Leonardo's Blog

A unified explanation of the four most-confused acronyms in deep learning compute: what they mean, how to count them, and how they show up in scaling laws, model cards, and hardware spec sheets.

Notes on how to use Apptainer (Singularity) for containerization on HPC clusters

Notes on .gitignore, pre-commit, and GitHub Actions checks

Rclone

Spatial Reasoning in VLMs

DiT, MMDiT, DiT-Air, UViT and PRX

On-Policy Distillation, ULD and GOLD

Random Thoughts

Linear RNNs and Attention

Bagel and LightBagel

Kaplan et al. and Chinchilla et al. Scaling Laws

Survey

Visual Geometry Grounded Transformer

The Intrinsic Dimension of Images and Its Impact on Learning

CUDA, Triton and flash attention

State of AI Report 2025

TOEFL

Spatial Reasoning

Representation Learning for Generation (Illustration)

Tokenizer training and inference code

Fat-tree

ColBERT and FILIP

Code practice for deep learning

A Sober Look at the Robustness of CLIPs to Spurious Features

Vision encoders should be image size agnostic and task driven

Unified Vision-Language Models

Huggingface Trainer

Pre-training is all about mode coverage, post-training is all about mode collapsing

Summary of Xiangyu Zhang's talk

GRPO

Principled & Automated Interpretability in Deep Learning

Robustness of VLM

Xet

Distribution Shifts

AliTok

WebDataset

MGVQ

KARL

Image tokenization

Accelerate

FlowTok

VA-VAE

CFM

Download ImageNet

Downsampling Regularization

FMM

Single-Step Generation via Self-Consistency

EDM

DiT and SiT

LDM

REPA

Stochastic Interpolants

Language Diffusion Model

Use t-SNE and UMAP for Visual Analytics properly

Approximating Language Model Training Data from Weights

Mean Flow Model

Research

FM

Inference-time Scaling

Thoughts

CNF

ODE and SDE

CoT

Generative Model for Vision

Diffusion Models

VAE

GAN

Flow-based Models, Energy-based Models

Self-Supervised Learning

KL, JSD, Wasserstein Distance, Fisher Divergence