Aaypom/factorjepa-cosmos-future-adapter

FactorJEPA to Cosmos future-latent adapter

This repository contains the small trainable adapter that maps frozen FactorJEPA future tokens into NVIDIA Cosmos continuous video latents. It predicts two future latent slots; a separately encoded last-context-frame anchor is prepended only for causal 4k+1 decoding. It does not contain the upstream FactorJEPA or Cosmos weights.

• Base FactorJEPA checkpoint: anonymousML123/factorjepa-outputs/outputs/full/vjepa_2_1_vitg_1B/train/m09c_surgery_3stage_DI_diheavy_encoder/m09c_ckpt_best.pt
• Cosmos tokenizer: nvidia/Cosmos-0.1-Tokenizer-CV4x8x8
• Best/latest metric payload: metrics/best.json
• Training code: experiments/jepa_cosmos in the FactorJEPA repository.