Image — characters become scenesLiloImage
One character, the same face from every angle
An in-house anime image model that draws character scenes
LiloImage is the in-house anime image model that draws character scenes for ONE AI Labs Hub. Every character has a bible (appearance and wardrobe notes) and a reference sheet; when a writer submits a one-line scene in Korean, five stages turn it into a 1664×2216 portrait frame. ① A vision-language model crops a single front-facing full-body cut from the reference sheet, ② the same model rewrites the scene into a 90–150-word English prompt, with the camera sentence taken from a fixed table rather than the model. ③ The composition is drawn first at about 1MP, ④ the face, hair color and outfit are transferred from two references, and ⑤ an anime-specific upscaler finishes the frame. Everything, prompt rewriting included, runs inside a single GPU with no external API calls — about 40 seconds and roughly $0.02 per scene. Every video on this page is real output used in the service. It is an internal-only model — weights and API are not publicly available.
Key features
- Identity transfer from two references — face, hair color and outfit hold across cuts
- Hair color and outfit come from the appearance line in the character bible — the model never guesses them from an image
- Camera sentences come from a fixed table — high angle, low angle, top-down or POV, exactly per scene
- Scene-type camera adapters (first-person, height, back view) are applied at the composition stage to lock the angle
- Korean scene description → 90–150-word English prompt, rewritten on the same GPU (zero external LLM calls)
- About 40 seconds per scene — one line of Korean becomes a high-resolution 1664×2216 portrait frame
- One model draws every character scene, so the anime art style stays consistent across the whole service
- Runs end to end on in-house GPUs with no external API — character settings and scene scripts never leave the company
- The upscaling stage never invents new detail, so line art and coloring stay crisp and true to the composition
- A drawn scene becomes a video's first frame, keeping the face and outfit steady to the end of a 5-second clip
What This Reference Proves
~40 seconds for a 1664×2216 frame
All five stages — composition, identity transfer and upscaling included — finish inside a single GPU. Prompt rewriting runs on the same GPU, so there is no waiting on external APIs: a one-line Korean scene becomes a 1664×2216 portrait frame in about 40 seconds.
Same face, whatever the scene
The face, hair color and outfit are transferred from two references, so the character bible's look holds even as the scene and composition change. Used as the first frame, the image keeps that face and outfit through to the end of a 5-second video.
Composition 4.8 → 7.0, face 8.4 → 8.9
Across 28 scene-opening composition directives, scored before and after on the same rubric: composition rose from 4.8 to 7.0, anatomy from 6.1 to 7.0, face from 8.4 to 8.9 and anime style from 8.2 to 8.6.
One cut as the reference
Feeding a multi-pose reference sheet as the character reference made 12 out of 12 images copy the sheet layout. So a single front-facing full-body cut is cropped first, and only that cut is used.
Compose small, finish large
Identity transfer duplicates or smears the character above 2MP. So composition and identity are settled at about 1MP, and a non-diffusion anime upscaler takes the frame to 1664×2216.