← Back to Blog · 2026.07.31
How to Organize Tag Captions for LoRA Training Datasets
When you're training a LoRA, the quality of your results depends as much on your captions (tag lists) as it does on the images themselves. How you organize your tags determines what the model treats as a "feature to learn" versus what it treats as a "fixed constant to ignore." Here are some guidelines worth keeping in mind when you're refining raw tags from an image tagger model into real training captions.
1. Character LoRA vs. Style LoRA Need Different Captioning Strategies
If you're training a Character LoRA for a specific character, it's common practice to leave out the fixed traits that character always has — hair color, eye color, signature hairstyle, and so on. If you leave them in the caption, the model can overfit, tying "this character" too tightly to "this exact hairstyle," which makes it harder to generate the character with a different look later. On the other hand, tags that vary from image to image — pose, outfit, expression — should stay in the caption, so the model learns to treat them as variable elements rather than part of the character's core identity.
A Style LoRA works almost in reverse. Keep detailed captions for character, composition, and outfit, but remove the tags that describe the style itself (color palette, line style, etc.) — this sends the model a strong signal that "this style is always rendered this way."
2. Sorting by Category Makes These Calls Much Easier
Tags from an image tagger model all come back mixed together in one flat list, which makes it surprisingly tedious to decide what to keep and what to cut. AnimeTag-AI's "Category View" automatically reorganizes the same tag list into Basic, Appearance, Outfit, Pose & View, Props & Features, Background, and Etc, so decisions like "this is a Character LoRA, so let's drop the whole Appearance category and keep only Outfit, Pose, and Background" become much faster at a category level.
3. Three Criteria for Filtering Out Unnecessary Tags
- Image-quality/watermark tags — tags like
watermark,signature, andlowresdon't help training, so drop them. - Low-confidence tags — nudging the Threshold up a bit to exclude ambiguous tags (around 0.3 and below) from the start saves cleanup time later.
- Tags unrelated to your goal — if you want varied backgrounds, keep background tags; if you're only ever going to use one composition, drop the composition tags. Judge each case against what you're actually trying to train.
4. A Real-World Workflow Example
1) Upload your image to AnimeTag-AI to extract tags → 2) Set the Threshold to around 0.3–0.4 to filter out an initial pass of noisy tags → 3) Switch to "Category View" and review whole categories that don't fit your LoRA's purpose → 4) Click chips to remove individual tags, then grab your final caption with "Copy Prompt." If you already have a tag list and just want to clean it up without an image, paste it into the Tag Analyzer for the same category-based breakdown.