<h1 id="manga-page-character-illustrations">Manga page → character illustrations</h1>Detects every character on a manga page, crops each one and redraws it with Qwen Image 2.1 as a complete illustration with a background, extending the framing naturally and completing what the panel hides.<h2 id="how-to-use">How to use</h2>1. Manga page: load the page (or a single panel).2. Detection - full page (true): best when characters break panel borders or fill the page. - per panel (false): best for small characters, each one inside its own panel.3. Characters to find: SAM3 text prompt, max 32 tokens. <code>:N</code> is the maximum detections per prompt and per image (per panel in per-panel mode). - <code>anime character:8</code> → everyone - <code>anime character:1</code> → the main one in each panel - <code>blonde girl:1, girl with long black hair:1</code> → specific characters4. Detection threshold: raise it (0.6–0.7) to drop background people and slivers; lower it if small characters are missed.5. Redraw: characters (true) redraws each character; whole panels (false) redraws each panel.6. Redraw style: pick <code>color</code> or <code>black_and_white</code>. The matching Base prompt always asks for a full background and a natural, wider framing (not forced full body). Refer to the crop as <code>&lt;image1&gt;</code>.7. Prompt upsampler (off by default): off keeps the redraw closest to the original drawing, with softer colors. On, Qwen3-VL 8B looks at each crop and rewrites the base prompt for it: viewpoint (front / profile / from behind), a color for every element that matches the manga values (solid black hair stays dark, white hair becomes blonde), knee-up or full-body framing, clothing for the hidden lower body, a specific background scene, and removal of sound effects. Its output is shown in Upsampled prompt inside Redraw; edit Upsampler system prompt there to change the rules. Off uses the base prompt as is: faster and more faithful, with weaker colors, and sound effects or back views are handled less reliably. Character notes: optional, one line per character, starting with how to recognize them, e.g. <code>Girl with long black hair: maid, black dress, white apron</code> or <code>Blonde boy with glasses: student, gakuran uniform</code>. The upsampler applies the line that matches each crop (role, colors, clothing) and ignores the rest; leave it empty to let it decide. With the upsampler off, the notes are appended to the base prompt as is, so keep them to the character(s) on this page; with several characters the model may mix their attributes. Output canvas: custom size (true) computes the canvas from each crop: crop aspect ratio × Canvas expand (width ×, height ×) at Canvas megapixels, rounded to multiples of 32. Expand 1 / 1 keeps the crop's exact aspect; e.g. height × 1.3 leaves 30 % more room below to complete the figure. Keep megapixels around 1 (Qwen Image 2.1 supports up to ~4). crop size (false) uses the crop's own size, better for whole panels.8. Run. Crops are saved to <code>output/manga_chars/crop_*</code>, illustrations to `output/manga_chars/

Files with SHA256: 421841b42a2a08f455280aa75468d88572c08db047f1d4030750e3665aafd5e7