Amit Thakur

🚀 @Microsoft has recetly released a rearchitectured version of @pyautogen v0.4 — the future of intelligent, modular multi-agent AI systems! To help you get hands-on, I’ve launched a new course: 🎓 Build Intelligent Multi-Agent Applications with AutoGen (v0.4)

Langevin dynamics in a nutshell: take the gradient of the output w.r.t. the input, then nudge the input slightly in the negative gradient direction to reduce the energy score. Small steps, guided by gradients, gradually move you toward lower-energy (more likely) states.

Energy based modela usage the concept which was originally formulated by Boltzmann in 1868. I love the fact that most of deep generative models use maths behind the scenes which were formulated long time ago, still very effective and it comes back again and again.

We are also releasing self-contained lecture notes that explain flow matching and diffusion models from scratch. This goes from "zero" to the state-of-the-art in modern Generative AI. 📖 Read the notes here: arxiv.org/abs/2506.02070 Joint work with @EErives40101.

High likelihood and great sample quality should be the aim.

🚀MIT Flow Matching and Diffusion Lecture 2026 Released (diffusion.csail.mit.edu)! We just released our new MIT 2026 course on flow matching and diffusion models! We teach the full stack of modern AI image, video, protein generators - theory and practice. We include: 📺 Videos: Step-by-step derivations. 📝 Notes: Mathematically self-contained lecture notes 💻 Coding: Hands-on exercises for every component We fully improved last years’ iteration and added new topics: latent spaces, diffusion transformers, building language models with discrete diffusion models. Everything is available here: diffusion.csail.mit.edu A huge thanks to Tommi Jaakkola for his support in making this class possible and Ashay Athalye (MIT SOUL) for the incredible production! Was fun to do this with @RShprints! #MachineLearning #GenerativeAI #MIT #DiffusionModels #AI

Starting my PhD in AI at BITS Pilani Dubai, researching diffusion models and flow matching for controllable generative AI.

Video lectures, MIT IAP 6.S183 A Practical Introduction to Diffusion Models winter 2026, by Chenyang Yuan, Cole Becker, Artem Lukoianov, Daniel Pfrommer, Christopher Scarvelis practical-diffusion.org/lectures/ .

@ssahoo_ @mbzuai You have any event here in Dubai?

Habibi, come to Dubai (or join the Silicon Valley Lab @mbzuai and they’ll fly you out every once in a while)

Which SDE book you found to be helpful while learning about the basics and use of stochastic differential equations in diffusion models and flow matching?

🚀 PhD Internship (will be hosting one person): Diffusion LLMs (DLLM) Looking for PhD students with: • hands-on DLLM research • First author published paper on DLLM (NeurIPS/ICLR/ICML/ACL/EMNLP) Send Google Scholar + a brief introduction → DM RT appreciated 🙏 #NLP #ML

Looks like a good resource to learn about stochastic processess . Will be helpful for diffusion model researchers. arxiv.org/abs/2501.04126

As promised, I have published a 3+ hours long new YouTube video where I explain the Vision Transformer (ViT) paper and implement it from scratch in PyTorch. youtu.be/t3L7DE_6GNE

I'm a diffusion model. I learn step by step.

This does NOT mean the model is acting as a classifier. It never predicts labels or probabilities. It only predicts noise. So the key idea is simple. CFG uses conditional and unconditional predictions, not classification.

These two predictions are combined. The difference between them pushes the image toward the prompt more strongly.

Classifier-free guidance often gets misunderstood. The name makes it sound confusing, so let’s clear it up.