 19 recordsLinked to original sources

arXiv2026

Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models

Text-to-image diffusion models exhibit remarkable generative capabilities, yet their internal operations remain opaque, particularly when handling prompts that are not fully descriptive. In such scenarios, models must make implicit decisions to generate details not explicitly specified in the text. This work investigates the hypothesis that this decision-making process is not diffuse but is computationally localized within the model's architecture. While existing localization techniques focus on prompt-related interventions, we notice that such explicit conditioning may differ from implicit decisions. Therefore, we introduce a probing-based localization technique to identify the layers with the highest attribute separability for concepts. Our findings indicate that the resolution of ambiguous concepts is governed principally by self-attention layers, identifying them as the most effective point for intervention. Based on this discovery, we propose ICM (Implicit Choice-Modification) - a precise steering method that applies targeted interventions to a small subset of layers. Extensive experiments confirm that intervening on these specific self-attention layers yields superior debiasing performance compared to existing state-of-the-art methods, minimizing artifacts common to less precise approaches. The code is available at https://github.com/kzaleskaa/icm.

cs.CV↗

arXiv2026

TADA! Tuning Audio Diffusion Models through Activation Steering

Audio diffusion models can synthesize high-fidelity music from text, yet achieving fine-grained control over specific musical attributes remains challenging, as their internal mechanisms for representing high-level concepts are poorly understood. In this work, we use activation patching to demonstrate that recent audio diffusion architectures exhibit a semantic bottleneck, where a small, shared subset of consecutive attention layers controls distinct musical concepts, such as the presence of specific instruments, vocals, or genres. Building on this, we systematically evaluate a broad spectrum of steering paradigms, comparing activation steering against prompt-level, score-space, and weight-space interventions, analyzing the interaction between the steering mechanism and the intervention site. Our new benchmark, supported by an extensive user study, demonstrates that localized activation steering establishes a new state-of-the-art in audio concept modulation.

cs.SD↗

arXiv2026

ELROND: Exploring and decomposing intrinsic capabilities of diffusion models

A single text prompt passed to a diffusion model often yields a wide range of visual outputs determined solely by stochastic process, leaving users with no direct control over which specific semantic variations appear in the image. While existing unsupervised methods attempt to analyze these variations via output features, they omit the underlying generative process. In this work, we propose a framework to disentangle these semantic directions directly within the input embedding space. To that end, we collect a set of gradients obtained by backpropagating the differences between stochastic realizations of a fixed prompt that we later decompose into meaningful steering directions with either Principal Components Analysis or Sparse Autoencoder. Our approach yields three key contributions: (1) it isolates interpretable, steerable directions for precise, fine-grained control over a single concept; (2) it effectively mitigates mode collapse in distilled models by reintroducing lost diversity; and (3) it establishes a novel estimator for concept complexity under a specific model, based on the dimensionality of the discovered subspace.

cs.LG↗

arXiv2025

Membership and Dataset Inference Attacks on Large Audio Generative Models

Generative audio models, based on diffusion and autoregressive architectures, have advanced rapidly in both quality and expressiveness. This progress, however, raises pressing copyright concerns, as such models are often trained on vast corpora of artistic and commercial works. A central question is whether one can reliably verify if an artist's material was included in training, thereby providing a means for copyright holders to protect their content. In this work, we investigate the feasibility of such verification through membership inference attacks (MIA) on open-source generative audio models, which attempt to determine whether a specific audio sample was part of the training set. Our empirical results show that membership inference alone is of limited effectiveness at scale, as the per-sample membership signal is weak for models trained on large and diverse datasets. However, artists and media owners typically hold collections of works rather than isolated samples. Building on prior work in text and vision domains, in this work we focus on dataset inference (DI), which aggregates diverse membership evidence across multiple samples. We find that DI is successful in the audio domain, offering a more practical mechanism for assessing whether an artist's works contributed to model training. Our results suggest DI as a promising direction for copyright protection and dataset accountability in the era of large audio generative models.

