Training & Using LoRAs

Train a custom AI model on your own images, then call it in generations with a trigger word. Training happens on the board, on the images you already have there; applying a LoRA happens in the chat bar; your finished models live under LoRAs in Home.

Start a Training Run

Training begins from a board selection, not from a blank form — the images are the starting point.

  1. Select 10–30 images on the board. Below that range the training rows stay greyed out and the menu says the range instead of failing on click.
  2. Open the node actions menu on the selection.
  3. Pick Train Style (art styles, design language, aesthetic) or Train Object (the exact same subject).

The choice is made before the dialog opens, and the dialog opens committed to it. Flux LoRA is the training model.

The dialog shows your selected images on the left and the settings on the right. Give the model a Name; everything else is pre-tuned, and a value is editable by clicking it.

Pointer: any image on the board can be a training image — generated, rendered, upscaled or uploaded. See Boards, media library & stock.

Captions

Good captions describe everything except what the LoRA should learn — the caption acts like a subtraction: whatever you describe becomes interchangeable at generation time, whatever you leave unsaid binds to the trigger word as the learned style.

The Captions row picks how they are written:

  • Auto — every image is captioned for you when training starts. Auto-captioning costs credits per image, folded into the cost on the Start Training button.
  • Custom — the image grid turns into caption fields and you write each one yourself.

Captioning runs on one model for everyone; there is no per-run model choice to make. Captions are optional — an uncaptioned image trains on the trigger word alone.

What to train

With Auto captions, the What to train row says what the model should pick up implicitly. Everything outside it stays describable, so it can be steered later:

  • Design language — silhouette, proportions, structure.
  • The art style — the look: stylization and graphic feel.
  • The medium — the technique: paint, render, photo.
  • Custom — describe what matters in your own words and the caption chat writes the captions. Its answer can also be applied as a blueprint, so each image gets its own variation of the same caption shape.

Trigger-word position

Where the trigger word sits in the caption changes how the style is applied — and you should prompt the same way you captioned. The Trigger position row sets it:

  • Start (head, In the style of <trigger>, <description>) — for detail-rich, multi-part subjects (interiors, complex products, vehicles). The trigger sets the style before each part is described, so each one is style-conditioned as it's built.
  • End (tail, <description>, in the style of <trigger>) — for simpler, holistic subjects (exteriors, painterly styles) where there aren't many discrete parts to condition.

Design language is structural and trains cleaner at the start; the dialog says so when you pair it with End.

Pointer: for the deeper method — captioning as subtraction, training one car class to generate another, and divergent vs convergent datasets — see Training a design LoRA — the method.

Trigger Word & Steps

Trigger word is the rare token that becomes the handle for your trained look; you type it in prompts later to summon the style. One is generated for you — click the field to type your own, or the wand icon to roll another.

Training steps is how long the model trains. Higher means a stronger grip but more overfitting risk; around 1,200 is the sweet spot, and the value the dialog opens with is derived from your image count and the training kind. The field accepts 1,000–2,000.

Submit & Track

Click Start Training at the bottom of the settings panel. The button carries the total credit cost, captioning included, and stays disabled until the run is valid.

Once submitted, a Training Progress section appears in the same panel with the status and a percentage. You can close the dialog; training continues.

Find Your LoRAs

Finished and in-flight models are listed under LoRAs in Home, where you can search them, filter by status (All, Completed, Drafts, Training, Failed), favourite them, share them with your organisation, make them public, rename them and delete them. Open one to see its training images, captions and settings.

A setup you saved without training is a Draft, and it sits in the same grid as a card marked Draft, showing the images it was set up with. Opening one takes you back to the board it was saved on, with the trainer reopened on it — a training set is chosen on a board, so it is edited there.

Apply a LoRA in Generation (Trigger Word)

To use a finished LoRA in a generation, go to the chat bar / generation controls (see Chat bar: generating images):

  1. Open the LoRA bubble in the chat bar's input row and use its + to open the LoRA browser.
  2. Pick from your own completed LoRAs or community LoRAs. Up to 4 LoRAs can be combined.
  3. Each selected LoRA shows its Trigger word (copy it with the copy icon) and a Scale control. The slider covers the comfortable range; type a higher value (up to the API max of 4) to over-push. A Total Strength indicator warns when the combined scale gets high.
  4. Put the trigger word in your prompt to activate the LoRA. When a LoRA with a trigger word is added, Framo can auto-insert the trigger phrasing (e.g. "in the style of <trigger>") for you.

Pointer: the trigger word controls when your learned style appears; a LoRA applied without it may bleed into every generation. Where the word sits in the prompt should match where it sat in the captions — see Trigger-word position above.