Understand what each knob does, check that intuition, then sweep rank, steps, dataset size, targets, and scale — building a contact sheet gate by gate.

Chapter 03: The Experiments

Objective

Build a contact sheet: same fixed prompt, every adapter variant labeled. Learn to name underfit vs overfit by eye. Your tutor runs this one experiment at a time — not all five overnight without looking.

Prerequisite: Chapter 02 baseline (mflux generates, train folder + captions, at least one completed or in-progress baseline run).

Concept: What the knobs actually do

You are not hunting a magic hyperparameter combo. You are building visual vocabulary for failure modes you will later see in text and tool-call fine-tunes.

Knob	Intuition
Rank	Capacity of the adapter. Too low → cannot express the subject. Too high + tiny data → memorizes training photos.
Steps / epochs	How long you press the patch in. Too few → underfit. Too many on few images → photocopy of the train set.
Dataset size	More varied images usually help identity; more near-duplicates do not. Capacity × data is a tradeoff.
Target modules	Which layers get LoRA. Attention-only is often enough; “all linear” is more capacity and more ways to overfit.
Adapter scale	At inference (or baked scale), how hard the adapter pushes. Adapters are dials, not on/off switches.

What you are looking for in images

Underfit: “a dog,” not your dog; token does little.
Sweet spot: recognizably your subject in new scenes.
Overfit: same pose/background as a training shot; token works only as a photocopy machine.

Write one prediction before each run (one sentence). After the contact sheet, score your predictions. The goal is calibrated judgment, not a perfect LoRA.

Concept check

Q1. You train rank 64 for a long time on 5 near-duplicate photos of a dog on the same couch. What do you expect to see when you prompt TOK dog on the moon, and why?

Q2. Why is a contact sheet (same prompt, every variant labeled) more useful for learning than a single “best” image you like?

Concept answer key — attempt first

Answer key (concept)

Q1

Model answer: Strong overfit: the dog (or couch/pose) from the training photos pasted into a weak “moon” scene, or collapse to training frames. High capacity + tiny/narrow data memorizes examples instead of general identity.

Pass criteria: predicts memorization/overfit; links high rank or long train + few similar images

Q2

Model answer: Side-by-side comparison teaches the deltas between knobs; a single pretty image does not build vocabulary for underfit/overfit you will need later on text.

Pass criteria: comparison / deltas / judgment over one-off aesthetics

Gate: Fix the evaluation protocol

Before sweeping, lock:

One prompt used for every variant (include your trigger token + a scene not in the train set).
One seed (or a fixed short list of seeds) so noise does not look like a hyperparameter effect.
Output folder layout, e.g. ~/lora-eyes-open/sweeps/rank-4/, .../rank-16/ — each experiment gets its own checkpoint.output_path so zips do not overwrite.
Baseline reference: your Chapter 02 adapter (or note if it is still training).
Resume habit: if a sweep run dies, continue with mflux-train --resume …/checkpoints/NNNNNNN_checkpoint.zip (see Chapter 02 — checkpoints are full training state; the data folder must still exist). Do not restart with --config unless you intend a fresh run.

Paste your fixed prompt to your tutor before the first sweep run.

Gate: The sweep (one experiment at a time)

Copy your working train.json for each variant; change one axis only. Train (or load checkpoints), generate with the fixed prompt, save labeled images.

Experiment	Vary	What you’re looking for
A	rank 4 / 16 / 64	Where does capacity stop helping?
B	steps / epochs: short / medium / long	Where does it flip from underfit to memorized?
C	dataset 5 / 15 / 30 images (from your raw→train pipeline)	The data/capacity tradeoff
D	attention-only vs broader linear targets	Is attention enough for your subject?
E	adapter scale 0.3 / 0.7 / 1.0	Adapters are dials, not switches

For E, bake or load a scaled adapter when your toolchain supports it, e.g.:

mflux-save --lora-paths /path/to/lora.safetensors --lora-scales 0.7

(Exact flags follow your installed mflux version — check --help if names differ.)

Pacing: finish experiment A generations and look at them before starting B. Your tutor should stop you from launching all jobs blind.

Gate: Build the contact sheet

Deliverable: one grid image (or a clearly labeled folder of a single index page) — same prompt, every adapter variant labeled.

Pin it somewhere. You will refer back when a text LoRA in Parts 2–5 behaves strangely and you need to name what you are seeing.

Exit criteria

You can look at a generation and say “overfit” or “underbaked” without checking loss curves.
You can predict, before running, roughly what rank 64 on 5 images will do.

The one failure mode that matters here

Spending three days on this. It is rung one. Get the contact sheet, move on.

Check your understanding

Q3. Why change only one hyperparameter per experiment when building the contact sheet?

Q4. How does “adapter scale” (experiment E) differ from “rank” (experiment A) in what it is controlling?

Answer key — attempt every question first

Answer key

Q3

Model answer: So you can attribute the visual change to that knob. Changing many variables at once makes the contact sheet unreadable as science.

Pass criteria: attribution / isolate variables / interpretability of the sheet

Q4

Model answer: Rank is capacity of the trained adapter (how much it can represent). Scale is how strongly that adapter is applied at use time (how hard it pushes), even for a fixed trained file.

Pass criteria: capacity vs application strength / train-time capacity vs inference dial

← Chapter 02 · Next: Part 2 →