ABOUT

I’ve been testing how far local AI video can be pushed beyond single impressive shots - and into actual narrative work.

This music video was built as an experiment around LTX 2.3, not just to generate a few strong clips, but to see whether a local model could hold together a coherent story across many different shots, moods, locations, and camera setups.

That was the real challenge.

Not character consistency alone - that part was helped by training LoRAs for both image generation and LTX 2.3 - but getting everything else to work together at the same time: the right framing, camera movement, performance, lighting, styling, continuity, and overall feel from shot to shot.

That’s where this kind of work stops feeling like “prompt in, video out” and starts feeling much closer to directing.

A single shot can look great on its own. But building a sequence that actually feels connected is a very different problem. If one part breaks - motion, camera language, environment, wardrobe, mood - the whole shot often needs to be rethought.

What makes this especially exciting to me is that the entire project was built locally on a single machine.

The music was generated with ACE-Step 1.5.

Images and edits were prepared with Z Image Turbo, FLUX.2 [klein], and Qwen Image Edit.

The core of the project - and the part that carried most of the weight - was LTX 2.3.

The only non-local tool used was ChatGPT 5.4, which helped with the lyrics and prompt development.

The bigger takeaway for me is this:

The gap between large-scale production workflows and what a single creator can now prototype, direct, and assemble is shrinking fast.

Not because the tools are perfect - they aren’t.

But because even local tools are already becoming capable of much more than quick visual experiments.

Still plenty of rough edges. Still a lot to improve.

But this feels like a very real step toward more serious solo AI-driven filmmaking.