V-JEPA-LM: 1B-8B video-language models with counterfactual dynamics supervision, 4B scores 72.1
_akhaliq · x · 2026-11-17
V-JEPA-LM (1B-8B) attaches a frozen V-JEPA 2.1 encoder to an LLM and trains with counterfactual supervision so the models actually read video dynamics — direction, speed, and order — rather than relying on single-frame captions.
- They lead every size class across 11 video benchmarks; the 4B model averages 72.1, beating all open models.
- Single training stage, 4.53M training samples, 32 visual tokens per frame.
- Open-sourced on Hugging Face.
Related event: V-JEPA-LM: Counterfactual Supervision Teaches LLMs to Understand Video(2 posts)→
More from Multimodal
- French creator builds an AI-made GTA-style Marseille while waiting for GTA 6 — Dimillian · 2026-11-17
- Savannah nocturne: a surreal AI-generated safari scene with cinematic lighting — cosmoxtraveller · 2026-11-17
- YouArt launches cheapest Seedance MCP at $0.0918/sec, selling below cost — nikola_mr64990 · 2026-11-17
- Could Minimax H3 serve as a latent upscaler for 20-64MP two-pass image pipelines? — Majestic-Ad-1336 · 2026-11-17
- Diffusers guide: run IDOL single-image-to-3D Gaussian splat with a few lines, no official integration needed — RisingSayak · 2026-11-17
- LichtFeld Studio adds collision support for walkable Gaussian Splat scenes — janusch_patas · 2026-11-17