Lip-sync faceoff: LTX-2.5 beats MiniMax H3, AI talking-head reels cost just $0.32 each
victor_explore · x · 2026-10-05
A developer building AI pregnancy-educator talking-head videos for his app TenKicks ran a lip-sync faceoff between two open video models using the same face photo and voice.
- MiniMax H3: needed a negative prompt and a checker on every clip, 7.5 min per take, and kept drawing subtitles/logos on screen, nodding without talking, and moving the camera.
- LTX-2.5: open weights since August, voice drives the mouth, one command per take; produced 63s of talking video in 16 min on a single RunPod GPU.
- Cost: about $0.32 of GPU time per reel.
LTX-2.5 won decisively — all 11 reels were made with it.
More from Multimodal
- AI-generated music video roasting Threads users is cracking people up — venturetwins · 2026-10-05
- Open-source cinetic skill turns your coding agent into a launch-film director at 60fps — soumitrashukla9 · 2026-10-05
- H3 prompt trick: give subjects a subtle action to stop unwanted speech in AI video — SSj_Enforcer · 2026-10-05
- Fable 5.5's Oil Painting Timelapse Stuns Users as First Model With Real Creativity — dreamwieber · 2026-10-05
- Team Preparing Open-Source Krea 2 LoRA, Solicits Community Input on Evaluations — AIPnely · 2026-10-05
- Test: Claude cut a 9-minute Cybercab review into a short reel in one shot — EricETesla · 2026-10-05