The Tuesday Art Drop - 003

This week, I've started doing full length Music Videos...

The Tuesday Art Drop - 003
New type of street light

Sept 1. September Mood.

Start to finish - song in, finished September Mood video out - took six hours and eighteen minutes, and almost none of it was hands-on.

The pipeline transcribed the track, cut it into 33 lattice-aligned shots, wrote the anchor-image prompts and the shot prompts, generated all 32 reference stills through the Gemini API, rendered every clip, verified that each one came back at exactly its planned frame count, and assembled the final cut with the song re-muxed - unattended.

The human contribution was creative judgement: which takes were wrong, and why.

About two hours forty of the total was GPU time that can't be compressed; the rest went on a single bug that convincingly impersonated five others. Lip-sync was failing, and the plausible culprits — the speed LoRA, mismatched lyric tags, competing actions in the direction, the sampler settings, my own prompt wording - were each genuinely a little bit broken, so fixing them kept producing improvement without producing a fix.

The actual cause was upstream of all of it: the reference images had the singer in profile, and MiniMax H3 drives lip-sync from the face in the reference frame. Given no visible mouth, it animates a camera move instead. One regenerated anchor, face-on to the lens, and the problem disappeared.

The lesson worth keeping is not about lip-sync at all - it's that I'd been verifying that her mouth moved when the question was whether it matched, and reviewing anchors for identity, wardrobe and continuity without ever asking whether the model could actually do its job with them.

6 hours 20 mins to create !


Aug 31. Heat Hit

Full Length music video with original music

How hot does your computer run?
Didn't see this in Florida's Dali Museum

Aug 30. New type of street light

The above inspired by Owl-City's FireFlies


Aug 29. No More Waiting

The outstanding entry for this challenge, in my opinion, was by super talented discord user Jules, who managed a full length animated style music video. Jules was kind enough to share his techniques with me.

For local video generation, we need the fastest GPU with most VRAM possible...
Done Waiting Tables...
Done Waiting for the wakeup alarm to ring...

This one below is a High Quality Upscaled Full Music Video - that I generated via Claude AI automation.

Full Length - Hi Def Music Video