Public Blind Test Crowns MiniMax H3 Over ByteDance Seedance 2.5 in Video Gen
葬AI · wechat · 2026-08-07
To break the monopoly of a few researchers on evaluating video models, the FuneralAI team introduced memebench, a blind testing benchmark based on memes and complex prompts. The test uses 50 popular memes from Bilibili and Douyin, inviting 400 members of the general public for blind A/B voting, calculating relative strength via the Elo mechanism.
The results were surprising: MiniMax H3 took first place, scoring slightly higher than the previously dominant ByteDance Seedance 2.0, while the cinematic-focused Seedance 2.5 only ranked third.
The analysis points out that although Seedance 2.5 produces the most photorealistic images, it severely regresses in complex action continuity, dubbed an "advanced PPT." MiniMax H3, on the other hand, caught up due to its extreme instruction-following capability and action coherence, showing significant advantages in intense action scenarios.
Related event: MiniMax H3 vs Seedance 2.5: Benchmark Tests and Blind Survey(6 posts)→
More from Multimodal
- Seedance 2.5 Deep Dive: From Short Video to Cinematic Quality, AI Video's Leap and Industry Impact — 数字生命卡兹克 · 2026-08-07
- Runway Releases Seedance 2.5: More References and Longer Clips — runwayml · 2026-08-07
- Runway Launches Seedance 2.5 with More References and Longer Clips — runwayml · 2026-08-07
- Magnific Integrates Seedance 2.5, Adds 3D Motion Control and 4K Upscaling — LudovicCreator · 2026-08-07
- Kyutai's MuScriptor Audio-to-MIDI Model Adds Automatic Tempo Detection — petewoodbridge · 2026-08-07
- Exploring KREA 2 LoRA Block Testing and Workflows — dariusredraven · 2026-08-07