Colorize Line Art from a Reference Drawing

Local color is the base identity of an object before scene lighting changes its appearance. Write it by object: dark auburn hair, cream cotton shirt, desaturated teal bag, warm gray wall, brushed steel buckle. Avoid a loose string such as “warm cozy cinematic colors,” because it does not say which region owns which hue. Three to six main colors are usually enough for a first test. If the character already has an approved palette, use the same object-to-color mapping and keep accent colors small.

Flat colors are not an inferior result. They are a diagnostic stage. You can compare whether skin, hair, costume pieces, props, and background are separated correctly before shadows conceal boundary errors. Krita’s official flat-coloring guide likewise treats clean flats as a foundation and explains workflows for filling beneath line art and later splitting color regions. Anime Maker does not return those regions as layers, so a clear flat proposal is useful primarily as a visual key for manual rebuilding.

Do not encode every color as a hex value and expect calibrated reproduction. Display profiles, provider interpretation, surrounding light, and generated texture can shift appearance. Use ordinary color names plus relationships—“the scarf is darker and warmer than the shirt,” “the bag is the only saturated accent”—and keep a real swatch file outside the generator as the production source of truth.

After local colors are assigned, define where the main light comes from. “Soft window light from upper left” tells the result where lit planes and cast shadows should relate. Add a shadow family such as cool blue-gray or warm brown, but avoid requesting sunset, neon signs, moonlight, sparks, fog, rim light, and studio fill in one pass. Multiple competing sources can make the drawing harder to read and make color boundaries look inconsistent.

Value structure matters more than the number of effects. Squint at the result or view it in grayscale: can you still separate the face from the hair, hands from clothing, and the subject from the background? If not, change the value relationship before adding detail. A light object in shadow should not automatically become the same color as a dark object in light. Tell the prompt which distinction matters and review the actual image rather than assuming the words were followed.

Use cast shadows to explain form and contact. A hand should cast a plausible shadow on the object it holds; a figure standing on the floor needs grounding; a leaf in front of a face should affect the light. Generated images can create attractive but contradictory highlights. Treat them as proposals, then repair the lighting logic during paint-over.

Material direction works when it changes edges, highlight size, texture, and translucency. Matte cotton produces broad soft changes with little mirror reflection. Glazed ceramic can carry a compact highlight and a softened reflection. Brushed metal often has directional value bands. Worn leather shows creases and a controlled sheen. A translucent leaf may glow near a backlit edge. Name only materials already visible in the drawing so the generator does not invent props to satisfy a material list.

Keep material detail subordinate to line readability. Dense texture can cover thin hatching and redraw seams. If preserving ink is the priority, ask for restrained rendering, readable black contours, and simple broad material cues. Run a material-rich pass only after a flat-color pass has established the right object ownership. The four-panel editorial study on this page demonstrates that sequence: local colors, warm light, cool shadow, then a restrained surface pass.

Write the most important invariants in concrete terms: keep the face proportions, hand pose, clothing seams, watering-can silhouette, window grid, and major leaf contours recognizable. “Keep everything exactly the same” is not a control the provider can guarantee. Rank the features instead. If the face and hand gesture matter more than background hatching, say so and simplify the background request.

Compare source and result side by side at the same scale. Check the outer silhouette first, then landmarks, expression, finger count and grip, costume construction, props, and finally small texture. Overlaying the two files at reduced opacity in an editor will reveal drift quickly. Do not accept a beautiful result if it quietly changes the storytelling action or the identity of a commissioned character.

For a production illustration, place the original line art above the generated color proposal and test a Multiply blend or other suitable compositing method in your editor. Adobe’s official layers documentation explains that layers separate components so they can be changed without altering the rest of the image. That editable separation is precisely what the generated flattened image does not provide. Rebuild the color and lighting on dedicated layers if later revisions matter.