cs.LG↗

arXiv2025

FoldSAE: Learning to Steer Protein Folding Through Sparse Representations

RFdiffusion is a popular and well-established model for generation of protein structures. However, this generative process offers limited insight into its internal representations and how they contribute to the final protein structure. Concurrently, recent work in mechanistic interpretability has successfully used Sparse Autoencoders (SAEs) to discover interpretable features within neural networks. We combine these concepts by applying SAE to the internal representations of RFdiffusion to uncover secondary structure-specific features and establish a relationship between them and generated protein structures. Building on these insights, we introduce a novel steering mechanism that enables precise control of secondary structure formation through a tunable hyperparameter, while simultaneously revealing interpretable block and neuron-level representations within RFdiffusion. Our work pioneers a new framework for making RFdiffusion more interpretable, demonstrating how understanding internal features can be directly translated into precise control over the protein design process.

q-bio.QM↗

arXiv2025

ExpertSim: Fast Particle Detector Simulation Using Mixture-of-Generative-Experts

Simulating detector responses is a crucial part of understanding the inner workings of particle collisions in the Large Hadron Collider at CERN. Such simulations are currently performed with statistical Monte Carlo methods, which are computationally expensive and put a significant strain on CERN's computational grid. Therefore, recent proposals advocate for generative machine learning methods to enable more efficient simulations. However, the distribution of the data varies significantly across the simulations, which is hard to capture with out-of-the-box methods. In this study, we present ExpertSim - a deep learning simulation approach tailored for the Zero Degree Calorimeter in the ALICE experiment. Our method utilizes a Mixture-of-Generative-Experts architecture, where each expert specializes in simulating a different subset of the data. This allows for a more precise and efficient generation process, as each expert focuses on a specific aspect of the calorimeter response. ExpertSim not only improves accuracy, but also provides a significant speedup compared to the traditional Monte-Carlo methods, offering a promising solution for high-efficiency detector simulations in particle physics experiments at CERN. We make the code available at https://github.com/patrick-bedkowski/expertsim-mix-of-generative-experts.

cs.CV↗

arXiv2025

Precise Parameter Localization for Textual Generation in Diffusion Models

Novel diffusion models can synthesize photo-realistic images with integrated high-quality text. Surprisingly, we demonstrate through attention activation patching that only less than $1$% of diffusion models' parameters, all contained in attention layers, influence the generation of textual content within the images. Building on this observation, we improve textual generation efficiency and performance by targeting cross and joint attention layers of diffusion models. We introduce several applications that benefit from localizing the layers responsible for textual content generation. We first show that a LoRA-based fine-tuning solely of the localized layers enhances, even more, the general text-generation capabilities of large diffusion models while preserving the quality and diversity of the diffusion models' generations. Then, we demonstrate how we can use the localized layers to edit textual content in generated images. Finally, we extend this idea to the practical use case of preventing the generation of toxic text in a cost-free manner. In contrast to prior work, our localization approach is broadly applicable across various diffusion model architectures, including U-Net (e.g., SDXL and DeepFloyd IF) and transformer-based (e.g., Stable Diffusion 3), utilizing diverse text encoders (e.g., from CLIP to the large language models like T5). Project page available at https://t2i-text-loc.github.io/.

cs.CV↗

arXiv2025

SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders

Diffusion models, while powerful, can inadvertently generate harmful or undesirable content, raising significant ethical and safety concerns. Recent machine unlearning approaches offer potential solutions but often lack transparency, making it difficult to understand the changes they introduce to the base model. In this work, we introduce SAeUron, a novel method leveraging features learned by sparse autoencoders (SAEs) to remove unwanted concepts in text-to-image diffusion models. First, we demonstrate that SAEs, trained in an unsupervised manner on activations from multiple denoising timesteps of the diffusion model, capture sparse and interpretable features corresponding to specific concepts. Building on this, we propose a feature selection method that enables precise interventions on model activations to block targeted content while preserving overall performance. Our evaluation shows that SAeUron outperforms existing approaches on the UnlearnCanvas benchmark for concepts and style unlearning, and effectively eliminates nudity when evaluated with I2P. Moreover, we show that with a single SAE, we can remove multiple concepts simultaneously and that in contrast to other methods, SAeUron mitigates the possibility of generating unwanted content under adversarial attack. Code and checkpoints are available at https://github.com/cywinski/SAeUron.

