Flow Matching for Generative Modeling: https://buff.ly/3T4bb9W
Latent Consistency Models: https://buff.ly/3COzdmh

What’s your crazy idea for AI-generated images? 🚀 What other techniques or tools have you come across?

What's exciting is that every leap forward isn't just about improving efficiency. It's about knocking down barriers so that our collective imagination can go further. Blockade Labs allows developers to generate 360-degree environments and Decart OASIS - a fully-generated AI game is already here!

Latent Consistency Models (LCM): While FM rewrites the rules, LCM focuses on optimizing the existing technique. By “teaching” a smaller, faster model to mimic diffusion models, LCM drastically reduces the time needed for inference. We go from 100s of denoising step to 1-4.

Flow Matching (FM): A reimagination of how image generation could work. Instead of step-by-step denoising, FM replaces it with a generalized approach that’s faster, more flexible, and requires fewer steps to generate an image. It's a new ground-up approach to image generation.

But these challenges are exactly what motivate brilliant minds to find solutions - which is what I came across in my recent paper quest. Researchers have come up with ways to solve the problem of image generation speed often with completely different approaches:

Idea: What if you could have a real-time world generated as you navigate it from within your AR/VR glasses!
Roadblock: Stable Diffusion actually takes a few seconds to generate images as it has to go through multiple denoising steps (upto 100)

Here's a link to the SoM paper for those who are interested in reading - arxiv.org/pdf/2310.11441
If you're interested in signing up for my course, use this link to get a discount - uplimit.com/go/ai-produc...

8/ P.S. I would also highly recommend a trip to Albania. It’s stunning, underexplored, and full of hidden gems. 🌍

7/ It’s also exactly the kind of thing I teach in my course, Building AI Products. We explore practical techniques for text, images, and other modalities—and how to turn them into real-world applications.
If you're interested, our next batch starts in December!

6/ I found this pretty cool because visual prompting is a whole new dimension compared to text-based prompting techniques like role-play or chain-of-thought. It opens doors to exciting possibilities in AI-powered tools.

5/ SoM prompting is brilliant because it transforms how AI interprets visual data. I guess you are giving cues to the model about what to focus on, boosting accuracy dramatically. In the paper, the researchers suggested automating this with Segment Anything Model (SAM).

4/ In the second attempt, I added four small markers to the image—highlighting key spots like cars and mountains! GPT-4o nailed it.
Not only did it correctly identify the location as Albania, but it also looked at number plates and suggested the exact area I was heading to!

3/ First, I uploaded the photo as it was.
The AI’s response was vague and unhelpful. No clear idea where the picture was taken.

2/ I came across an interesting research paper that describes a technique called Set-of-Mark (SoM) prompting, designed for LMMs like GPT-4o.
SoM is a way to guide AI’s “vision” by adding small visual markers to an image. Curious, I decided to test it out.