If color crosses a boundary, inspect the source for a gap, faint contour, or ambiguous overlap. If the face changes, reduce style language and repeat only the essential facial landmarks. If the palette drifts, map colors to named objects and remove extra mood words. If the image becomes too painterly, ask for clean flats, restrained cel shading, and readable ink contours. If the background takes over, lower its detail and contrast rather than adding more instructions to the character.

Change one variable per test and keep a short log: source version, prompt version, what remained correct, largest mismatch, and next change. A failed result does not justify replacing every color, light, material, and preservation instruction simultaneously. Controlled comparison turns generation into art direction; uncontrolled rewriting produces attractive samples with no reliable lesson.

Some failures are workflow failures rather than art failures. The upload requires sign-in and accepts JPG, PNG, or WebP within the server file limit. Reference tasks require a public URL and APIMart image configuration. A missing-reference error means no image was attached; a “not configured” response means the server cannot run this reference workflow. Repeated prompting cannot repair missing service configuration.

Use the generated image as a color key, paint-over base, or conversation artifact. For production, bring the original lines and chosen result into an illustration editor. Build separate layers for flats, shadows, light, effects, and corrections. Restore line weight and facial landmarks from the source, clean edge halos, correct hands and props, and remove any accidental marks. If a client needs revisions, this structure is far more useful than a single merged image.

Check the image at its delivery size. Thin lines that look elegant at 200% may disappear in a thumbnail; subtle shadow differences may collapse on a phone. Test grayscale, mirrored view, and a contrasting background. Add exact text, logos, and accessibility copy outside the generated art. If transparent output is needed, create a deliberate mask and inspect fine edges rather than assuming the background is removable.

A small editorial team has an approved black-and-white drawing of a botanist in a greenhouse. The deadline decision is not final rendering; it is whether the character should wear a deep green apron, whether terracotta can balance the foliage, and whether warm morning light keeps the face readable. The team uploads a cleaned copy, asks for those local colors and one upper-left light, and requests that the face, watering can, window grid, and large leaf shapes remain recognizable.

The returned image gives the art director a concrete color conversation. The team rejects a changed hand grip and an invented hanging plant, but keeps the sage–terracotta relationship and warm reflection on the watering can. A colorist then rebuilds the flats under the original line layer, paints a simpler shadow family, and corrects the face. In this case the generator saved exploration time without becoming the production source file. That is a realistic success criterion for reference-guided colorization.

Possessing an image file does not automatically grant permission to create and publish a derivative color version. Use your own drawing, commissioned work with suitable rights, licensed material, or a source for which you have explicit authorization. Keep the creator, source date, permission, project purpose, and uploaded version together. Do not remove a signature to disguise authorship, and do not use this page to recolor a known character for a commercial release without a rights review.

The U.S. Copyright Office’s 2025 Copyright and Artificial Intelligence, Part 2 report explains that copyrightability of AI-assisted output depends on human authorship and is assessed case by case in the United States. It discusses human-authored inputs, selection, arrangement, and modification rather than treating prompts alone as automatic authorship. Other jurisdictions differ. Preserve the original drawing, prompt, chosen output, repaint layers, and decision notes, and obtain qualified advice when the release is consequential.

Krita Manual — Colorize Mask: https://docs.krita.org/en/reference_manual/tools/colorize_mask.html. This documents key strokes, gap handling, edge detection, mask updates, conversion to a paint layer, and splitting colors. Krita Manual — Flat Coloring: https://docs.krita.org/en/tutorials/flat-coloring.html. This compares editor-based flatting methods and emphasizes correcting the final regions. These sources describe Krita, not Anime Maker, and are useful when exact editable regions are required after concept exploration.

Adobe Photoshop — Layers overview: https://helpx.adobe.com/uk/photoshop/desktop/create-manage-layers/get-started-layers/layers-overview.html. It explains why separate image, fill, adjustment, type, and shape layers support non-destructive changes. U.S. Copyright Office — Copyright and Artificial Intelligence: https://www.copyright.gov/ai/. Use these references to plan the hand-finishing and records that a single flattened generated image cannot supply.