3 ms·
Looks great, but looking at the benchmark, can’t help but think about how crazy good dots-ocr is as a model. Too bad they’re not as open as the Deepseek team be
by bugglebeetle 1y ago
Looks great, but looking at the benchmark, can’t help but think about how crazy good dots-ocr is as a model. Too bad they’re not as open as the Deepseek team because its so crazy good and would love to know how it was trained.
- rfoo 1y agoIf you look you'd notice that it's the same Haoran Wei behind DeepSeek-OCR and GOT-OCR2.0 :p
- bugglebeetle 1y agoOh you’re right! Good catch!
- bethekind 1y agoDid we read the same graph? DeepSeek Gundam 200 dpi appeared to get similar perf as dots-ocr, but with less tokens needed. The x axis is inverted, descending with distance from the origin.