How to pick images for a character LoRA
The dataset decides almost everything. This short guide covers how many images you need, which ones to keep, which ones to throw away, and how to caption them.
Works for Qwen-Image, Flux, SDXL and most modern models. Updated 6 November 2026.
Your LoRA learns your mistakes too
A LoRA copies whatever your images have in common. If most of your images are close-ups in warm light, it learns "this character is always a warm close-up." If the face slowly drifts across your images, it learns the drifted face. Training settings can't fix a bad dataset, but a good dataset works with almost any sensible settings.
Fewer, better images
Twenty strong, varied images beat sixty similar ones. When you're unsure about an image, leave it out.
Mostly close, some far
The face and the small details are the hardest part, so they get the most images. Full-body shots teach proportions and outfit.
For 24 images, that's about 10 close-ups, 8 head-and-shoulders or half-body shots, and 6 full-body shots.
Change everything except the character
The character should be the only thing every image has in common. Change everything around it.
- Angle
- Front, three-quarter left and right, side profile, back view, from slightly above and below
- Expression
- Neutral, smiling, laughing, angry, sad, surprised, thinking
- Lighting
- Soft daylight, warm lamp, cool night light, dramatic side light
- Background
- Plain, indoors, outdoors, busy, simple. Avoid the same wall in every shot.
- Pose
- Standing, sitting, walking, pointing, holding objects
- Outfit
- The signature outfit in most images, plus 2–4 others so the outfit can be changed later
Write down the must-keep details
Before you collect anything, list the 3–6 details that make the character recognizable, such as eye colour, a scar, a hair shape, an ear notch or a tail tip. Then check every image against that list.
The keep-or-cut test
Keep
- Sharp, well lit and clearly the character
- The face is big enough to see the details (200+ pixels across)
- Every must-keep detail is correct
- It adds an angle, expression or setting you don't have yet
- At least 1024 pixels on the short side when possible
Cut
- Blurry, heavily compressed or tiny
- The face is cut off by the frame or hidden
- Other characters prominent in the shot
- Text, logos, watermarks or signatures
- Heavy filters or a very different art style
- A near-copy of an image you already kept
- Any must-keep detail is wrong
Caption what changes, not what stays
Start every caption with a rare trigger word, then describe only the things that change between images: the shot, the pose, the expression, the setting and the outfit. Leave the fixed features out. The LoRA then links them to the trigger word, so you never have to describe them again.
Pick a rare trigger and test it. A common word like "Anna" pulls results toward its usual meaning. On Z-Image, wr3nn sometimes printed as text on clothes, so we test full names too.
Only train on images you're allowed to use
- Your own art, your own generations, or images you have written permission to use.
- Never a real person without their clear consent.
- No characters owned by studios or brands if you plan to share or sell the LoRA.
- Check that your base model's licence allows what you plan to do with the LoRA.
Don't want to do all of this yourself?
Bravikron Labs is getting ready to train custom character and style LoRAs for you. We'll handle cleaning, captioning, training and testing, then deliver the LoRA file with sample images and prompts. Join the waitlist to hear when orders open.