---
license: cc-by-nc-4.0
tags: [diffusion, dit, synthetic-data, generalization]
---
# Visual Generative Lab (VGL) checkpoints

Checkpoints behind Table 2 of *Do Diffusion Models Learn to Generalize Basic Visual Skills?*
Code, datasets and evaluation: https://github.com/AmishSethi/visual-generative-lab

Layout: `table2/<skill>/<variant>/seed_<n>/final.pt` with the run's `run_config.json` beside it.
Skills: size, position, rotation, count. Variants: baseline, sinusoidal, rotary, adaln, vae, flow, unet, dit_large
(rotation also has `baseline_nz`, the ten-seed zero-null baseline). Each `.pt` holds `model`, `ema`, `opt` and the
training state; load with the training scripts in the repository. License CC BY-NC 4.0 (inherited from DiT).