bobthebuilderinternational/inr-dit-video

INR-DiT Video

A from-scratch generative video pipeline that diffuses a neural representation, not pixels.

Faces generated from pure noise — only a grid of neural-network codes was diffused.

Pipeline: noise → DiT (generates a latent grid) → INR decoder (renders patches) → video

1,000 steps	6,000 steps	12,000 steps

Structure emerging over training.

file	what it is
bet2_grid.pt	shared INR decoder + 337 latent grids (8×8×8 × 32-dim)
dit_012000.pt	diffusion transformer over the grid cells

Code & full write-up: https://github.com/BOB-THE-BUILDER-in/inr-dit-video Trained on HDTF (not redistributed): https://github.com/MRzzm/HDTF