Research map / Vision-language and multimodal / Multimodal LLMs
Scaling Language-Free Visual Representation Learning
David Fan, Shengbang Tong, Jiachen Zhu, Koustuv Sinha, Zhuang Liu, Xinlei Chen, Michael Rabbat, Nicolas Ballas, Yann LeCun, Amir Bar, Saining Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train DINOv2 on web data containing text (heavy filter, 1.3% of MC-2B) instead of full data (ablation) | Web-DINO ViT-2B · Cambrian-1 OCR & Chart | OCR & Chart VQA: 26.8 -> 40.4 | Improved |
| light text filter (50.3% of data) instead of full data (ablation) | Web-DINO ViT-2B · Cambrian-1 OCR & Chart | OCR & Chart VQA: 26.8 -> 33.2 | Improved |
| heavy text filter instead of full data (ablation) | Web-DINO ViT-2B · Cambrian-1 average VQA | average VQA: 50.8 -> 53.7 | Improved |
| heavy text filter instead of full data (ablation) | Web-DINO ViT-2B · Cambrian-1 General | General VQA: 72.8 -> 70.7 | Mixed |
| scale model size from 1B to 7B (ablation) | Web-DINO · Cambrian-1 VQA | OCR & Chart and Vision-Centric VQA | Improved |
| scale CLIP model size beyond 3B (ablation) | CLIP · Cambrian-1 VQA | VQA categories | No clear effect |
| increase examples seen from 1B to 8B (ablation) | Web-DINO ViT-7B · Cambrian-1 VQA | OCR & Chart VQA | Improved |
| increase examples seen beyond 2B to 4B (ablation) | Web-DINO ViT-7B · Cambrian-1 General, Knowledge, Vision-Centric | VQA | No clear effect |
| pretrain on ImageNet-1k instead of MC-2B (ablation) | Web-DINO 1B-3B · VQA and ImageNet-1k linear probe | VQA / linear probe | Hurt |
| use MAE instead of DINOv2 as SSL objective (ablation) | Web-MAE 1B-5B · Cambrian-1 VQA | VQA | Mixed |
| increase input resolution 224 to 518 via short fine-tuning (ablation) | Web-DINO ViT-7B · Cambrian-1 average VQA | average VQA: 55.2 -> 59.9 | Improved |
| train on web-scale MC-2B vs LVD-142M (DINOv2 ViT-g) | Web-DINO ViT-7B · Cambrian-1 average VQA | average VQA: 47.9 -> 55.2 | Improved |
Closest papers by meaning
- Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision (ICML 2021 · similarity 0.87)
- Unsupervised Visual Representation Learning by Graph-Based Consistent Constraints (ECCV 2016 · similarity 0.85)
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · similarity 0.83)
- Sequential Modeling Enables Scalable Learning for Large Vision Models (CVPR 2024 · similarity 0.83)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.83)
- Language-Informed Visual Concept Learning (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
