Guides

Training a Character Style Without Leaving Your Mac: On-Device LoRA Fine-Tuning

20 September 2026 · RenderBob team

The LTX-2.3 MLX port can train a style or subject LoRA on the Mac. Reference clips stay local. It covers the small adapter job, not a foundation-model run.

An earlier note on in-house LoRAs treated training as a burst workload: useful for a character or a house style, and usually rented, because a full fine-tune is hungry and a campaign only needs it sometimes. The LTX-2.3 MLX port changes the small end of that range. Its train.py adapts the video model on the Mac, with no PyTorch, and the README's example run is 1,000 steps at rank 32.

A LoRA is a small adapter, not a new model

A LoRA does not retrain LTX-2.3 from scratch. It trains a compact adapter that nudges the existing model toward a style or subject. For image-model LoRAs, the earlier note cited a working range of roughly 15 to 50 images and 1,000 to 3,000 steps. The MLX video path is the same idea pointed at clips: precompute text and video embeddings from a local dataset, train the adapter, then pass the resulting safetensors file to generate.py. The repository does not publish a training-memory figure, so plan from the inference requirements already stated, about 10 GB at 4-bit and 14 GB at 8-bit, and measure the train run on your machine.

The reference set does not have to leave the Mac

Skipping the rental is the obvious gain: no instance to provision, no queue, no hourly bill for a short job. The sharper gain is where the pictures go. Unreleased client frames, a person's likeness, or brand assets used as the training set stay on the machine for the local train path. A cloud trainer can promise the same isolation. It still means an upload, an account, and a deletion policy. Local training skips that hop. Likeness material still needs consent. Keeping the files local does not replace that.

This fits the common adapter, not a foundation run

On-device training here is aimed at a character, a brand style, or a visual treatment on a modest set. It is not a substitute for training a foundation model, or for a high-volume run that wants a cluster. The code is marked for research use, and the base weights follow Lightricks' licence. The heavy, rare job still belongs on burst compute. The common job, one adapter for one project, can start on hardware the studio already owns.

In a timeline editor that can already generate on the Mac, that adapter is a session action: gather references from the project, train locally, and use the result on the next clip. RenderBob's local-or-cloud split is the same idea at shot scale. Keep the small, sensitive work on the machine. Burst when the model or the deadline outgrows it.