About this role
Summer Semester 2027 – fixed term of 3–6 months How can generative models create synthetic data that is not only realistic, but also reliable and useful for training machine learning models? In this thesis, you will investigate how simulation and generative models can be combined to produce controllable training data for a specific downstream task. The main challenge is to increase visual diversity and reduce the visual gap between synthetic and real images without compromising scene structure, annotation validity, or model performance.
Simulation provides accurate annotations and precise control over scene composition, while generative models can increase visual diversity and reduce the visual gap between synthetic and real images. However, visually realistic images are not necessarily useful for model training. Generative transformations may alter object boundaries, visibility, geometry, or scene content, making the original annotations unreliable. In addition, standard image quality metrics do not necessarily translate into better model performance on real-world data.
The objective of this thesis is to investigate how generative pipelines for synthetic data can be controlled and evaluated with respect to a specific downstream task. Potential inputs for guiding generation include simulation metadata, visual concepts, text prompts, reference images, segmentation masks, edge maps, depth maps, or geometric constraints.
YOUR TASKS
- Investigate selected approaches for steering generative image synthesis while preserving task-relevant scene structure and annotation validity.
- Evaluate generated data in terms of visual quality, annotation consistency, and task-specific performance on real-world data.
- Identify which control mechanisms and generation strategies improve the synthetic data pipeline, critically analyze their limitations, and implement a validated prototype.
- The expected outcome is a validated methodology and prototype for controllable, task-aware synthetic data generation.
- A critical assessment of the underlying assumptions, limitations, and potential failure cases will be an essential part of the thesis.
YOUR PROFILE
- You are currently pursuing an advanced master's degree in Computer Science, Artificial Intelligence, Data Science, Mathematics, Robotics, Physics, Engineering, or a related field.
- You have a good understanding of machine learning and deep learning, as well as practical experience with Python and computer vision.
- Familiarity with PyTorch, diffusion models, simulation, or synthetic data is an advantage.
- You are interested in reviewing scientific literature, designing controlled experiments, and critically evaluating results beyond individual quantitative metrics.
- You can quickly grasp complex technical challenges, work independently and systematically, and communicate your findings clearly within an interdisciplinary team.
- You have a strong interest in generative AI, synthetic data, sim-to-real transfer, and applied industrial research.
- You have good written and spoken English skills.
You don't meet every requirement yet? We still encourage you to apply. We value curiosity, a strong willingness to learn, and the ability to approach complex problems systematically. You can also develop relevant skills during your thesis. Contact: Sarah Disch