🔥Telegram İfşa

An impressive paper that explores the balance between generalization and memorization in diffusion models. The authors argue that diffusion models do not globally underfit their training objectives but instead exhibit selective underfitting: they fit the empirical score function

really cool work!

Work with Yuki if you get the opportunity, he's a great researcher and a wonderful person :)

reward hacking 101

been thinking about recurrence for stupidly long video understanding. will need the right incentives and arch tweaks, but if done correctly would be beautiful

Why add REPA when you can be explicit and use the VLM representation to generate? 🤔 We found the semantic encoder already has the right priors. Train it to sample in its native latent space + lightweight pixel decoder = unified vision model. But naively using the semantic

Introducing nanosft, a clean single file implementation of finetuning for chat style model. Loads gpt2-124M weights on nanogpt and does supervised finetuning using just pytorch. a side project that I made recently for some prep. link below :) qts/rts appericiated

really good thread on how Sinkhorn-Knopp ended up powering current SSL methods

motivation is a scam, discipline is the way. don't be a slave to your moods. don't be a slave to your environment. be disciplined, be free.

Excited to introduce DiffuseNNX, a comprehensive JAX/Flax NNX-based library for diffusion and flow matching! It supports multiple diffusion / flow-matching frameworks, Autoencoders, DiT variants, and sampling algorithms. Repo: github.com/willisma/diffu… Delve into details below!

Introducing Representation Autoencoders (RAE)! We revisit the latent space of Diffusion Transformers, replacing VAE with RAE: pretrained representation encoders (DINOv2, SigLIP2) paired with trained ViT decoders. (1/n)

!!!

Rohan Choudhury I think the point is that we can use encoders like dino, siglip and that semantic representations arent as lossy as we had thought initially - cnn, vit doesn’t really matter

"humanity's last last exam. i swear, this is the last one, trust me bro"

🚀 New preprint! We present NP-Edit, a framework for training an image editing diffusion model without paired supervision. We use differentiable feedback from Vision-Language Models (VLMs) combined with distribution-matching loss (DMD) to learn editing directly. webpage:

It's the "model" in model-based RL

Chris Offner Alexandre Morgand That's a curve ball question but here's my intuition/hypothesis: 1. Locality of Task: Depth estimation driven by priors can be primarily thought of as a local task, i.e., given semantic context of things in the image you can predict the relative depth - hence why also linear

someone took 3 classes of intro to prob. cool stuff

really cool work on entropy-based tokenization!!

there are cathedrals everywhere for those with the eyes to see