Free LoRA caption checker

What does the LoRA caption checker check?

Nine things on every line: the exact trigger word at the start, words about her hair and eyes, her face and skin, her body, age or ethnicity, duplicates, length, a framing word, light, and odd characters or stray spaces. It also checks the trigger word itself. Rule breaks are marked Fix; judgement calls are marked Check.

Framing words it knows: close-up, upper body, cowboy shot, three-quarter body, full body, plus selfie and mirror selfie. The framing goes last, or first when it’s part of the shot, as in “close-up portrait looking up” or “mirror selfie in a bedroom”. Both are what our Dataset Maker writes. A caption with several lines (common in files from an AI captioner) is joined into one line in the fixed version.

Why leave her face and hair out of LoRA training captions?

Because the trainer ties whatever the captions don’t mention to the trigger word. Leave out her face, hair and eye colour, and they become part of “zvx woman”. Write “long brown hair” in every caption, and the LoRA learns hair as a separate, changeable detail, so it can drift from image to image.

A caption that passes every check:

zvx woman, standing and looking slightly to the left, soft smile, navy blazer over a white top, train station, overcast daylight, upper body

Our Dataset Maker writes one .txt caption per image in this order when it builds a 50-image datasetfrom one face: the trigger word, then the template photo’s own caption. In an older lesson the captions came from Gemini, in batches of 10. Whichever captioner you use, read what it gives you: an AI captioner usually describes hair and eye colour unless it’s told not to, and that’s the first thing this checker catches.

Why do prompts get hair and eyes when captions don’t?

Captions teach; prompts ask. When you train, leaving hair and eyes out ties them to the trigger word. When you generate, our videos add one short hair-and-eyes line after the trigger word, which keeps those two details steady from image to image. Her face shape, makeup, skin and body stay out of both.

The same scene as a training caption, then as a prompt:

zvx woman, standing and looking slightly to the left, soft smile, navy blazer over a white top, train station, overcast daylight, upper body

zvx woman, long wavy dark brown hair, middle part, hazel eyes, standing and looking slightly to the left, soft smile, navy blazer over a white top, train station, overcast daylight, upper body, candid smartphone photo, natural skin texture

So don’t run your prompts through this checker: it will flag the hair-and-eyes line, and that line belongs there. The freeAI influencer prompt generator builds prompts in the right format.

What makes a good LoRA trigger word?

A made-up token the model has never seen, plus one plain word for what she is. Our default is “zvx woman”. The token carries no old meaning, so nothing leaks into her face, and “woman” tells the model what it’s looking at. Use the exact same spelling in every caption, in training and in every prompt. More on picking one, and what AI Toolkit does with it, in how to pick and use a LoRA trigger word.

• Not a real word or name. “Emma” or “velvet” already mean something to the model, and that meaning pulls on her face.
• Lowercase, one space, no commas. A comma splits it in two. Capitals are one more thing to mistype across 50 files.
• Short. Three or four letters is plenty. You’ll type it hundreds of times.
• “woman”, not “girl”. Characters on this site are always clearly adult, and the trigger word should say so.

How do you fix captions the checker flags?

Fix the trigger word first, because it touches every line, then work down the Fix rows and judge the Check rows yourself. The fixed version handles the mechanical part: trigger first, hair, eye, face and skin words removed, characters cleaned. You still read it before training.

1. Type your trigger word exactly as you set it for training. In the Dataset Maker that’s the trigger_word field.
2. Paste your captions, one per line, or pick all the .txt files from your dataset folder.
3. Read the trigger word notes. A weak trigger is worth fixing before anything else.
4. Go through every Fix row, then decide on each Check row: “petite” might be her body or a clothing size.
5. Copy or download the fixed captions. With files, the download is one .txt with each caption under its file name: paste each one back into its own file.
6. Keep each caption file named like its image (image_01.png and image_01.txt), then run the check again until it’s clean.

What can go wrong?

Most failures come from one bad file in a set of 50, or from trusting a word list too far. Here are the usual ones, and what to do about each, before you spend GPU time and money on a training run.

• One file has a typo in the trigger word. That caption doesn’t match the rest, so that image teaches the trigger word less. The checker names the line or file.
• The captioner described her hair in half the files. Her hair starts changing between images. Strip it from every caption, not just most.
• The checker can’t see your images. It can’t tell whether image_01.txt describes image_01.png, or whether the outfit is right. Read a few captions against their photos.
• It’s a word list, not a reader. It may flag a word used another way, or miss a description phrased oddly. That’s why body, age and ethnicity words are Check, not Fix.
• A fixed caption lost a whole part. A part that only describes her hair, like “long wavy brown hair”, is dropped whole. A pose like “hand in her hair” stays, because it doesn’t say what her hair looks like. If a fix leaves the caption short, describe a new pose.
• Hidden characters from copy and paste. Word processors and chat apps add curly quotes and non-breaking spaces. The fixed version swaps them for plain ones.