character-consistency
Installation
SKILL.md
Character Consistency
Keep a recurring character on-model across generations and edits: proportions, colors, distinctive marks, finger counts, facial structure — every trait that makes the character that character.
Two principles govern everything in this skill:
- Simplest fix first. There is a ladder of approaches from a one-line guideline to a fine-tuned model. Always start at the lowest rung that plausibly solves the user's problem, and add a heavier layer only when the previous one demonstrably falls short and the new layer solves a clear, measurable problem.
- Aim for the smallest possible edit. Image models preserve identity best when the reference material is already close to the target and the prompt only describes the delta. Every approach on the ladder is a way of shrinking what the model has to invent.
The ladder of approaches
| # | Approach | Effort | What it is | Reference file |
|---|---|---|---|---|
| 1 | Corrective guidelines | Minutes | Short, targeted rules for what the model gets wrong; delta-only prompting | references/approach-1-guidelines.md |
| 2 | Reference images & atlas | An hour | Attach a master reference / character atlas; with a few references, pick the closest one per request | references/approach-2-references.md |
| 3 | Full dataset system | Hours–days | Curated 10–20 image dataset + tags + closest-reference selection per generation | references/approach-3-dataset-system.md |
| 4 | Pipelines & model routing | Hours | Two-stage identity/composition split, systematic model comparison, narrow routing rules | references/approach-4-pipelines-routing.md |
| 5 | Post-processing & training | Days+ | Refinement pass with an editing model; character LoRA or fine-tune | references/approach-5-postprocess-training.md |