7 Dataset Mistakes That Quietly Ruin Your LoRA (and How to Catch Them)

When a LoRA comes out wrong, most of us start tweaking learning rates and epochs. But very often, the real problem is the dataset. Here are seven mistakes I kept running into, and how to spot them before you hit “train.”

1. Near-duplicate images
Resized, recompressed, or slightly blurred copies of the same image add weight to one pose without adding information. Keep real variations (different expressions, angles), but remove the copies.

2. Too many plain backgrounds
If almost every image has a white or gray background, your LoRA learns that background too. Then it struggles to put the character into a scene. Add a few images with real backgrounds.

3. Mixed style words in captions
One caption says “anime illustration,” another says “soft painting,” another says “sketch.” If the images actually share one style, unify the wording. If they don’t, consider removing the outliers.

4. Inconsistent tags
“black ribbon,” “black hair ribbon,” and “long black ribbon” are treated as different tags, so the meaning gets split. Pick one and use it everywhere.

5. Missing or misplaced trigger word
If you use a trigger word, it should be in every caption, ideally at the start. One or two missing captions are easy to miss by eye.

6. Invisible characters (BOM)
Some editors save captions with an invisible BOM character at the start. Your trigger word can then become “\ufeffmychar” instead of “mychar.” You’ll never see it, but the trainer will.

7. Not knowing your real step count
With buckets, repeats, and batch size, the actual number of steps can be very different from what you expect. Check it before training, not after.

How I check all of this now

I got tired of checking these by hand, so I built a small browser tool: LoRA Dataset Checker.

• Drop in your training folder and get a score out of 100, with a breakdown for duplicates, quality, resolution, captions, tag consistency, and balance

• See flagged images with the reason for each one

• Fix captions in bulk: trigger words, duplicate tags, BOM removal, and tag normalization

• Runs entirely in your browser. Your images never leave your PC, and original files are never modified

The free version has no image limit and covers detection and caption fixing. The full version adds exporting a ready-to-train folder (kohya_ss / ai-toolkit / OneTrainer), step and bucket calculation, and an improvement plan.

Download (free version available):
https://yuri-lab.booth.pm/items/9062642

Feedback and bug reports are very welcome in the comments!