Tips and tricks

How to Tag Your LoRA Training Images (And Why Getting This Wrong Breaks Everything)

Most LoRA problems I see aren’t dataset problems. They’re caption problems. The model trained fine — it just learned the wrong thing because the tags were inconsistent.

Tagging is the text attached to each training image. The model doesn’t learn faces. It learns the relationship between words and pixels. Mess up the words, and it doesn’t matter how good your photos are.

image ↔ tags

Clean, structured tags → the model locks onto your subject. Sloppy or inconsistent tags → it doesn’t know what it’s supposed to be learning.

The Core Idea: Split Your Information

Before writing a single tag, understand that not all information belongs in every image. You need three buckets:

1. Identity — always present, never changes
2. Attributes — sometimes present (body, hair)
3. Context — varies by image (outfit, pose, environment)

Basic Tag Structure

[trigger_token], [identity], [hair], [face], [body], [clothes], [pose], [environment], [lighting]

Follow this order consistently. The model picks up on tag position, not just the words.

1. Trigger Token

This is your activation keyword — the one token that fires up the identity you trained. Every single image needs it. No exceptions.

raduwoman

Pick something that doesn’t already exist in the base model’s vocabulary. A made-up word works better than a common name.

2. Identity Tags

These describe your subject’s fixed features — the stuff that doesn’t change between photos. Write them once and copy-paste them into every caption.

young woman, caucasian woman, straight dark brown hair, shoulder-length hair, brown eyes

Don’t rewrite these. Don’t paraphrase them. The closer they are across images, the stronger the identity lock.

3. Body Tags

Use the same phrasing every time. This is where a lot of people slip up.

feminine physique, slightly accentuated curves, defined lower body

Swapping between “curvy,” “thick,” and “slim” across different images doesn’t just introduce variation — it actively confuses the model about what the subject looks like. Pick your phrasing and stick to it.

4. Clothing Tags

Unlike body and identity, clothing should vary. That’s the point. Each image shows a different outfit.

pink leggings, sports bra
jeans, crop top
summer outfit
gym outfit

5. Pose Tags

standing
walking
sitting
3/4 view
side profile
back view

6. Environment Tags

white background
gym
bedroom
outdoor
cafe
urban street

7. Lighting Tags

soft lighting
natural light
studio lighting
window light

Full Example Tag (One Real Image)

raduwoman, young woman, caucasian woman, straight dark brown shoulder-length hair, brown eyes, feminine physique, slightly accentuated curves, defined lower body, pink leggings, sports bra, standing, full body, white background, soft studio lighting

What Not to Do

Most LoRA training failures come down to a few repeating mistakes:

• Changing identity tags between images
• Over-describing (“beautiful stunning sexy attractive woman with amazing body” — the model can’t do anything useful with this)
• Mixing contradictory terms across the dataset

Keep identity stable. Keep body phrasing consistent. Let clothes and environment do the varying.

Advanced: What Experienced Trainers Actually Do

Keep Captions Short

Compact tags outperform verbose ones. The model doesn’t need a novel — it needs a signal.

Bucket Consistency

Group similar images and give them similar tags. All gym shots should share gym-specific language. All studio shots should share studio-specific language. This helps the model separate environment from identity.

Token Weighting (Optional)

Some training tools support weighting syntax to push certain tokens harder:

(raduwoman:1.2)

Use this on your trigger token if identity feels weak after training.

Example Dataset (3 Images)

Image 1 — Studio

raduwoman, young woman, straight dark brown hair, feminine physique, pink leggings, sports bra, standing, white background, studio lighting

Image 2 — Gym

raduwoman, young woman, straight dark brown hair, feminine physique, gym outfit, standing, gym, soft lighting

Image 3 — Street

raduwoman, young woman, straight dark brown hair, feminine physique, jeans, crop top, walking, street, daylight

The Actual Thing LoRA Is Doing

LoRA isn’t memorizing images. It’s learning which tags reliably predict which visual features. Show the same token next to the same features enough times, and the model binds them. That’s the whole mechanism.

Tag quality is model quality. Consistent identity tags, stable body tags, varied context tags. That’s the job.