AnimateDiff’s 7.3 versus Wan v2.6 Image to Video’s 15.8 is not a close result; it’s a separation in kind. Model A can hold a shot together over time, but in both tests it defaulted to safer, more generic imagery instead of executing the prompt. Model B consistently translated specific written direction into visible, cinematic choices on screen. In Cinematic motion , Wan v2.6 did the real directing work: a courier in a narrow wet alley at blue hour , wearing the rain cape , carrying pastry box...
source & further reading
runtimewire.com — original article
Happy Horse routs AnimateDiff Turbo on prompt fidelity
Zyphra Releases ZONOS2, an Open-Weight Real-Time Voice-Cloning Model
RuntimeWire — Weekly Report Week 2 · June 5 – June 11, 2026