cs.LG↗

arXiv2024

Joint Diffusion models in Continual Learning

In this work, we introduce JDCL - a new method for continual learning with generative rehearsal based on joint diffusion models. Neural networks suffer from catastrophic forgetting defined as abrupt loss in the model's performance when retrained with additional data coming from a different distribution. Generative-replay-based continual learning methods try to mitigate this issue by retraining a model with a combination of new and rehearsal data sampled from a generative model. In this work, we propose to extend this idea by combining a continually trained classifier with a diffusion-based generative model into a single - jointly optimized neural network. We show that such shared parametrization, combined with the knowledge distillation technique allows for stable adaptation to new tasks without catastrophic forgetting. We evaluate our approach on several benchmarks, where it outperforms recent state-of-the-art generative replay techniques. Additionally, we extend our method to the semi-supervised continual learning setup, where it outperforms competing buffer-based replay techniques, and evaluate, in a self-supervised manner, the quality of trained representations.

cs.LG↗

arXiv2024

Mediffusion: Joint Diffusion for Self-Explainable Semi-Supervised Classification and Medical Image Generation

We introduce Mediffusion -- a new method for semi-supervised learning with explainable classification based on a joint diffusion model. The medical imaging domain faces unique challenges due to scarce data labelling -- insufficient for standard training, and critical nature of the applications that require high performance, confidence, and explainability of the models. In this work, we propose to tackle those challenges with a single model that combines standard classification with a diffusion-based generative task in a single shared parametrisation. By sharing representations, our model effectively learns from both labeled and unlabeled data while at the same time providing accurate explanations through counterfactual examples. In our experiments, we show that our Mediffusion achieves results comparable to recent semi-supervised methods while providing more reliable and precise explanations.

cs.CV↗

arXiv2024

There and Back Again: On the relation between Noise and Image Inversions in Diffusion Models

Diffusion Models achieve state-of-the-art performance in generating new samples but lack a low-dimensional latent space that encodes the data into editable features. Inversion-based methods address this by reversing the denoising trajectory, transferring images to their approximated starting noise. In this work, we thoroughly analyze this procedure and focus on the relation between the initial noise, the generated samples, and their corresponding latent encodings obtained through the DDIM inversion. First, we show that latents exhibit structural patterns in the form of less diverse noise predicted for smooth image areas (e.g., plain sky). Through a series of analyses, we trace this issue to the first inversion steps, which fail to provide accurate and diverse noise. Consequently, the DDIM inversion space is notably less manipulative than the original noise. We show that prior inversion methods do not fully resolve this issue, but our simple fix, where we replace the first DDIM Inversion steps with a forward diffusion process, successfully decorrelates latent encodings and enables higher quality editions and interpolations. The code is available at https://github.com/luk-st/taba.

cs.CV↗

arXiv2024

Low-Rank Continual Personalization of Diffusion Models

Recent personalization methods for diffusion models, such as Dreambooth and LoRA, allow fine-tuning pre-trained models to generate new concepts. However, applying these techniques across consecutive tasks in order to include, e.g., new objects or styles, leads to a forgetting of previous knowledge due to mutual interference between their adapters. In this work, we tackle the problem of continual customization under a rigorous regime with no access to past tasks' adapters. In such a scenario, we investigate how different adapters' initialization and merging methods can improve the quality of the final model. To that end, we evaluate the naive continual fine-tuning of customized models and compare this approach with three methods for consecutive adapters' training: sequentially merging new adapters, merging orthogonally initialized adapters, and updating only relevant task-specific weights. In our experiments, we show that the proposed techniques mitigate forgetting when compared to the naive approach. In our studies, we show different traits of selected techniques and their effect on the plasticity and stability of the continually adapted model. Repository with the code is available at https://github.com/luk-st/continual-lora.

cs.LG↗

arXiv2024

Particle identification with machine learning from incomplete data in the ALICE experiment

The ALICE experiment at the LHC measures properties of the strongly interacting matter formed in ultrarelativistic heavy-ion collisions. Such studies require accurate particle identification (PID). ALICE provides PID information via several detectors for particles with momentum from about 100 MeV/c up to 20 GeV/c. Traditionally, particles are selected with rectangular cuts. A much better performance can be achieved with machine learning (ML) methods. Our solution uses multiple neural networks (NN) serving as binary classifiers. Moreover, we extended our particle classifier with Feature Set Embedding and attention in order to train on data with incomplete samples. We also present the integration of the ML project with the ALICE analysis software, and we discuss domain adaptation, the ML technique needed to transfer the knowledge between simulated and real experimental data.

