330MB EmbeddingGemma2 indexes 6,000+ screenshots: text beats visual descriptors in cosine similarity
jupiterbjy · reddit · 2026-10-19
The author tested EmbeddingGemma2 (330MB at Q8, 768-dim) via llama.cpp as a text-search-over-images engine on 6,000+ visual novel screenshots (English/Korean/Japanese, 22,448 total indexed).
Key findings:
- Very usable for its size; indexing at 5.9 imgs/sec on an RX 7900GRE via Vulkan (ROCm avoided)
- Sensitive to wording (blond vs blonde)
- Text dominates cosine similarity: dialog text scores 0.7596 vs 0.6821 for "blonde girl"; adding visual traits can even hurt similarity (chair in background: −0.00225)
- Single rare keywords outweigh full sentences — a word appearing in only 2/3003 images is the strongest discriminator
Setup: Q80 quantization, KV/batch/ctx all 4096 (embedding model restriction). Author admits the test set isn't scientific but shares full configs, timings and query tables.
More from Infra
- Bindu Reddy: NeoClouds Like CoreWeave May Be Better Off Declaring Bankruptcy — bindureddy · 2026-10-19
- Qwen3.5-0.8B 4-bit spotted running on Rabbit R1 via wllama — ngxson · 2026-10-19
- ORL trains diffusion LMs on their own trajectories, boosting a 4B model's throughput by 6.9x past speculative decoding — burny_tech · 2026-10-19
- 4 free CUDA resources to learn GPU programming from scratch, from NVIDIA to Oxford — techNmak · 2026-10-19
- Running Moss image model on a 4GB RTX 3050, optimized down to 62s per image — sanketshinde2004 · 2026-10-19
- Barclays: pure-play data center bond spreads hit ~209bps as market prices in contagion — SumitGup · 2026-10-19