The pipeline explained
How OCR, correction, translation and inpainting chain together — statuses, credits and retries.
6 min read
How it works
Every page goes through three stages: OCR detects and reads text blocks, translation renders each block in your target languages, and inpainting cleans the artwork behind the text. Each stage is a background job — you can close the tab, everything keeps running.
The chapter view shows live progress per page: which stage is running, what completed, what failed. Failed stages can be retried individually without redoing the whole chapter.
Credits and costs
Each stage draws from your plan credits, exactly like the app: OCR costs 1 credit per page, translation depends on the model (free if you bring your own key), inpainting 1 credit. Renders and exports are free — you already paid for the pipeline that produced them.
Per-chapter model selection lets you trade speed for quality: a fast model for a first pass, a premium model for the chapters that matter.
- 1Open the chapter and watch the status chips fill in
- 2Click any failed stage to see the error
- 3Use Retry failed to re-run only what broke
- 4Switch the chapter model at any time before retranslating
BYOK (your own API key) makes translation free — only OCR and inpaint consume credits.
Tuning quality
OCR confidence thresholds, fuzzy deduplication and minimum text length are tunable in Settings → OCR. LLM post-correction can clean up OCR mistakes before translation. And a glossary keeps the terminology locked across the whole series — see the Consistency guide.
Video coming soon
Watch: following a chapter through the pipeline (3 min)
Adding pages after the fact
Chapters are not frozen once translated. Add pages to an already-translated chapter and only the new pages run through the pipeline — existing translations, story context and credits stay untouched.
And if a page ever gets stuck in “processing”, the pipeline recovers it on its own — nothing stays blocked waiting for a manual fix.