hep-ex↗

arXiv2024

Machine-learning-based particle identification with missing data

In this work, we introduce a novel method for Particle Identification (PID) within the scope of the ALICE experiment at the Large Hadron Collider at CERN. Identifying products of ultrarelativisitc collisions delivered by the LHC is one of the crucial objectives of ALICE. Typically employed PID methods rely on hand-crafted selections, which compare experimental data to theoretical simulations. To improve the performance of the baseline methods, novel approaches use machine learning models that learn the proper assignment in a classification task. However, because of the various detection techniques used by different subdetectors, as well as the limited detector efficiency and acceptance, produced particles do not always yield signals in all of the ALICE components. This results in data with missing values. Machine learning techniques cannot be trained with such examples, so a significant part of the data is skipped during training. In this work, we propose the first method for PID that can be trained with all of the available data examples, including incomplete ones. Our approach improves the PID purity and efficiency of the selected sample for all investigated particle species.

physics.ins-det↗

arXiv2024

Generative Diffusion Models for Fast Simulations of Particle Collisions at CERN

In High Energy Physics simulations play a crucial role in unraveling the complexities of particle collision experiments within CERN's Large Hadron Collider. Machine learning simulation methods have garnered attention as promising alternatives to traditional approaches. While existing methods mainly employ Variational Autoencoders (VAEs) or Generative Adversarial Networks (GANs), recent advancements highlight the efficacy of diffusion models as state-of-the-art generative machine learning methods. We present the first simulation for Zero Degree Calorimeter (ZDC) at the ALICE experiment based on diffusion models, achieving the highest fidelity compared to existing baselines. We perform an analysis of trade-offs between generation times and the simulation quality. The results indicate a significant potential of latent diffusion model due to its rapid generation time.

physics.data-an↗

arXiv2024

Deep Generative Models for Proton Zero Degree Calorimeter Simulations in ALICE, CERN

Simulating detector responses is a crucial part of understanding the inner-workings of particle collisions in the Large Hadron Collider at CERN. The current reliance on statistical Monte-Carlo simulations strains CERN's computational grid, underscoring the urgency for more efficient alternatives. Addressing these challenges, recent proposals advocate for generative machine learning methods. In this study, we present an innovative deep learning simulation approach tailored for the proton Zero Degree Calorimeter in the ALICE experiment. Leveraging a Generative Adversarial Network model with Selective Diversity Increase loss, we directly simulate calorimeter responses. To enhance its capabilities in modeling a broad range of calorimeter response intensities, we expand the SDI-GAN architecture with additional regularization. Moreover, to improve the spatial fidelity of the generated data, we introduce an auxiliary regressor network. Our method offers a significant speedup when comparing to the traditional Monte-Carlo based approaches.

cs.LG↗

arXiv2024

GUIDE: Guidance-based Incremental Learning with Diffusion Models

We introduce GUIDE, a novel continual learning approach that directs diffusion models to rehearse samples at risk of being forgotten. Existing generative strategies combat catastrophic forgetting by randomly sampling rehearsal examples from a generative model. Such an approach contradicts buffer-based approaches where sampling strategy plays an important role. We propose to bridge this gap by incorporating classifier guidance into the diffusion process to produce rehearsal examples specifically targeting information forgotten by a continuously trained model. This approach enables the generation of samples from preceding task distributions, which are more likely to be misclassified in the context of recently encountered classes. Our experimental results show that GUIDE significantly reduces catastrophic forgetting, outperforming conventional random sampling approaches and surpassing recent state-of-the-art methods in continual learning with generative replay.

cs.LG↗

arXiv2024

Particle physics DL-simulation with control over generated data properties

The research of innovative methods aimed at reducing costs and shortening the time needed for simulation, going beyond conventional approaches based on Monte Carlo methods, has been sparked by the development of collision simulations at the Large Hadron Collider at CERN. Deep learning generative methods including VAE, GANs and diffusion models have been used for this purpose. Although they are much faster and simpler than standard approaches, they do not always keep high fidelity of the simulated data. This work aims to mitigate this issue, by providing an alternative solution to currently employed algorithms by introducing the mechanism of control over the generated data properties. To achieve this, we extend the recently introduced CorrVAE, which enables user-defined parameter manipulation of the generated output. We adapt the model to the problem of particle physics simulation. The proposed solution achieved promising results, demonstrating control over the parameters of the generated output and constituting an alternative for simulating the ZDC calorimeter in the ALICE experiment at CERN.

cs.LG↗