Revisiting Feature Prediction for Learning Visual Representations from Video
Paper • 2404.08471 • Published • 1
V-JEPA 2 ViT-L Cell Cycle Classification
This repo is the planned home for the V-JEPA 2 ViT-L fine-tuned for cell cycle state prediction on CTC Fluo-N2DH-GOWT1. Fine-tuning has not yet run; only the model card lives here.
For the trained baseline (U-Net + BiLSTM), see
DnaRnaProteins/unet-bilstm-cell-cycle-baseline.
Plan when training lands
- Single-channel microscopy input (3-channel patch embedding weights summed to 1-channel)
- Encoder frozen for 10 warmup epochs, then last 4 of 24 ViT-L blocks unfrozen, lr 1e-5
- Classification head: temporal mean-pool + MLP over patch tokens, 3-way output
Base model reference
Bardes et al., "Revisiting Feature Prediction for Learning Visual Representations from Video." arXiv:2404.08471. V-JEPA 2 released at ICLR 2026.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support