Band's local-AI music video hits 200k views: 900 generations kept 64, 3-second clips beat face drift
Hwang0500 · reddit · 2026-10-27
A shoegaze band's self-made AI music video passed 200k YouTube views (previous best: 3k). The maker, on an RTX 3060 12GB + 32GB RAM, shared the full workflow.
Stills (all local)
- SDXL base + a style LoRA trained on 40 of the band's own film photos (kohya, 15 epochs, a full weekend)
- A character LoRA from 25 photos of the singer for cross-shot consistency
- ComfyUI + ControlNet depth for specific compositions (empty pool, motel hallway)
- 900 generations, kept 64
Motion (cloud): local i2v yielded only 2 usable seconds in a week of trying, so image-to-video ran on Dreamina (near-free first month); faces drift after 3 seconds, so every clip is cut at 3 seconds and singer shots were rerolled often
Editing: free DaVinci Resolve, cuts on the snare; film grain and halation over everything to hide the AI shimmer; 3 real garage-performance shots mixed in — viewers guess those are the AI ones
Lessons: 40 training photos was barely enough for the style LoRA; plan shots before generating (300 generations wasted "exploring"); 2–3 seconds is the sweet spot before clips melt.
More from Multimodal
- Magnific teases October 28 release: 'We're making it move' — cuenca · 2026-10-27
- Midjourney --sref demo: retro Americana Batman art series — gizakdag · 2026-10-27
- AI video is a new medium, not cinema or animation — and it's perfect for depicting dreams — Kyrannio · 2026-10-27
- Dad turns his toddler's 31-year-old quote into a song via SUNO kids' choir — NathanpmYoung · 2026-10-27
- Adobe Firefly video editor walkthrough: one-sentence Quick Cut turns 32s into 20s — LudovicCreator · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26