VGI-Bench: Seedance Leads Visual Reasoning Benchmark for Video Models
青稞AI · wechat · 2026-09-01
To test if video models possess "visual reasoning via generation," researchers built VGI-Bench with 27 tasks and 810 samples. Results show models have emerging reasoning skills but lack stability.
Top 3: Seedance 2.0 (51.0), MiniMax-H3 (44.4), Kling 3.0 (44.0).
Key Findings:
- Visual appearance significantly impacts scores and rankings.
- Fine-tuning on large-scale synthetic data boosts performance by enhancing reasoning and reducing rule violations.
- Internal denoising traces show models form hypotheses early with minimal self-correction later.
- High scores in some closed-source models likely stem from external VLM prompt rewriting, not native ability. Open-source MiniMax-H3 nearing SOTA suggests scaling laws apply to video reasoning.
More from Multimodal
- Luma shares filmmaking-style workflow: motion-capture real stunts then restyle with video-to-video — lmoroney · 2026-10-18
- MOSS-Image-SFT at 4 Steps Draws Hype: 320x320 to 2K Without Artifacts — Powerful_Evening5495 · 2026-10-18
- MOSS SFT vs Flux 3 Klein vs SD 3.0: Same-Prompt Comparison Exposes SD3's Age — ZootAllures9111 · 2026-10-18
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Omni 1'1 flash video demo generates atmospheric foggy Victorian haunted house — Dense-Instruction210 · 2026-10-18
- Flux 3 klein distilled tested: int8 workflow delivers surprisingly good 2MP images — Ok_Tale7582 · 2026-10-18