Training Captioner Prompts

Prompts

The caption style comes entirely from the prompt you send with each image. The Captioning Prompt panel on the left holds two parts:

Field	What it is
System prompt	Standing instructions: role, style, what to include or avoid, output format
User message	The short request sent along with every image (e.g. "Write a detailed caption for this image.")

Both are sent for every image, together with the picture.

Saved prompts

The drop-down at the top lists your saved prompts. Four are included:

Name	Output
Natural language	One detailed descriptive paragraph, no preamble
Short caption	One sentence under 30 words
Booru tags	Comma-separated lowercase tags (subject, attributes, clothing, pose, setting, style)
Character / person LoRA	Describes clothing, expression, pose, background and lighting but not permanent features such as hair colour or face shape, so the trigger word learns the identity

Buttons under the editor:

• Save: overwrite the selected prompt with the current text
• Save As…: store the current text under a new name (an existing name is replaced)
• Delete: remove the selected prompt (asks first)

Prompts live in %APPDATA%\RulzCaptioner\prompts.json. You can back them up or share the file. The last prompt you used is selected on the next start.

Changing the text without saving still works for the current run. It simply isn't stored until you press Save or Save As….

Prefix and suffix

Under Output in the same panel:

• Caption prefix: text added before every generated caption, e.g. ohwx person, (include the trailing comma and space yourself)
• Caption suffix: text added after, e.g. , high quality

They are added by the app, not the model, so they are always exact.

Writing good caption prompts

• Say what to leave out. "Do not mention that this is an image or photo" and "no preamble, lists or markdown" remove most unwanted filler.
• Match your trainer. Natural-language trainers (Flux, Qwen-Image, SD3 style) like full sentences; SDXL/Pony-style trainers often want tags.
• Describe what you want the model to not learn as part of the subject. For a character LoRA, caption clothing and setting but not the fixed traits.
• Keep it short. Small models follow a few clear rules better than a page of them.
• Test on a few images first. Select 3–5 images, press Caption Selected, tweak the prompt, repeat.
• Length is capped by Max new tokens in Settings. Raise it for long paragraphs.

Example prompts

Style-neutral, detailed

System: You are an expert image captioner creating training data for text-to-image models. Describe only what is visible: subject, clothing, pose, expression, setting, lighting, camera angle and style. Do not mention that this is an image or photo. Output one paragraph with no preamble. User: Write a detailed caption for this image.

Short tags

System: Output 15 to 30 comma-separated lowercase tags for diffusion training. No sentences. User: Tag this image.

Art style dataset

System: Caption images from an art dataset. Start with the medium and style, then the subject and composition, then colours and mood. One sentence, no preamble. User: Caption this image.