VLM3D MICCAI 2026 Challenge Win 🎉
Our team, FakeBiomedImage, won 1st place in MR report-to-volume generation and 2nd place in CT report-to-volume generation at the Forithmus VLM3D Challenge. I will present the winning solution at MICCAI 2026.
AI PhD Researcher @ IDEA Lab, FAU Erlangen-Nuremberg | Generative models & Foundation models
About me
Hello! My name is Teodora Spasojevic, and I am a PhD Researcher at IDEA Lab, FAU Erlangen-Nürnberg, supervised by Prof. Bernhard Kainz. I research generative and foundation models for medical imaging as part of the BAIOSPHERE Foundation Model Project and the Borderless Research Group at FAU and Imperial College London.
I completed my Master’s degree at Friedrich-Alexander University Erlangen–Nuremberg with majors in Artificial Intelligence and Autonomous Driving. Throughout my studies, I focused strongly on research and worked primarily under the supervision of Prof. Dr. Vasileios Belagiannis. Our collaboration included a research-oriented project thesis on uncertainty estimation for generative models, as well as my final master’s thesis, on improving temporal consistency of video generation, which was conducted mainly externally.
Along with my studies, I completed an internship and master's thesis in Mercedes-Benz Autonomous Driving R&D, where I focused on diffusion-based video generation and world-models for autonomous systems. In my master’s thesis conducted with the team, I proposed a simple, yet effective fine-tuning framework for improving temporal consistency of Stable Video Diffusion. This work led to our scientific publication, accepted at IEEE Intelligent Vehicles (IV) 2026.
In addition, I served as a Research Assistant in an Information Theory project, under the supervision of Prof. Dr. Laura Cottatellucci.
News
New paper accepted at IEEE IIV 2026 🎉
We propose Localized Semantic Alignment (LSA), a simple yet effective framework for fine-tuning pre-trained video generation models. LSA enhances temporal consistency by aligning semantic features between ground-truth and generated video clips.
Experience
PhD Researcher
IDEA Lab FAU Erlangen-Nuremberg
My research is focused on generative and foundation models as part of the BAIOSHPERE project, the first large-scale Bavarian multimodal healthcare foundation model, in collaboration with the Technical University of Munich (TUM) and Ludwig-Maximilians-Universität München (LMU). The project aims to advance scalable, trustworthy AI systems for medical data across imaging, clinical, and multimodal representations.
Master's Thesis Student
Mercedes-Benz AG, Autonomous Driving R&D Team
Developed a master’s thesis on improving temporal consistency in video generation through an end-to-end multimodal fine-tuning framework for pretrained models (Stable Video Diffusion). The method enforces dynamic, semantic, and depth consistency of the generation via set of specialized loss functions that make use of DN-DETR, DINO, SAM, and DepthAnything. Designed distributed training and evaluation pipelines on GCP with Flyte and Accelerate support.
Additionally, contributed as a first author to a scientific paper accepted at the IEEE IV 2026 conference, by extending the thesis.
Intern
Mercedes-Benz AG, Autonomous Driving R&D Team
Conducted research on diffusion-based video generative models, world models (WMs), and conditioning mechanisms (e.g., ControlNet), contributing to the formulation of a master’s thesis proposal. Experimented with large-scale models including Stable Video Diffusion (SVD), Vista WM, Cosmos WM, and Ctrl-V (SVD with ControlNet).
Utilized GCP infrastructure with Flyte and Weights & Biases for experiment management, and developed a GCP workflow setup tutorial which is now used by the team.
Research Assistant
Friedrich-Alexander University Erlangen-Nuremberg
Implemented a new initialization method for Bilinear Expectation Propagation (EP) algorithm, part of a Cell-Free 6G Information Theory Project. New initialization led to an increase in accuracy of the EP algorithm. Experiments were conducted on a high-performance computing (HPC) cluster.
Software Engineer Intern
TTTech Auto
Completed my bachelor thesis on the topic of Real-time Traffic Light Detection in the CARLA Simulator. The work included usage of signal processing tools with, at the time, SOTA object detection model (YOLO) with real-time capabilities, and integrating it into CARLA simulator.
Undergraduate Teaching Assistant
University of Belgrade, School of Electrical Engineering
Worked as an undergraduate teaching assistant on courses in Signals and Systems and Automatic Control Systems, where I assisted with laboratory education and assessed students’ homework and lab performance.
Education
M.Sc.
Friedrich-Alexander University Erlangen-Nuremberg
Area: Artificial Intelligence & Autonomous Driving
GPA: 1.4
Thesis: Consistency-Aware Video Generation: Dynamic, Semantic, and Geometric Constraints for Improved Temporal Coherence
Scientific Project: Uncertainty Estimation of Video Generation Models
Minor Projects: Vision Transformer, LSTM, RNN implementation; Imitation Learning
Courses: Computer Vision, Advanced Deep Learning, Reinforcement Learning
B.Sc.
University of Belgrade, School of Electrical Engineering
Area: Electrical Engineering & Computing
GPA: 9.33/10.00 (equivalent to 1.5 in German grading system)
Thesis: Real-time Traffic Light Detection in the CARLA Simulator
Activities: Undergraduate Teaching Assistant for Signals and Systems and Automatic Control Systems
Courses: AI, ML, Computer Science, Mathematics, Signal Processing
Publications
LSA: Localized Semantic Alignment for Enhancing Temporal Consistency in Traffic Video Generation
Localized Semantic Alignment (LSA) improves temporal consistency in diffusion-based traffic video generation within a single fine-tuning epoch and with no inference-time control signals.
IEEE Intelligent Vehicles (IV) 2026
Awards & Scholarships
- — BAYHOST Scholarship
- — Scholarship for Exceptionally Gifted Students — Ministry of Education, Science and Technological Development of Republic of Serbia
Contact
Please feel free to get in touch! You can email me at spasojevic.teodora@gmail.com, DM me on LinkedIn or use the contact form below. 😊