The complete flow for creating a production-ready character — new design, full turnaround sheets, and targeted fixes — running entirely on a local GPU. Every step below is a tab in the Studio, and every claim was earned by experiment.
Point the Train tab at a folder of 15–40 existing characters from one game. A local VLM writes
one caption per image using a closed trait vocabulary — the same words the Builder's dropdowns use
(hero / villain, tiny / tall…), fixed before captioning, because sliders
only work if the training words match them. One overnight run (~26s/step on an RTX 3060) produces a
style LoRA that appears in the Builder automatically.
Pick traits from structured axes (gender, age, build, alignment, mind, role), add a free-text line, and anchor the style: a single reference character at low reference-pull (copies the drawing language without cloning the character), the trained LoRA, or both. Generate 2–4 candidates and pick.
The winner becomes a reference. Its declaration is a table, not a prompt: each feature says which views it is visible from, and — critically — what is there instead where it is hidden. The model cannot subtract; absence must be stated (“the backs of the trousers are plain — no stripe”). Style is declared explicitly too, or a glossy 3D reference comes back as flat 2D.
Each view renders separately (~80s) and passes gates: geometry in plain code (shared height, ground line, stance), and content checks by a local VLM judge — calibrated on known answers before it earned a veto. Failed views re-roll automatically; the designer accepts or rejects; layouts are assembled by code.
Local defects get local fixes: instruction edit (change one thing, keep the rest), head swap between characters, outpaint to grow a canvas, and identity-aware Character fix — paint a region, describe what belongs there, the character stays intact.
| Layer | Who does it | Why |
|---|---|---|
| Geometry, colour, layout | plain code | exact and free — never ask a model to count pixels |
| Where is a feature | grounding VLM (boxes) | only after its phrase is validated on the reference |
| Counts & verdicts | local VLM judge | chosen by calibration on known answers (10/10) |
| Taste & truth | the designer | the loop's corrections become permanent system rules |
Hardware: one RTX 3060 12GB. Generation ~80s per view; LoRA training overnight via WSL2. No cloud, no per-image cost, no licence fees.