Tag
作者测试了 Embedding Gemma 2（Q8 约 330MB、768 维）在视觉小说截图上的文搜图能力，发现其对跨语言文本含义有一定理解且表现可用，但文本内容往往主导余弦相似度、视觉特征影响较弱，且需要精细的措辞。
MMSearch-OPD 提出了一种结合可控数据精炼与联合强化学习和在线策略蒸馏的方法，以提升长时程多模态搜索智能体的能力，通过将简单问题扩展为困难且可学习的训练样本取得显著改进。
Google promotes how developers can combine EmbeddingGemma 2 with Google AI Edge to build on-device apps that search through photos, videos, and audio using natural language descriptions.
Feiteng Li 分享了一份针对 EmbeddingGemma2 的开源资源清单，涵盖 DigUp、Recall 等本地多模态搜索应用、决策模型微调方案，以及 llama.cpp、MLX、LiteRT、Transformers.js 等多端部署路径。相关仓库已开源并欢迎社区贡献。
SCM (Screen Memories) is a local-first macOS app that adds AI-powered search over every photo and video frame in any folder, with scene-level video search, OCR, Whisper-based dialogue search, and opt-in local LLM chat — all running offline on-device. It is open-source, installable via Homebrew, and built with Bun, Electron, and a vision model (CLIP by default).