121140828-ship-it/stylize-photo-lofi-doodle

Stylize Photo as Lo-fi Doodle

Create a mixed-media image in which the environment remains photographic and only the chosen person becomes a modest, imperfect drawing. Favor selective human decisions over uniform filters, polished manga rendering, or synthetic texture.

Workflow

1. Identify the edit target and the intended person.
   • If no image is available, ask the user to upload one.
   • If one foreground person is clearly dominant, proceed without asking.
   • If several people are similarly prominent, ask which person or people to redraw.
   • Treat printed people, posters, screens, reflections, statues, and distant bystanders as background unless the user selects them.
2. Inspect the target image before editing.
   • Use view_image for a local file.
   • Note the crop, camera angle, perspective, lighting, occlusions, out-of-frame limbs, and major background landmarks.
3. Build an edit prompt from references/prompt-template.md.
   • Label every image as edit target, style reference, or supporting input.
   • Describe visible facts from the target rather than inventing details.
   • Convert visual references into generic traits. Do not require a named artist's exact style.
4. Use the built-in image-generation editing tool by default.
   • For a local target, pass its path after inspecting it.
   • For a conversation-only attachment, include the smallest number of recent images that contains the target.
   • Produce one balanced draft unless the user requests variants.
5. Inspect the result against the validation checklist.
6. If needed, make one targeted revision that changes only the failed dimension. Do not rewrite the whole art direction after every comment.
7. Return the generated image and its saved path. Preserve the source file and never overwrite it by default.

Lock the Photograph

Preserve the original crop, aspect ratio, perspective, camera angle, lighting, color cast, and all non-target elements as real photography. Explicitly name important background objects in the prompt so they are not redrawn.

Do not:

• stylize the entire image;
• replace or beautify the location;
• redraw distant people or printed characters accidentally;
• add a sticker border, drop shadow, halo, title, UI, logo, or watermark;
• extend cropped hands or limbs beyond the original frame.

Generative edits may not preserve the background pixel for pixel. Describe this limitation only when exact pixel identity matters to the user's use case.

Lock the Subject

Preserve the person's placement, scale, silhouette, pose, gaze, expression, crop, overlaps, clothing structure, hairstyle, accessories, and held objects. Retain recognition through these identity anchors rather than by rendering a photorealistic face.

For difficult anatomy, prioritize the source pose and original cropping. Do not invent hidden fingers, hands, feet, or body parts.

Default Art Direction

Use a quick notebook drawing made with a slightly dry ballpoint pen and two or three fading markers.

• Lead with uneven, economical linework.
• Make some contours dark and retraced; leave others faint, broken, incomplete, or slightly overshot.
• Include a few searching lines and corrections, but keep large quiet blank regions.
• Keep facial features readable yet mildly asymmetric and unpolished.
• Use loose hair clumps and a limited number of unresolved strands.
• Render skin mainly as warm paper-white with almost no modeling.
• Describe clothing with imperfect seams, folds, patterns, and sparse scribbles rather than solid digital fills.
• Apply muted color selectively. Allow fills to stop short, cross a contour, or leave accidental gaps.
• Vary detail density: concentrate lines around the face and identity anchors; simplify secondary areas.
• Keep the character visibly flat against the detailed photographic environment.

Preserve the source palette when it is distinctive, but compress it to approximately two to four muted drawing colors plus dark ink and paper-white.

Avoid Synthetic AI Signals

Do not simulate hand drawing by adding the same noise everywhere. Avoid:

• uniform crosshatching across the whole person;
• perfectly continuous vector-like contours;
• symmetric facial features and glossy anime eyes;
• evenly spaced hair strands or clothing patterns;
• fully resolved anatomy and consistent finish in every area;
• smooth gradients, airbrush skin, cel shading, 3D volume, or painterly polish;
• random grunge that does not follow deliberate line decisions.

The target is not merely minimalist. Keep enough uneven observation to describe the person, clothing, and pose. Favor incomplete specificity over blank iconography.

Revision Controls

Interpret feedback narrowly:

• More lo-fi / less AI: reduce uniform finish; add selective omission, mild asymmetry, broken contours, and imperfect color registration. Do not add dense line noise everywhere.
• More line character: replace smooth fills with a few purposeful contour, seam, fold, and fray lines; vary pressure and continuity.
• Too minimal: restore moderate facial, hair, clothing, and accessory information without restoring polished rendering.
• Too polished: simplify beauty rendering and remove gradients, glossy highlights, perfect curves, and evenly finished details.
• Background changed: repeat the background invariants and list the affected photographic objects explicitly.
• Identity drift: reinforce hairstyle, glasses, clothing, accessories, expression, pose, and silhouette before adding facial detail.

Validation Checklist

Before returning the image, verify:

• Only the selected person is illustrated.
• The background still reads as an ordinary photograph.
• Crop, pose, silhouette, and important overlaps match the source.
• Clothing, hairstyle, accessories, and held objects remain recognizable.
• Line weight, continuity, and detail density are visibly inconsistent in a human way.
• Blank space and imperfect color registration are present.
• The result is neither polished manga nor ultra-minimal icon art.
• No extra body parts, text, UI, logos, or watermarks were introduced.