DINOv3 embeddings survive 1-bit quantization for image search with re-ranking, tests show
mgostIH · x · 2026-11-04
The author found DINOv3's semantic embeddings good enough to spot what two images have in common — something even Opus 5.5 High missed at first glance and only got by analyzing the images' cluster; SigLIP 2 was too coarse semantically.
Practical tips: storing embedding vectors at 4-bit quantization is essentially lossless for search; even 1-bit or lower works well if you re-rank a larger retrieved set, at extra inference cost. Reference throughput: 78 img/s for DINOv3 on a 9070XT GPU.
More from Infra
- Unsloth Dynamic v3.1 GGUFs boost Qwen3.8-Flash-Next accuracy 10-15% over GSQ-RCO — danielhanchen · 2026-11-05
- Trained an MoE across two GPU clusters 25 Gbps apart — the compute islands guide — _arohan_ · 2026-11-05
- Arena data: Gemini 4 Carbon scales cheapest on hard tasks, rivals cost 3-4x more — arena · 2026-11-05
- Apple warns holiday-quarter gross margin will slide to 45% as memory chip prices surge — rohanpaul_ai · 2026-11-05
- AI infra is token-centric, but pixels, video and 3D deserve first-class status — prateekj · 2026-11-05
- 3x leveraged chip ETF SOXL hits $3.4B daily volume, topping most crypto perp markets — tomwanhh · 2026-11-05