Zero Python. Zero servers. Kites detects text balloons, cleans foreign lettering with AI inpainting, and typesets authentic comic dialogue right inside the artwork.
Experience clean polygon inpainting with zero blurred boxes, balanced line-breaks, and authentic comic lettering.
👈 Drag slider to reveal AI inpainting & typography in real-time 👉
No white squares. Erases strictly inside character strokes.
From subtle in-page hover triggers to an interactive full-screen studio editor.
Hover over any comic panel on Twitter/X, Pixiv, or manga reader sites. An unobtrusive floating translation button appears seamlessly using modern CSS Anchor Positioning. Click it, and the panel translates right inside the artwork.
Switch target languages, toggle reading modes, or trigger on-demand viewport translations in one click.
Configure PaddleOCR models (v3–v6), inpainting algorithms (LaMa / AOT-GAN), WebGPU acceleration, and custom API keys.
Inspect clean background plates, resize speech balloons with 8-point handles, and export 1:1 lossless PNGs.
Kites eliminates heavy Python dependencies and expensive cloud servers by running optimized neural networks directly in Chromium via WebGPU.
Runs ONNX Runtime Web directly on your graphics card with seamless WASM fallback. No local CUDA or Python installation needed.
Reconstructs reading order and merges fragmented text lines into unified dialogue balloons using graph-theoretic minimum spanning trees.
LaMa Manga and AOT-GAN neural inpainters operate on localized dynamic-bucketed crops at 1:1 scale, avoiding ugly gray blur boxes.
Automatic failover across on-device WebLLM, our zero-cost Cloudflare community pool, Google Translate, or your private API keys.
Your images never touch our servers. All OCR detections, clean backgrounds, and translations are stored locally in IndexedDB with 7-day TTL pruning.
Licensed under GPLv3. Auditable source code, zero hidden analytics, zero telemetry beacons, and zero monthly subscriptions.
