Omni World Model

Evaluating a trained world model: evaluate_only workflow, sequential vs full_sequence sampling, PSNR/SSIM/LPIPS/FID(+FVD) definitions, and headline numbers on DINO-WM and RT-1 checkpoints.

Design spec for adding an optional Wan2.2-TI2V-5B video backbone to OmniWM: OmniWan model class, f0 token surgery, Wan VAE latent pipeline, flow-matching joint loss, prefill-cache policy sampling, and LoRA compute plan. All additive; NanoWM/Omni paths untouched.

Training a Nano World Model: workflow, Hydra config system, and the three ablated design axes (prediction target, action injection, model scale) with result tables and shipped checkpoints.