namithap/teconerv-models

TeCoNeRV Model Checkpoints

TeCoNeRV uses hypernetworks to predict implicit neural representation (INR) weights for video compression. A patch-tubelet decomposition enables hypernetworks to scale to high-resolution video prediction. The temporal coherence objective reduces redundancy across consecutive clips, enabling compact residual encoding of per-clip parameters.

This repository contains hypernetwork training checkpoints for the three model families described in the paper.

Model families

nervenc — Baseline NeRVEnc hypernetwork. Predicts full-resolution clip reconstructions directly.

patch_tubelet — Proposed patch-tubelet hypernetwork. Predicts parameters for spatial tubelets; full frames are reconstructed by tiling. Supports resolution-independent inference.

teconerv — Proposed method. Initialized from a patch_tubelet checkpoint and finetuned with a temporal coherence objective.

Getting started

See the GitHub repository for full documentation on setup, training, and evaluation. Checkpoint download instructions are in docs/models.md.

git lfs install
git clone https://huggingface.co/namithap/teconerv-models

Citation

@article{padmanabhan2026teconerv,
  title={TeCoNeRV: Leveraging Temporal Coherence for Compressible Neural Representations for Videos},
  author={Padmanabhan, Namitha and Gwilliam, Matthew and Shrivastava, Abhinav},
  journal={arXiv preprint arXiv:2602.16711},
  year={2026}
}