I ran the whole video through H3’s Turbo LoRA from start to finish, and the settings were super lean: 10 steps, 0.4MP. The first three shots used image-to-video, and the last two used first-and-last-frame interpolation. The author himself said it’s still a work in progress, not a finished product, but what came out was already timeline-ready.
The speed difference is night and day—before, for the same length, I’d have to leave the machine running and go grab a meal before coming back to check. There are issues too: in some audio sections, it sounds like it’s been stretched, and the rhythm drifts a bit. It’s not noticeable in short shots, but it’s obvious in longer ones.
The music at the end was something I layered in during post; the visuals didn’t need a single manual fix. The official version hasn’t even warmed up yet, and the community’s acceleration solution is already out—this iteration speed is honestly kind of terrifying.
