Generative AI (cont.) - Diffusion Transformer models (DiT)

AGENDA
This month's session is about Diffusion Transformer models (DiT).

Last time we built up the math framework of flow and diffusion models and used a U-Net inside it as a concrete implementation to generate images. This time we'll check if we can use transformers instead of the U-Net with convolutions. To understand the benefits of transformers over convolution in generative AI, we'll first discuss differences between pixel vs. latent space diffusion through an image generation example.

We'll start using transformers from the base architecture of vision transformer (ViT) models and walk through our way to the topic of how to apply them in a diffusion setting instead of the usual autoregressive setups.
Finally, we'll give a quick peek into multi-modal image generation - where we condition our model with complex textual prompts instead of a class label only.

The session will discuss the content mostly of the following papers: Scalable Diffusion Models with Transformers --- https://arxiv.org/pdf/2212.09748 Scaling Rectified Flow Transformers for High-Resolution Image Synthesis --- https://arxiv.org/pdf/2403.03206 High-Resolution Image Synthesis with Latent Diffusion Models --- https://arxiv.org/pdf/2112.10752

+1 thing: Watch out for the opening announcement of our Slack channel where you'll soon be able to recommend or vote future papers & topics. Details will be available on the meetup series' page: https://www.meetup.com/turbine-ai-meetups/

ENTRY DETAILS
Please look for and download the QR code attached this event page. The code will appear by the week of the event. The gates are quirky, so please increase your phone's screen brightness to maximum and hold it at a 5-10 cm distance from the scanner to help it focus. Please hold onto the codes as you will need them both in and out!

We will be waiting for you in the lobby to help.

PAST EVENTS
Materials from past events are listed here:
https://www.turbine.ai/ai-meetup-presentations