Research & Development

Solving complex challenges in entirely new ways

Whether you are an innovative startup, an enterprise-level world-changer, or anything in between, the power of reaching for the unknown is at your disposal.

We support you in creating world-changing solutions from scratch, where new discoveries and scientific breakthroughs are required. With world-class research and development services, disrupting your market was never that easy.

Start innovating

Latest research

Unifying Deep Stochastic Processes for Image Enhancement

Unifies diffusion models, Ornstein–Uhlenbeck processes, and diffusion bridges under a single SDE framework
Consolidates 11 existing methods into one structured mathematical taxonomy
Benchmarks all methods across four tasks using identical backbones for fair comparison
Ships as ItoVision, an open-source modular PyTorch library

Possible applications

image super-resolution, colorization, low-light enhancement, deraining

Read the paper

Unifying Perspectives: Plausible Counterfactual Explanations on Global, Group-wise, and Local Levels

First gradient-based method unifying local, global, and group-wise counterfactual explanations in a single framework
Plausibility built directly into the optimization – no post-hoc filtering required
Applicable to any differentiable model without architecture-specific constraints
Covers all three granularity levels previously addressed only by separate, incompatible methods

Possible applications

explainable AI for credit scoring, healthcare, hiring, and regulatory audits

Read the paper

UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning

Hypernetwork generates LoRA weights on-the-fly from CLIP embeddings, replacing static per-concept fine-tuning
Trains in ~3 hours vs. 24+ hours for comparable methods
Generalizes zero-shot to unseen synonyms without per-concept retraining
Supports simultaneous erasure of 100+ concepts across Stable Diffusion and Flux architectures

Possible applications

content safety in text-to-image systems, copyright and identity removal, regulatory compliance

Read the paper

Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers

Novel approach to continual learning using Auxiliary Classifiers (ACs)
Improves accuracy by ~10% without changing core training methods
Enables early-exit inference, reducing compute by 10–60%
Compatible with multiple architectures (ResNet, VGG19, ViT, WideResNet)

Possible applications

Scalable continual learning in vision tasks, efficient AI systems for resource-constrained environments

Read the paper

Hypernetworks for Image Recontextualization

Novel approach to using AI for recontextualizing images
The technique requires no fine-tuning to work properly
The system delivers more accurate adjustments
Based on Low-Rank Adaptation and Hypernetworks

Possible applications

E-commerce, fashion, and apparel industries to reduce the cost of photomanipulation

Read the paper

Divide and not forget: Ensemble of selectively trained experts in Continual Learning

Overcoming the catastrophic forgetting in continual learning
A novel approach to overcoming catastrophic forgetting
The technique leverages the Mixture of Experts to make training the neural network easier
The technique injects new skills into the network while keeping its capabilities untouched

Possible applications

Reducing the time and cost required to upskill an existing neural network

Read the paper

Modeling Uncertainty in Personalized Emotion Prediction with Normalizing Flows

Detecting hate speech with more accuracy
Enhancing sentiment analysis with context detection
The model is capable of building the user profile from its messages and later taking it into account when analyzing the next messages.
The approach significantly reduces false positives and false negatives when detecting hate speech on digital platforms.

Possible applications

Hate speech detection
Sentiment analysis

Read the paper

HyperShot: Few-Shot Learning by Kernel HyperNetworks

A novel approach for few-shot problems that combines the benefits of using hypernetworks and kernel methods
The role of the hypernetwork is to produce weights dedicated to the considered task
The model can be applied to image classification problems where only a few examples are labeled
The model works well both for in- and cross-domain problems

Possible applications

Working in industries with a scarcity of labeled data, for example healthcare and life sciences

Read the paper

Diffused Heads: Diffusion Models Beat GANs on Talking-Face Generation

Our solution is capable of hallucinating head movements, facial expressions, such as blinks, and preserving a given background.

Possible applications

Media
Entertainment
Education

Read the paper

PluGeN: Multi-Label Conditional Generation From Pre-Trained Models

A novel approach for attribute manipulation in images, but not just this – we also modify proteins
The model works as a plug-in – we can link it with large pretrained generative models
The model is tiny and can be easily trained on a single GPU, but it is able to cooperate with large generative models, like StyleGan

Possible applications

Delivering precisely defined images by manipulating parameters for the sake of advertising, art or product development

Read the paper

Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations

A novel approach to eliminate the need for manual annotations to train computer-aided diagnosis tools in digital pathology
Application of mixed Deep Learning techniques with histopathological data
Automatic analysis of clinical data including imagery and text

