Build a 15–20 image Dreambooth dataset with a trigger token and run your first mflux training from a JSON config.

Chapter 02: Train an Adapter

The Dataset

15–20 images of one subject. A person, an object, a place. Pick something you have a lot of photos of and can judge quality on instantly — that instant judgment is the whole point of doing images first.

Consistent subject, varied background/lighting/angle
512×512 (mflux’s training default)
Captions using a rare trigger token (TOK, sks, whatever) so the model binds the concept to a handle it isn’t already using

Training

Dreambooth training in mflux is driven entirely by a JSON config — training parameters cannot be passed as CLI flags. Start from the bundled example:

cd mflux
mflux-train --train-config src/mflux/dreambooth/_example/train.json

Copy that config, point it at your dataset, edit the examples section. Default is 512×512, batch size 1.

Expect several hours for a full run depending on machine and step count. Use --low-ram if memory pressure goes yellow — it constrains the MLX cache and releases the text encoders and transformer components after use.

While It Trains

Read the mflux source (Chapter 01’s homework). Or write down your predictions: what do you expect rank 64 on 5 images to produce? What about 2000 steps on 15 images? You’ll check these predictions in Chapter 03.

Known Failure Modes

“None of my LoRAs look good.” Known territory — early Dreambooth-on-MLX results were widely reported as poor, and adapters trained in one tool didn’t always load in another. Don’t over-index on quality; you’re here for the deltas between runs, not a shippable adapter.
Slowdowns after interleaving generation with training. Documented issue. Train, then generate. Don’t interleave.

← Chapter 01 · Next: Chapter 03 →