What this guide covers
“The first image was perfect, the second one is a different person.” “Extending the video changed her face.” These are the most common frustrations when working with a character.
This guide explains why it happens, from the research side, and shows which methods are strong and weak at what, including the limitations the papers themselves state. Finally, we built one character on EGAKU and tested text only, a reference image, a character sheet, video and extensions in turn, and we show with images and frames where it held and where it broke. All 25 cited papers come with title, authors, venue and link.
Why characters drift (in one paragraph)
An image model draws every image again from noise. It does not remember “the person from before”. If you write “a woman with short black hair”, countless people match that description, and each time a different one of them gets drawn.
So to keep drawing the same person, you have to give the model some clue that identifies “this person” in addition to the text. The ways research has tried to give that clue fall into four broad groups.
A map of the four approaches
| Approach | What you do | Strength | Weakness (as the papers state) |
|---|---|---|---|
| 1. Teach it | Train on a few images of the character (LoRA, etc.) | Strong likeness | Training effort; too-similar photos get copied back[2] |
| 2. Give a reference image | Feed a reference image every time | No training, easy | Tends to stop at “resembles”[7] |
| 3. Make them together | Generate several at once, letting them refer to each other | No training; suits series | Mixing styles breaks consistency[10] |
| 4. Hold it in video | Keep referring to the first frame or a reference | Holds through motion | Drift accumulates as it gets longer[14][15] |
The paid part goes through the four one by one: what the research does, its limits, and how that translates into a creator's practice. It also covers how “looks like the same person” is measured, and a worked example we made and checked on EGAKU (7 figures).
La suite de ce guide est payante
Dans la partie payante
- 1. Teaching it — a keyword, and an add-on
- 2. Giving a reference image — easy, but “resembles”
- 3. Making them together — series and storyboards
- 4. Video — why extending breaks things
- 5. How “looks like the same person” is measured
- 6. A worked example on EGAKU — made and checked
- 7. Checklist
- A note on the research organisations
- References
Achat unique, lisible tout de suite et conservé sur votre compte.
Confirmation de votre achat…