SAM3 vs Astra head-to-head: Astra nails player roles at 42/45 where SAM3 fails
TheMoonMidas · x · 2026-09-11
A hands-on comparison of SAM3 and Astra on soccer match frames shows the two are close on simple noun queries ("referee": 16/20 for SAM3 vs 20/20 for Astra), but Astra pulls far ahead on contextual understanding:
- For "the players involved in this attack", SAM3 returns every player, while Astra picks 4 of 16 with accurate roles: on the ball, passing option, closing down.
- For "the player with the ball at his feet", SAM3 returns all 18 players; Astra returns the right one, scoring 42/45.
- SAM3 can't handle "number 10", while Astra reads jersey numbers.
The author concludes Astra is remarkably good at understanding image context and plans to build further projects on top of it.
More from Models
- Persimmon unveiled: first large-scale model to simulate human conversation — niloofar_mire · 2026-09-11
- DeepSeek V4 training details: RL beyond collapse, WSD schedule, 1M context over 10T tokens — stochasticchasm · 2026-09-11
- Neel Nanda replicates Astra system card: no-CoT reasoning jumps 1.75x over next-best models — NeelNanda5 · 2026-09-11
- OpenAI pauses Pro subscriptions due to overwhelming demand — Charuru · 2026-09-11
- Artificial Analysis isn't broken: self-funded benchmarks, $13k spent on one model — Antblue · 2026-09-11
- Claude checkpoint diagnostician jokes Opus 3 is 'the cure' after newer model quirks — repligate · 2026-09-11