Computer Vision Paris Meetup #27

### The Computer Vision Meetup Paris is back 🔥

For this 27th edition, the Meetup will be held on Tuesday, December, 12, 7p.m at 1 Rue Ambroise Thomas, 75009 Paris at Mobiskill’s premises.

As usual, we have 2 special guests, Galem Kayo, CEO of Automi AI and Antoine Sueur, Research Engineer chez Yokai.

🎤 Here are the talks of the evening (the talk are in French 🇫🇷 with English supports 🇬🇧) :

• Scaling computer vision in industrial environments with giga-models presented by Galem Kayo who will be remote

Computer vision solutions are increasingly sought after in the manufacturing sector, but their Return on Investment (ROI) often remains modest due to scalability challenges.
Industrial environments exhibit a high degree of variability, complicating the widespread application of computer vision. Traditional algorithms struggle in the face of this variability. While deep learning has improved adaptability and robustness, it comes with significant training costs.
This presentation explores the potential of giga-models combining vision and language modalities to enhance ROI in computer vision applications.
We will delve into giga-models and effective deployment strategies for scalable computer vision solutions in industrial environments. Emphasis will be placed on practical solutions that address the unique challenges of these contexts.

• Image generation models: origins, functionality, and limitations presented by Antoine Sueur

This presentation will focus on a thorough analysis of recent Text-to-image model architectures, aiming to detail their functionality and capabilities. It will begin the presentation with a brief recap of advancements in image generation over the past few years, from the emergence of GANs to the rise of diffusion models. It will then conduct a detailed examination of the StableDiffusion model's architecture, the most widely used open-source model for image generation. This analysis will cover the various constituent blocks of this model and their respective functions. It will also address the training methodology of this model, while identifying the limitations and biases that may arise from it. It will explore advanced techniques such as ControlNet, Cross-Image Attention, and RichText, aiming to enhance control over this model or even enable it to assimilate new concepts, akin to LORA and Dreambooth. Finally, it will present certain research avenues considered providing solutions to some inherent issues in current models.

If you wish to attend this conference, register now !
🚨 The number of tickets is limited to 50 🚨

We look forward to seeing you 👋