Yumina

Guides

How to keep an AI character consistent with a LoRA (2026)

Train a LoRA from one portrait so your AI roleplay character keeps the same face, hair and clothes in every picture. The steps in Yumina Atelier, the pictures to prepare, and the limits.

To keep an AI character consistent, train a LoRA on that character and use it for every picture. In Yumina Atelier one portrait is enough to start: Atelier draws 18 training pictures from it, you pick the good ones, and training takes about 15 minutes. The whole run costs 2,300 mushies, refunded if it fails.

What is a LoRA?

An image model draws a new person each time you ask, even when the description is the same. A LoRA is a small model file that teaches the image model one character or one art style. With the LoRA loaded, the model draws that character with the same face, hair and clothes every time. You call the character by a trigger word, a made-up word the LoRA learned as that character's name. Atelier makes one for you, "ym" followed by six letters.

What pictures should you prepare?

You have two ways to start.

One portrait Several pictures
What you bring 1 picture of the character 8 to 40 pictures of one character or one art style
What Atelier adds 18 drawings of the character in new views, poses, expressions and places nothing, it trains on your pictures
Cost 300 mushies for the drawings, then 2,000 to train 2,000 mushies to train
Time about 3 minutes to draw, about 15 to train about 15 minutes

For one portrait, choose a picture that shows the face, the hair and the outfit clearly. A vision model reads the clothes the portrait does not show, so every full-body drawing invents the same ones.

For several pictures, variety matters more than count. The 18 drawings Atelier makes are a good checklist: a full-body front view, a three-quarter view, a side profile, a view from behind, sitting, walking, close-ups that laugh and frown, a view from above and one from below. Atelier shrinks any picture larger than 1,536 px on its long side before training.

How do you train a character LoRA in Atelier?

  1. Open Atelier and go to LoRA training. Click Train a LoRA.
  2. Under "Train from", choose One portrait or Several pictures. One portrait always trains a character. With Several pictures you also choose Character or Art style.
  3. Pick the portrait from your pictures or your character portraits, or upload one. Give the LoRA a name.
  4. Click Start training. The line above the button shows the total: about 18 minutes and 2,300 mushies. Atelier charges 300 now and draws the training pictures, up to 18.
  5. Pick the training pictures. Keep the ones that look like your character and drop the rest. You need at least 8.
  6. Click Train (2,000 mushies). The stages are Preparing pictures, Waiting for a GPU, Training with a percentage, Saving, and Drawing samples.
  7. When it finishes, you see three sample pictures and the trigger word. The LoRA is in My LoRAs.

If a character on one of your cards already has a portrait, the LoRA window on that character has a shortcut, Train one from the portrait, which uses that picture.

A character LoRA trains for 800 steps at 1,024 px. In Yumina's test on 19 pictures from one portrait, that kept the face, hair, glasses and outfit in every new scene with the trigger word alone. Doubling the steps made the character stiffer and no closer.

How do you use the LoRA so the character stays the same?

  1. On a character in a card: open the character's settings and choose the LoRA under Illustration LoRA. Every story illustration that includes the character uses it.
  2. On yourself: choose it for your persona under Personas, so you look like yourself when you are in the picture.
  3. On a whole card: Card settings → Story illustrations → Art style LoRA applies one art style to every illustration.
  4. In Atelier: open My LoRAs and click Draw with it, or add it to a Custom image. A Custom image takes up to 3 LoRAs at 2 mushies each, and marks a LoRA that does not fit the base model you picked.

Strength starts at 0.9 for a character and 0.8 for a style. A character LoRA brings its original outfit along. When the card's clothes matter more than the outfit in the portrait, lower the strength to between 0.6 and 0.8.

What limits does LoRA training have?

  • 8 to 40 pictures per training.
  • One training at a time per account. A second one is refused with "One is already training".
  • Your LoRA list has a cap, 20 in the default setup. Past it, Atelier asks you to delete some first.
  • Story illustrations take Anima LoRAs only. LoRAs trained in Atelier are Anima LoRAs. An imported LoRA for another model is refused with a message saying so.
  • A LoRA imported from Civitai needs an author who allows use on generation services.
  • Two characters with LoRAs in one picture mix into each other. The Vivid illustration style draws one person per picture.
  • Changing a character's LoRA on a published card sends the card back to review.

You do not have to train. The LoRA square lists LoRAs other people have shared, and Add to my LoRAs puts one in your list.

How to make AI video of your character starts from a picture like the ones your LoRA draws, and What is Yumina Atelier? lists the other studios.

FAQ

How many images do I need to train a character LoRA?

One portrait is enough in Atelier, because it draws 18 more for you. If you bring your own set, you need 8 to 40 pictures.

How long does LoRA training take?

About 3 minutes to draw the training pictures and about 15 minutes to train. A busy GPU queue adds waiting time.

How much does it cost to train a LoRA on Yumina?

2,300 mushies from one portrait: 300 for the drawings and 2,000 for training. From your own pictures it is 2,000. A failed or cancelled training is refunded.

Can I use a LoRA from Civitai?

Yes. Paste the Civitai link or upload the .safetensors file in the LoRA picker. For story illustrations it has to be an Anima LoRA, and its author must allow use on generation services.

Sources

Checked 15 November 2026. What a LoRA does and where to attach one: Playing a world and the Illustration LoRAs section of Visuals and audio. Training from one portrait and the 2,300-mushie price: the update log at yumina.io/app/updates, entry of 15 November 2026. Picture counts, step names, times, the 18 drawings and the training test: the LoRA training screen and its settings as shipped that day.

Worlds you can play now

See every world on Yumina

More on atelier

Read next