Possible applications

Boosting the AI-powered diagnostics development process

Read the paper

FlowHMM: Flow-based continuous hidden Markov models

A novel approach for modeling sequential data
Hidden Markow Models enriched with normalizing flows to model high non-Gaussian distributions for different states
Two techniques for training the model that depend on the data’s nature were proposed
The model can be applied to various sequence modeling tasks, including speaker separation, activity recognition, part-of-speech tagging, and many others

Possible applications

Delivering better predictions for time series data
Modeling risk in the energy sector, utilities
Predicting stock prices

Read the paper

On Analyzing Generative and Denoising Capabilities of Diffusion-based Deep Generative Models

A heavy modification of the existing generative model that enrich the diffusion paradigm
The impact of noise in diffusion models is examined using various experimental settings
Gaining more insights into how diffusion models work and at which point they switch from generating structure with noise, to pure denoisingA novel approach for modeling sequential data

Possible applications

Boosting generative AI capabilities for design or art

Read the paper

Continual Learning with Guarantees via Weight Interval Constraints

A new training paradigm for continual learning
The approach assumes the hyperrectangle weight allocation of upcoming tasks
The proposed method overcomes the catastrophic forgetting problem while training the model in an incremental setting

Possible applications

Boosting overall ML model performance
Reducing the time and cost of training and maintaining ML models

Read the paper

Multiband VAE: Latent Space Alignment for Knowledge Consolidation in Continual Learning

A new method for unsupervised generative continual learning through the realignment of Variational Autoencoder’s latent space
Controlled forgetting of new samples allows the reduction of the reconstruction error in generative models
Multiband VAE is the first method to show forward and backward knowledge transfer in generative continual learning

Possible applications

General model performance improvement

Read the paper

CoNeRF: Controllable Neural Radiance Fields

We extend neural 3D representations to allow for intuitive and interpretable user control beyond novel view rendering (i.e. camera control)
We show a first time novel view and novel attribute re-rendering of scenes from a single video

Possible applications

Entertainment
Movie production
Advertising business

Read the paper

Two-headed eye-segmentation approach for biometric identification

We delivered architecture tailored for eye segmentation, done purely for person-identification purposes
Quality was ensured by evaluating segmentation to learn the informative segmentation needed for identification

Possible applications

Security
Person-identification

Read the paper

General hypernetwork framework for creating 3D point clouds

A new generative model for 3D point clouds that introduces the concept of HyperFlows
Set of practical methods that enable the generation of high-quality meshes using only point cloud data during training
The framework may be applied to generate complex 3D objects

Possible applications

Industrial design
3D processing

Read the paper

Non-Gaussian Gaussian Processes for Few-Shot Regression

Model that Predicts using only a few examples for a given task
Compared to most works, we are focused on regression problems (continuous outputs)
Since giving precise output values based on few labeled observations is not a good approach, we postulate to also model the uncertainty of the output (probabilistic regression)
The model can be used in many applications, including head pose trajectory modeling, power consumption prediction, or EEG signal analysis

Possible applications

Demand forecasting in retail
Stock price changes
Accurate estimations of power consumption

Read the paper

UCSG-Net–Unsupervised Discovering of Constructive Solid Geometry Tree

The model takes 3D shapes and decomposes them into simple geometrical shapes called primitives
The model also delivers a roadmap (Constructive Solid Geometry Tree) that tells us how to construct a complex shape using the primitives
The model is trained in an unsupervised manner, which practically means that it observes only 3D shapes without any information about the true primitives

Possible applications

Automated manipulation of scanned 3D objects
Preprojection of parts in a larger 3D system
Reducing super-complex structures (for example anatomy) into easier, engineerable parts for further analysis
The model can tell you how to deconstruct furniture to deliver it to a client

Read the paper

Hypernetwork approach to generating point clouds

A generative model for 3D point clouds
Using hypernetworks to generate 3D objects represented by various numbers of point clouds
The approach can learn to create meshes through seeing only point cloud data during training
The proposed approach may be used as a component for complex 3D scene generation

Possible applications

3D modeling
Reducing the computational cost of 3D and LIDAR data processing

Read the paper

Adversarial Autoencoders for Compact Representations of 3D Point Clouds

Framework for complex 3D point cloud processing
The set of models for complex 3D modeling includes point cloud: generation, representation, retrieval, clustering
The model can learn compact binary representations for point clouds

Possible applications

Working with 3D point cloud data
Autonomous vehicles
Industrial scanning

Read the paper

Conditional invertible flow for point cloud generation

A novel approach for generating 3D objects represented by point cloud
We can quickly create point clouds with various densities and details
The model is based on discrete normalizing flows and can be trained for a variety of objects and shapes

