← IndexCase 03 / 10Image Model2026

LiloImage

An in-house image pipeline that draws the same character in every scene

A one-line scene a writer types in Korean goes through five stages and becomes a 1664×2216 portrait shot. Everything, prompt rewriting included, runs on one GPU with no external API calls. About 40 seconds and about $0.02 per scene. Every video below was actually used in production.

Visit LiloImageai-labs.to.nexus/ko/apps/lilo-image
Project Detail
Role
Model · pipeline design · engineering
Contribution
100%
Year
2026
Platform
Internal image model · single GPU
Background

General image models changed the character's face every shot

Ask a general image model to redraw the same character and the face, hair color and outfit change from shot to shot. For a serialized character service, that is fatal. I redesigned the pipeline to remove, one by one, every place the model had to guess.

Approach3 Points
  1. Point 01

    Identity transfer from two references

    First we crop one front-facing full-body shot from the reference sheet. When we fed the whole sheet, 12 out of 12 outputs copied the sheet layout too.

  2. Point 02

    Looks from canon, camera from a fixed table

    Hair color and outfit come from the appearance lines in the character bible, which is canon. Camera lines come from a fixed table instead of the model, so high angle, low angle and POV are exact.

  3. Point 03

    Compose small, finish large

    Identity transfer above 2MP duplicates or smears the subject. We lock composition and identity at about 1MP, then scale to 1664×2216 with an anime upscaler that has no diffusion step.

Process5 Stages
  1. 01

    Crop

    A VLM extracts one front full-body shot from the reference sheet

  2. 02

    Rewrite

    Korean scene → 90–150 word English prompt (same GPU)

  3. 03

    Compose

    About 1MP composition + per-scene-type camera adapter

  4. 04

    Transfer

    Face, hair color and outfit from two references

  5. 05

    Upscale

    Finished at 1664×2216 with an anime upscaler

Result

Results

Per scene1664×2216 portrait shot
Cost per imagePrevious pay-per-use engine $0.09
External API callsRuns entirely on one GPU
Scenes used in productionNEXUS AI Studio logs
Before and after 28 direction cues (same rubric, out of 10)Baseline · beforeAfter
View as table
Before and after 28 direction cues (same rubric, out of 10): before and after
MetricBeforeAfter
Composition4.87.0
Anatomy6.17.0
Face8.48.9
Anime style8.28.6
10 Items
  • 01 — LiloImage generated shot
  • 02 — LiloImage generated shot
  • 03 — LiloImage generated shot
  • 04 — LiloImage generated shot
  • 05 — LiloImage generated shot
  • 06 — LiloImage generated shot
  • 07 — LiloImage generated shot
  • 08 — LiloImage generated shot
  • LiloImage key visual
    LiloImage key visual
  • LiloImage character cast sheet
    LiloImage character cast sheet
  • Tech referenceai-labs.to.nexus/ko/apps/lilo-image
  • Tech blogai-labs.to.nexus/ko/blog/nexus-talk-and-lilo-image-models
Next Case — 04

NEXUS AI Studio →

A character IP engine that turns one seed line into world, scenes, video and chat

GY.KIM