Fine-tuned Qwen3-VL 8B reads camera shots at 69% accuracy, beating Opus and Gemini
ravisparikh · x · 2026-11-20
Developer jhh3 released Shot Reader, a fine-tuned Qwen3-VL 8B model that actually "watches" video to reverse-engineer shot construction: camera height, lens, angle, camera movement, plus who's in frame and where they move.
- Frontier models struggle here: even Opus 5.5 and Gemini 4 with purpose-built prompts get camera height and movement both right less than half the time; out-of-the-box open models only hit 5%.
- Shot Reader reaches 69%, runs 2.4x faster than Opus, costs 30x less, and is small enough to run on your own GPU.
- As ravisparikh notes, models like Fable and Astra only inspect frames — they can't watch video.
The pitch: AI can already write Blender scripts, but rebuilding a shot starts with reading it — see a shot, get the plan, then rebuild in 3D, reshoot, remix. Open source.
More from Multimodal
- Redditor Shares 'Entrega Express', a Short Film Made With Seedance 2 — Ok-Nerve941 · 2026-11-20
- HeyGen Releases Avatar 6, Shows Side-by-Side Realism Leap Since Avatar III — HeyGen · 2026-11-20
- User calls Google Genie's new upgrade impressive, diving back in ASAP — AI_Andrew · 2026-11-20
- 'Johnny' — stunning Johnny Depp-style AI portraits showcase top image generation quality — egeberkina · 2026-11-20
- MBZUAI's PAN 2 World Model: 29B Params, Fully Open with Weights, Inference and Training Code — ZhitingHu · 2026-11-20
- Gemini 4 Neon codes a Chamonix ski sim from scratch: ~10k lines, zero assets — flngr · 2026-11-20