How to Save on AI Credits
Coloring Book Engine is local-first, so wasted credits require deliberate effort: most tasks can run on Local AI for free, and credits are only ever spent where you choose. Here’s the full picture of where money goes — and how to spend less of it.
Where Credits Go (and Don’t)
Section titled “Where Credits Go (and Don’t)”| Spends credits | Free (local) |
|---|---|
| Sketch generation & edits | Sketch Process and Vectorize |
| Coloring via API (& edits) | Coloring via Local AI, palette/adjust, Finalize |
| Matter & cover AI generation | Book export, validation, 3D preview |
| Plan chat & one-time image analysis (cheap — text-scale) | The entire Assets & Videos pipeline |
The Habits
Section titled “The Habits”- Run Rest is the default. It only fills empty slots — finished pages are never re-billed. Reach for Run All only when you genuinely want to redo everything (and it will warn you and back up the old images first).
- Draft on Cost, finish on Quality. Iterate your prompts and plan on the cheap tier; flip the toggle for the final pass on pages that earned it.
- Fix pages one at a time. A page you dislike costs one credit to redo — click that card, don’t re-run the batch.
- Refine instead of regenerate. The edit dialog’s “Additional requirements” mode adjusts the existing image; it’s more predictable than a fresh roll of the dice, so you converge in fewer attempts.
- Front-load the Plan stage. Every hour in Plan (free) sharpens prompts that would otherwise burn credits on mediocre generations. Mention only the characters a page needs — each
{Name}reference adds a little to the request. - Let Local AI do the work. If your device has a capable GPU (~8 GB+ VRAM), route Sketch, Coloring, and Book to Local AI — it generates on-device for zero credits, and it’s the app default. Keep the API for the final, sellable pass, where line quality earns its small cost.
The Built-in Protections
Section titled “The Built-in Protections”Even if you ignore all of the above: finished pages are never silently regenerated, Run All always warns and backs up, two consecutive failures pause the queue instead of burning through retries, and generation is paced to respect provider rate limits.