How training works

Last updated: October 15, 2026

Training shows a base model your images over and over, nudging it each time until it can reproduce what they share. Knowing what the settings control, and what failure looks like, tells you what to change when a finetune comes back wrong.

How it works

An image finetune is a LoRA: a small set of extra weights added to a base model, trained while the base model itself stays unchanged. At each training step the model looks at one image and its caption, tries to reproduce the image, measures how far off it was, and adjusts the LoRA a little. Thousands of these steps add up to a finetune that pulls the base model towards your set whenever it is applied.

The trainer works through the set one image at a time, at a fixed resolution. You don't set the batch size or the resolution; both are fixed. What you control is the dataset, the captions and a few settings.

The settings

Every one of these has a sensible default for the base model you pick, and Automatic sets them for you. The exact ranges and defaults are in Training settings and Trainable models.

Epochs

How many times the model sees the whole set. Ten images for one epoch is ten views; the same set for many epochs is the same ten images seen again and again. More epochs mean a stronger hold on the concept, and past a point, memorising the images instead of learning from them.

Steps

The total number of training steps: epochs multiplied by the number of images. Steps are what a run costs and how long it takes, so a large set at high epochs is a long run. Each trainer caps the total, and Layer lowers the epochs to fit when a big set would go over; see Training cost and time.

Learning rate

How big each nudge is. A higher rate learns faster and more coarsely; a lower rate learns gently and needs more epochs to get there. The two go together: when you raise epochs, lower the learning rate, and the other way round.

LoRA rank

The capacity of the LoRA: how much detail it can hold. A higher rank can capture finer detail and more varied subjects, makes a bigger file, and memorises small sets more readily. A lower rank holds the broad strokes and generalises well, which suits a single character or a simple style.

Checkpoints and preview prompts

Some trainers save the finetune part-way through as well as at the end. At each checkpoint, Layer renders your Preview prompts with it, so you can watch the concept come in over the run. Preview prompts don't change what is learned. The finished version uses the last checkpoint.

Write preview prompts that test what you care about: one close to the training images, one that asks for something the set doesn't show (a new pose, a new subject in the style). The second tells you whether the finetune generalises.

Reading the result

Underfitting

The finetune hasn't learned enough. Results look like the base model with a hint of your set: the character's colours are right but the face drifts, the style's line quality comes and goes. Raising the weight helps only a little. The usual causes are too few epochs or too low a learning rate for the set, captions that describe the very traits you wanted learned, or a set too varied to have one clear concept.

Overfitting

The finetune has learned too much. Results copy training images: the same pose, the same background, the same crop, whatever you prompt. Colours go harsh and textures look burnt, and the prompt stops being followed, especially at full weight. The usual causes are too many epochs or too high a learning rate for the set size, a rank high enough to memorise a small set, or a set whose images all share a background or framing you never meant to train. See Finetune results look overcooked or ignore the prompt.

What to fix first

When a finetune isn't right, change things in this order:

  1. The dataset. Off-model images, duplicates, blurry or upscaled files, and anything every image shares by accident (one background, one pose) are trained in as faithfully as what you meant. Removing a bad image or adding variety does more than any setting.
  2. The captions. They decide what is tied to the trigger word and what stays promptable. A trait you describe in every caption becomes optional; a background you never mention becomes part of the concept. See Captions and trigger words.
  3. The settings. Only once the set and captions are right. Change one thing at a time, compare the versions side by side, and keep the better one active with Use this one.

Settings tuned around a flawed set only trade one problem for another: fewer epochs to stop it copying a background also weaken the character.

Before you retrain, check the weight

A finetune that looks overcooked at full weight is often right at a lower one, and a weak one sometimes needs a higher weight rather than a new run. Try a few weights in the generation form before you spend Creative Units on a retrain. See Generate with a trained reference set.

Related