Possible applications

The method may be helpful in 3D design, virtual reality, and related topics
Transferring 3D point clouds
Storage optimization
AI-powered processing of point-cloud shapes

Read the paper

Bingan: Learning compact binary descriptors with a regularized GAN

A model for representing images using a minimal number of bits
We can search for similar images in a database more efficiently
The model is trained entirely in an unsupervised manner (no labels needed) and has high retrieval accuracy
We make use of the discriminator of the regularized GAN to obtain an informative binary representation

Possible applications

Reducing the carbon footprint of image-reliant operations
Optimizing the performance of image-heavy workflows

Read the paper

Quantitative spatial analysis of hematopoiesis-regulating stromal cells in the bone marrow microenvironment through 3D microscopy

A novel approach to quantifying microscopy data in the immune system
3D segmentation of bone marrow tissue samples after applying tissue-clearing techniques
Application of spatial statistics to characterize the organization of the biological system

Possible applications

Reducing the carbon footprint of image-reliant operations
Optimizing the performance of image-heavy workflows

Read the paper

Long-term tracking of budding yeast cells in brightfield microscopy: CellStar and the Evaluation Platform

A novel algorithm to segment yeast cells based on raycasting
Benchmark for evaluating SOTA methods
Automatic Evaluation Platform for scoring cell based segmentation assays

Possible applications

Diagnostics
Research in biology, pharmacology, chemistry, etc.
Microscopic observations

Read the paper

Our grants

Examode–H2020

Duration

2019–2023

Industry

Healthcare

Project budget

4M EUR

Tooploox allocation

600K EUR

Consortium leader

University of Applied Sciences, Bern, Switzerland

Learn more

Our team

With professors, PhDs and PhD candidates we will get support you in breaking the rules and reaching for what was seen as unreachable before. We have a whole scientific institute ready to challenge reality for you.

Tomasz Trzciński, Prof. Chief Scientist

LinkedIn Google Scholar

Ivona Tautkute-Rustecka, PhD Senior AI Engineer

LinkedIn Google Scholar

Maciej Zięba, Prof. Head of R&D

LinkedIn Google Scholar

Rafał Topolnicki, PhD Senior ML Engineer

Rafał Nowak, PhD ML/AI Researcher

LinkedIn Google Scholar

Paweł Lorek, Prof. Data Scientist

Google Scholar

Joanna Kaczmar-Michalska, PhD candidate Senior ML Engineer & Researcher

LinkedIn Google Scholar

Adam Słucki, PhD candidate ML Engineer

Mateusz Kobos, PhD ML Engineer

LinkedIn Google Scholar

Mirek Korzeniowski, PhD Data Scientist

LinkedIn Google Scholar

Szymon Stoma, PhD Co-founder Tooploox Research

LinkedIn Google Scholar

Let our team rethink your toughest challenges

Research

Solve problems once deemed unsolvable by letting our team take a crack at your challenges. Our minds will deliver solutions that will shatter the status quo in your industry and make you a shaper of tomorrow.

Prototyping and POC

We deliver proof that our ideas and concepts are real and applicable. Get ready to see your impossible idea ready and working.

Feasibility study

Sometimes it is better to let an idea go. But when? With our feasibility study, you will get the answers to your business questions and mitigate risks well before they arise.

MVP

Let the market verify your vision. Our research and development team, supported by software engineers and designers, will prepare you a Minimum Viable Product so you can enter the market and change the world right away!

Key milestones in Solvd research department

2019

Microscope IT represents Polish tech at CERN

Microscope IT, acquired by Tooploox, now part of the Solvd team, joined a delegation of Polish companies that presented their work and innovative, research-driven ideas to the international scientific community.

2019

Strategic partnership with Microscope IT

By joining forces with Microscope IT – a dev shop providing AI and R&D services for Life Science, Healthcare and Medtech, Tooploox strengthened its capabilities in applied research and advanced engineering, laying the foundation for a dedicated research team.

2025

Tooploox joins Solvd

Years of research and development practices were unified into a single research department, with the belief that this merger would not only make us stronger but also help us to fulfill our mission to improve lives through cutting-edge products and solutions.

Solving complex challenges in entirely new ways

Latest research

Unifying Deep Stochastic Processes for Image Enhancement

Unifying Perspectives: Plausible Counterfactual Explanations on Global, Group-wise, and Local Levels

UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning

Our grants

Examode–H2020

Our team

Let our team rethink your toughest challenges

Research

Prototyping and POC

Feasibility study

MVP

Key milestones in Solvd research department

Ready to take your business to the next level?

Solving complex challenges in entirely new ways