Train your own LoRA

A LoRA is your own small model. Train it on your pictures, then use it any time you generate.

Start training See a real training
What can I train?

My models

Follow the trainings in progress, pick the epoch to keep from a finished one, publish it.

A real training, start to finish

Our mascot, trained in the Studio exactly the way you would. Every picture below is the real thing.

1

Gather your pictures

A dozen or more drawings of what the model should learn, varied in pose and framing. Here, the mascot’s reference sheet from the OC Creator: turnarounds, faces, expressions and a few scenes, sent to the Studio in one click.

2

Labels, minus what never changes

Every picture came labelled. What changes from one to the next is tagged: the pose, the place, the expression. What never changes, the black hair, the red eyes, the white shirt, is left out, so the trigger word ashi_girl learns it.

ashi_girlfull bodystandinglooking at viewerrainnightumbrellatransparent umbrellaoutdoorsstreetneon lightsreflection
ashi_girlupper bodysittingfrom sideindoorscafewindowcupholding cupday
ashi_girlportraitfrom sidelooking at viewerangryclenched teethv-shaped eyebrowssimple backgroundwhite background
3

About 30 minutes on our GPUs

Every epoch renders the three sample prompts, so you watch the model learn. The early ones barely know her. By the last, it is her, in places the dataset never showed.

  • BaseIllustrious
  • Steps1,200
  • Epochs10
  • Cost2,000 coins
Epoch 1Epoch 3Epoch 6Epoch 10School uniform, classroomKimono, summer festivalArmor, castle
4

Then she is one word in your prompt

Keep the epoch that looks right, and ashi_girl works in the studio like any LoRA: any outfit, any place, next to other characters and styles.

Train yours
ashi_girl, maid, maid headdress, holding tray, cafe, smile, cowboy shot
ashi_girl, hoodie, sweatpants, headphones, bedroom, sitting on bed, holding controller, night
ashi_girl, winter coat, scarf, snowing, city, night, smile, upper body
ashi_girl, witch hat, long dress, cape, riding broom, flying, night sky, full moon

What you can train

Three kinds, one wizard. Each gets its own settings and its own kind of trigger.

Character LoRA

Lock in a face, a hair, an outfit and the small signature details, so the same character comes back in any scene, angle or lighting. For original characters and series work.

Start training

Style LoRA

Capture a way of drawing: the linework, the shading, the palette, the texture. Applied to any subject, it gives a whole set of pictures one hand. No trigger word needed.

Start training

Concept LoRA

An object, an outfit, a pose, a prop the base model does not know. Summoned by its word, it drops into any character's picture.

Start training

How training works

  1. Prepare the pictures

    Between 4 and 200 drawings that show what the model should learn, varied in pose and framing. Photos of real people are refused.

  2. Pick the kind and the base

    Character, style or concept, on Illustrious or Anima. The settings follow from your image count; the hands-on path lets you change them.

  3. Tag

    Every picture gets suggested booru tags; keep, fix or add, with the same autocomplete as the prompt box.

  4. Start

    The run costs coins, from 2,000, shown before you confirm. Previews arrive as it trains.

  5. Keep an epoch and generate

    Choose the epoch whose previews look best. The model is private to you unless you publish it on your page.

Why train here

In the browser

No tools to install, no GPU on your side, nothing to configure.

For everyone

First-timers take the quick path; regulars edit tags and settings.

A set of your own

Characters, styles and concepts you reuse, private or published.

Same tools after

A trained model works with pose control, casts, inpainting, animation.

More AI tools

Everything else ashigen does, one click away. See them all in the Studio.

AI anime generatorAnime art from tags or plain words, free every day.OC CreatorDesign an original character with dropdowns, see them in seconds.Reference sheetFifteen views of your character: turnaround, faces, scenes.ProImage to videoA few seconds of motion from one picture and a line of direction.ProUpscaleMore detail and a bigger picture, in one pass.ProRemove backgroundCut your character out onto transparency.ProInpaintPaint over one part of a picture and redraw just that.RelightMove the light on any anime picture, with shadows, rim light and coloured lights. Free.

Questions

What is LoRA training?

Teaching the model a new thing, a character, a style or a concept, from your pictures, so it can draw it on demand. The result is a small file that loads next to the base model.

How many pictures do I need?

Four at the very least, around twenty for a character, up to two hundred. Quality beats quantity: clear subject, varied poses and framings, the same style throughout for a style model.

What should the trigger word be?

A made-up word nobody else would type, like a name. It is what you write in a prompt to summon the character or the concept; a style needs none.

How long does it take?

About half an hour on our GPUs for a set of a dozen or two pictures, longer for big sets and long runs. Previews arrive during the run.

What does it cost?

Coins, priced per run from 2,000 and shown before you confirm. Generating with the trained model costs the same as any image.

What pictures are allowed?

Drawings and renders of fictional characters. Photos of real people are screened out, and the content policy applies to everything.

Can I download the file?

Not from the site; the model lives in your account and runs here. Publishing shares it with other ashigen users.

Can I retrain later?

Yes. Train a new version with more pictures or better tags; each run keeps its epochs so you can compare.