Write captions

Describe what should change from picture to picture. Leave out what the LoRA should learn.

Skye DeskPoint here to try the real app

1. 1
   Select a photo and type its caption in the box on the right. It saves when you leave the box or press ⌘S.
2. 2
   Start every caption with the trigger word; it is highlighted as you type.
3. 3
   Keep it short. The word and token count above the box turns yellow when a caption gets long.

Skye DeskPoint here to try the real app

1. 1
   Click Trigger word problems on the left to see captions that miss the word or spell it another way.
2. 2
   Click Find/Replace and open the Trigger tab.
3. 3
   Choose Rename from…, type the old spelling (vel_trench) and check that the trigger word field says vell_trench.
4. 4
   Read the preview: each caption before and after. Only a whole comma-separated item that matches the old spelling (capitals or not) changes. If a caption then has the trigger word twice, only the first is kept.
5. 5
   Click Apply. ⌘Z undoes the whole batch.

More on the Trigger tab

• Add where missing puts the trigger word at the start of every caption that has none. It does not fix a misspelt one.
• Move to the start moves the word to the front of each caption; Remove takes it out.

With Find & replace

Find & replace works for any text. On its tab, find the misspelt word (vel_trench), replace it with your trigger word (vell_trench) and tick Whole words, then check the preview and click Apply.

Skye DeskPoint here to try the real app

1. 1
   Open Tags under the grid. The bar shows what share of captions uses each word.
2. 2
   Click a word to see the photos that use it.
3. 3
   Right-click to rename or delete a word, or open All… to merge several into one.
4. 4
   Drag words to reorder them; the trigger word stays first.

If one background or pose is in most captions, the LoRA will learn it too. Rewrite those captions or shoot more variety.

Coming later

None of this is in the first beta. It is here so that you can plan the disk space and check the licences.

Which models

Skye Desk offers a small first set: about 3.6 GB in all, with the Python they run in (about 1.5 GB).

Larger models are listed too, for when you have the space:

JoyCaption follows the Llama 3.1 Community License, which adds conditions – credit, naming and an acceptable-use policy. Read the licence text before you use it. Models whose licence does not allow commercial use, such as RMBG-2.0, Qwen2.5-VL-3B and the clothes and face parsing models, are not offered.

Download and remove

• Nothing downloads until you confirm. First Skye Desk shows each model’s name, size and licence, and how much free space will be left.
• Models come from Hugging Face. A download can be stopped and resumed, and every file is checked before it is used.
• They live in their own folder, ~/Skye Desk/models, outside your workspace and with their own Python: nothing else on your Mac changes.
• To free the space, quit Skye Desk and delete the model’s folder there.

What needs them

• Auto caption on the Dataset toolbar writes drafts for the photos you choose. They are a preview: nothing reaches your captions until you click Apply, and ⌘Z undoes the whole batch.
• Masks cuts out the clothes, the face or the hair, so that training can focus on them. Masks of the whole subject need BiRefNet from the larger set.
• Without the models everything else works as before, and you write the captions yourself.

Clothing

my_coat, a trench coat on a dress form, side view, plain grey background

Describe: the pose, the person or dress form, the background, the light, the camera angle.

Leave out: the fabric, colour, cut and details you want the coat to keep – the LoRA should learn those.

Characters

my_char, sitting on a bench, three-quarter view, soft evening light, city street

Describe: what the character is doing, the pose, the clothes if they should change, the place and the light.

Leave out: the face, the hair, the colours and markings that make the character who they are.

Hair

my_hair, woman seen from the side, indoor window light, plain wall

Describe: who wears it, the angle, the light, the background, the clothes.

Leave out: the cut, the length, the colour and the texture of the hair itself.

Products

my_bottle, on a stone plinth, top view, morning light, soft shadow

Describe: where it stands, the angle, the light, the props and the background.

Leave out: the shape, the label, the colours and the logo – the parts that must stay the same.

Styles

my_style, a lighthouse on a cliff, wide view

Describe: the subject of each picture – it should be different in every one.

Leave out: how it looks: the brush, the palette, the line – that is what the LoRA learns.