Training cost and time
Last updated: October 15, 2026
How a training run is priced in Creative Units, when the units are reserved, billed and released, and what decides how long a run takes. For the price of each model, see Creative Units and layer.ai/pricing.
How a run is priced
Image and edit finetunes
Charged per training step. Steps are epochs multiplied by the number of images (or pairs) in the run, so doubling the set or doubling the epochs doubles the steps. Each model has its own price per step, and some have a minimum charge per run.
Each trainer caps the total steps. When epochs times images would go over, Layer lowers the epochs to fit, so a very large set trains fewer passes rather than costing more. fal also rounds the step count to the nearest hundred.
The trainer changes the price: Kohya and AI Toolkit cost more per step than fal on the same model.
Preview images
The images rendered from your Preview prompts at each checkpoint are charged as ordinary generations on that model, on top of the training. Fewer preview prompts cost less.
Voices, motions and image styles
A flat charge per run, whatever the number of clips, references or the length of the description.
Seeing the cost before you start
The train button shows the estimated cost for the current settings, and updates as you change the images, the epochs or the trainer. If the workspace doesn't have enough Creative Units, the button says Not enough creative units to train this model. Click to top up.
When units are reserved, billed and released
When the run starts
Layer reserves the estimated cost from the workspace balance. The reserved units can't be spent on anything else while the run is going.
When the run finishes
The run is billed for the steps it actually ran, and the reservation is cleared. A run that trained fewer steps than estimated costs less than the estimate.
When the run fails or you stop it
The reservation is released and the training is not billed. Anything the run had already used, such as preview images already rendered, stays charged. Stop training on a running run asks you to confirm first.
When you delete a version
Delete version removes a trained version from the set. It doesn't refund the run: Creative Units already spent aren't returned.
Captioning before a run
Auto-captions are written before the run starts. If captioning fails, the run doesn't start and no units are reserved.
How long a run takes
What drives the time
The number of steps, which is epochs times images; the base model, since larger models take longer per step; the trainer; and how many checkpoints and preview images the run renders at the end.
While a run trains, its banner shows how long ago it started, the progress and an estimate of the time left. Near the end it reads Finalizing… while the finetune is saved and the previews render.
Timeouts
Each run has a time limit. A run that goes over it fails with Training took too long and timed out. Try again, or reduce the number of images or steps. Adjust settings on the failed run opens the settings with the epochs lowered for you. A timed-out run isn't billed.
A run that seems stuck
If a run goes far past a normal duration with no estimate of the time left, the banner changes to Taking longer than expected and offers Stop training. Stop it and train again; see Training failed.
Several runs at once
You can start a run while another is in progress on the same model; Layer asks Start another training run? first, because the second run spends Creative Units again. Each person can start a limited number of training runs per hour; past it, wait a while and try again.
Notifications
You get a notification when a run starts, when it finishes and when it fails, so you don't need to keep the set open. A failure notice says why the run failed.