Seminar: Open Generative AI Systems: LLMs, Video Models, and World Models

Administrative information

Seminar course for Bachelor students (IN0014) and Master students (IN2107).

  • Organizers: Tobias Lasser, David Frank
  • Sessions: the course takes place online.
    • Block 1+2: LLM architectures, diffusion and flow models, and efficient inference
      • Wednesday, October 7, 2026, 13:30 - 17:30
      • Thursday, October 8, 2026, 13:30 - 17:30
    • Block 3+4: world models, JEPA, physical AI, and open generative AI systems
      • Wednesday, December 9, 2026, 13:30 - 17:30
      • Thursday, December 10, 2026, 13:30 - 17:30
  • Course language: English

Pre-course meeting

A pre-course meeting will take place on Thursday, July 2, 2026, at 11:00 online (BBB room).

Registration

  • Registration takes place via the IN.TUM Matching System from July 10 to 15, 2026.
  • Additionally, please fill out the registration form until July 15, 2026.
  • If you do not fill out the registration form, you will not be matched to the course!

The seminar is open to all students of Informatics, Mathematics, Physics and related fields.
The number of students is limited to 8.

Course overview

Generative AI is rapidly moving beyond single model families. Current frontier systems involve large language models, diffusion and flow-based image and video generation, and increasingly world models that aim to predict and simulate interactive environments. In this seminar, we will study these developments and efficient inference techniques through recent arXiv papers, technical reports, and open-source model releases.

The seminar is organized around four themes. First, we examine modern LLM architectures, including autoregressive transformers, attention variants, mixture-of-experts models, speculative decoding, quantization, and serving systems. Second, we study diffusion and flow-based generative models for image and video, with a focus on recent work on efficient video generation, self-forcing-style methods, causal generation, and inference optimization. Third, we investigate world models, including interactive video models, JEPA-style latent predictive architectures, omnimodal physical-AI models, and world-action models for planning and control. Finally, we connect these topics through open-source reproducibility, evaluation, and hands-on system design.

Each participant will give two in-depth presentations in 45-minute slots, structured as a 30-minute talk followed by 15 minutes of discussion. The presentations will be based on recent scientific literature and should connect technical mechanisms to concrete open models, repositories, or systems that can be inspected or tried out. Interactive sessions and audience engagement are highly encouraged to foster a collaborative learning environment. To conclude the seminar, each student will submit a concise 3-page written summary covering the key ideas, papers, and discussions from all of the presentations. The goal is to create an individual reference that captures both the technical content and the open questions around current generative AI systems.

Aims of the course

This seminar aims to give students a solid understanding of current open generative AI systems, from LLMs to diffusion models to video generation and world models. Students will learn how modern model architectures are designed, how inference is accelerated, how open-source releases can be evaluated, and how generative models are beginning to support simulation, interaction, and physical AI. The course fosters critical reading of recent papers, technical presentation skills, and collaborative discussion. Students are encouraged to analyze not only model quality, but also reproducibility, efficiency, architectural tradeoffs, and practical usability.

Prerequisites

Participants should have a basic understanding of deep learning concepts. Familiarity with transformers, diffusion models, or generative AI is helpful but not required. A strong interest in current AI research and enthusiasm for the interactive, presentation-based seminar format are highly recommended.