Captions and trigger words

A caption is the text a finetune trains on next to each image. Captions decide what the finetune ties to your trigger word or character name, and what stays free for your prompts to change.

How it works

During training, the model sees each image with its caption and learns to connect the two. Whatever the caption names, the model learns as something a prompt controls. Whatever the caption leaves out, and every image shares, the model learns as part of the concept itself, tied to the trigger word or the character's name.

So the rule is: the finetune learns what you leave out. Describe what varies from image to image, and leave out what should always be there. A character set's captions name the pose, the setting and the framing, but never the character's hair or costume. A style set's captions name the subject, but say little about the rendering.

Captions show under each image when you open an image finetune in the set's editor (Add a caption… on an empty one). Image finetunes train each image with its own caption. Capitalisation makes no difference: training text is lowercased.

Auto-captions

When you start training, Layer captions every image that has no caption, using a vision model and the set's category. Captions you wrote are never overwritten. The new captions are saved to the set, so you can read them, fix them and retrain. If captioning fails, the run stops before it starts and no Creative Units are spent.

Treat auto-captions as a starting point. They describe each image well, but for style, object and scene sets they can mention the art style, which you may want to trim (see below). The agent can also draft captions for you and show them in a review card, where you accept, edit or regenerate each one; see Review and edit training captions.

What to describe, by category

The set's category decides how Layer captions it, and the same guidance works when you write captions yourself.

Art style and general sets

Describe the content of each image: the subject, the view or angle, the composition, the colours, the lighting. Keep any words about the rendering (brushwork, line, finish) short and the same on every image, or leave them out, so the look attaches to the trigger word instead of to a phrase you must repeat. Leave out the background colour unless it changes and you want to control it.

Character sets

Refer to the character by the set's name, spelled exactly the same way every time, as part of the sentence. Describe only what changes: the pose or action, the setting, the shot (close-up, full body), the expression, and props or clothes the character doesn't usually have. Leave out everything that makes the character who they are: face, hair, body, colours, marks, signature outfit, and the art style. If you describe the character's red scarf in every caption, the scarf becomes optional; leave it out, and it comes with the name.

Object sets

Describe the view and the specific item: what it is, what it is made of, its condition, any variation in shape or colour between items. For a family of icons or items, name the item and leave the shared frame, outline and lighting out, so every new item comes back in the family look.

Scene sets

Describe the composition, the key elements, the time of day and the atmosphere. Leave out what defines the place across every image (the architecture of a recurring town, say) when the set is one location; describe it when the set is a type of scene and the elements change.

Edit pairs

An edit finetune trains on an editing caption per pair (Describe the edit under the pair), not on a description of either image. Write it as the instruction you will give when you use the finetune: "add a thick black outline and a soft drop shadow", "turn the summer scene into winter, with snow on every surface". Use the same wording for the same edit across pairs; Apply to all pairs writes one description to every pair. Every pair you train needs one.

Trigger words and character names

Trigger word

A word the finetune learns to associate with the concept, set under Show advanced settings when you customise training. Pick something that isn't an ordinary word, so it carries no meaning of its own ("mystyle", "frostink"). Layer puts the trigger word in front of every caption during training, and in front of your prompt when you generate with the set attached, so you don't type it into captions or prompts. If a style, object or scene set has no trigger word, Layer trains on an internal one and applies it the same way.

Character name

Character sets anchor on the set's name instead of a trigger word. The name goes in every caption, in place of a description of the character, and Layer adds it to the front of any caption that lacks it. Choose a name that is short and unusual enough not to collide with a common word ("Captain Rhea" rather than "Hero"), and keep it once you train: renaming the set later is safe, since a finetune keeps the name it was trained with.

Always applied text

Text you want in every generation, such as a framing note or a required background, belongs in the set's Always applied prompt. It is added when you generate, and is not part of the training text, so changing it never needs a retrain.

Worked examples

A character set

A set named "Captain Rhea", an illustrated ship's captain with silver hair and a long blue coat.

Weak: "Captain Rhea, a tall woman with short silver hair and a scar over her left eye, wearing her long blue naval coat with gold buttons, standing, detailed digital painting, cel shaded."

This describes her identity and the art style, so the finetune learns them as things a prompt can change, and the caption says almost nothing about this image's pose or setting.

Better: "Captain Rhea leans on the ship's rail at dusk, looking out to sea, seen from behind at waist height."

Only the pose, setting, time and framing are named. Her face, hair, coat and the painting style come with the name.

A style set

A style set with the trigger word "frostink": soft ink lines and pale washes, many subjects.

Weak: "Beautiful fantasy art in a painterly watercolour style with soft ink lines and a muted blue palette, masterpiece, a fox."

The style words compete with the trigger word that Layer adds in front, and "a fox" is all the model learns about the subject.

Better: "A fox curled asleep on a snowy tree stump, three-quarter view from above, pale morning light."

The subject, view and light vary from image to image and stay promptable; the ink-and-wash look attaches to the trigger word.

Writing a caption set

• Keep one structure, vary the words. Cover the same things in the same order on every caption (for a character: name, action, setting, shot), but write each sentence fresh. A template with one slot changed teaches the template.
• One or two sentences. Most important words first. Specific beats long: "kneels to tie a boot" is worth more than a paragraph of adjectives.
• Describe what is there. A caption that names something missing from the image, or misses something prominent, teaches the wrong link.
• Read them all before you train. One caption that describes the character's outfit, among twenty that don't, can be enough to loosen it.

When an LLM helps

A language model is useful for drafting a whole caption set to a fixed pattern: give it the pattern, the list of what changes in each image, and the things it must never mention, and ask for one or two sentences per image. Then read every caption against its image. Don't ask it to describe everything it can see, and don't accept long captions that restate the style or the character's fixed traits in every line: they work against the finetune, which learns that detail better from the images when the captions leave it out.

Related

• Review and edit training captions
• Finetunes
• Train a character finetune
• Train a style finetune
• A trained character drifts off-model