DDRL scores lower reward than DanceGRPO yet wins human preference by 10.5 pts
gowthami_s · x · 2026-10-15
Experiment details from the DDRL paper: full finetuning of Cosmos2.5 pre-RL checkpoints (2B/14B), 93 frames at 720p, 1,044 test prompts. For 2B text-to-video, DDRL gets a lower VideoAlign reward than DanceGRPO (0.604 vs 0.715), yet leads human preference by 10.5 percentage points across 9.6k comparisons. Caveat: only 15 raters, roughly half of them authors — broader human eval would be welcome.
Related event: DDRL keeps diffusion loss during RL to fix reward hacking in video models(5 posts)→
More from Multimodal
- FastVideo's FastH3-Omni 8-step distilled model trends on Hugging Face with ComfyUI support — FastVideo · 2026-10-23
- Kylon Launches Annotations to Analyze and Remix Any Video Into Your Own Brand Version — testingcatalog · 2026-10-23
- Droste-effect AI image with infinite nested scenes stuns: 'You can see everything?' — yungcontent · 2026-10-23
- AI-generated Slavic folklore horror anthology 'Midnight Ryazh' debuts episode 1 — Creative_Dependent_6 · 2026-10-23
- Japan's ai& launches video generation with MiniMax H3: 12-second clips from $0.0125/sec — DavidBennett__ · 2026-10-23
- Canva unveils Craft, its first in-house image model that plans before generating — The Verge AI · 2026-10-23