lov-riso-portrait

lov-riso-portrait — Riso portrait

Redraw a single-person photo directly into a Riso illustration suitable for use as an avatar: the person remains recognizable at a glance, and the image uses limited color inks, halftone dots, paper grain, thick lines, and slight misregistration, rather than applying a retro filter over the original photo.

Triggers

Activate when

• The user says "make this photo into a Riso portrait", "generate a stencil-print-style avatar", or "make a two-color halftone avatar".
• The user asks to preserve the person's identity, pose, and clothing characteristics while changing it into a Riso, Risograph, or stencil-print visual.
• The user asks to “create a Riso portrait”, “turn this photo into a Risograph avatar”, or “make an identity-preserving Riso profile picture”.

Do not activate when

• The user only wants skin smoothing, brightening, background replacement, or a professional headshot; use lov-professional-portrait。
• The user wants to distill an arbitrary visual style from multiple sample images, rather than generate a Riso portrait; use the style analysis capability.
• The user wants to create a poster, infographic, or communication material with text; use the corresponding design Skill.
• The current runtime cannot call gpt-image-2 for image editing; Canvas, Sharp, CSS, two-color mapping, or ordinary filters must not be used to disguise it as an equivalent result.

User Profile (cross-session)

On each run, read skill.yaml declared by user-profile/v1, and resolve color scheme and cropping preferences according to the current request, project context, skills.lov-riso-portrait.records, shared Preferences, and safe defaults. Only long-term preferences directly declared by the user are saved via scripts/profile_store.py record --confirm ; photos, personal identity information, temporary Prompts, access credentials, and generation results must not be written to the Profile.

Skill Group Composition

Before running, read references/skill-composition.md. This Skill exclusively owns the result of "identity-faithful Riso avatar"; adjacent capabilities are only optionally handed off via the original image or the completed PNG, and do not constitute a hidden dependency.

Implementation paths are alternatives

Read references/implementation-options.md. Programmatic filters, generative model redrawing, and AI parameter analysis combined with deterministic rendering are three parallel approaches, not three steps that must be executed in sequence. The default delivery of this Skill is still gpt-image-2 direct redrawing, because it best reorganizes the person, color blocks, lines, and halftone dots; only when the user asks to research the implementation method, pursue parameter controllability, or require batch consistency should other approaches be discussed, and a two-color filter must not be passed off as the default Riso portrait result.

Workflow (MANDATORY)

Step 0: Resolve root and runtime

1. Resolve the Skill root directory, read skill.yaml、references/riso-art-direction.md、 references/quality-gate.md、references/implementation-options.md and references/skill-composition.md。
2. confirm the runtime can view the input image and can use gpt-image-2 to perform genuine raster image editing.
3. If the model is unavailable, stop explicitly and state that gpt-image-2is missing; do not silently substitute a programmatic filter, another model, or deliver only a Prompt.

Step 1: Inspect the source as the identity reference

• Treat the photo uploaded by the user as the sole source of truth for the editing target and the person's identity.
• Confirm the input is the original photo selected by the user or a verified authoritative original image, not an existing Riso result, a chat thumbnail, or an incorrectly cropped intermediate product. When there are multiple candidate images, record the source and role of each one, and do not infer the original image from a finished product.
• Check the face shape, facial feature proportions, hairstyle outline, expression, gaze, pose, clothing, accessories, held objects, and fingers.
• Identify whether the image is suitable for avatar cropping: whether the face is clear, whether the top of the hair is complete, and whether there is whitespace in front of the gaze.
• Photos with multiple people are not processed by default; first ask the user to specify the main subject or provide a single-person photo.
• When the use is only head or avatar cropping, first compare recognizability after circular cropping, gaze clearance, background noise, and thumbnail contrast, then decide on the source image; do not rank only by which looks "best" as a whole.
• Do not publicly upload the photo, results, or personal information unless the user separately authorizes it.

Step 2: Resolve the smallest sufficient brief

Default output is a 1:1 close-up portrait, suitable for circular cropping. Use deep charcoal black, teal green, vermilion, and warm paper color; retain paper grain, halftone dots, thick lines, not entirely even ink edges, and slight misregistration. The current request may override the color scheme, shot size, or background, but must not silently change the person's identity, age, pose, clothing, or accessories.

Ask only one question when missing information would clearly change the person or the avatar use; ordinary color schemes and cropping are filled in by safe defaults.

Step 3: Build the identity-locked edit prompt

Read references/riso-art-direction.md, and make the Prompt explicitly include:

1. style-transfer use case and 1:1 social profile avatar asset type;
2. the input image is the edit target and identity reference;
3. the person's identity, pose, gaze, clothing, and key objects remain unchanged;
4. Riso's limited color inks, halftone dots, paper, hand-drawn lines, and slight misregistration;
5. circular avatar safe zone, complete hairstyle outline, and clear space in front of the gaze;
6. no photographic feel, plastic skin, 3D, ordinary anime-fication, text, Logo, watermark, or decorative border.

The user's additions are only used as Creator note appended to the end of the template, and must not override the identity lock and factual constraints.

Step 4: Edit directly with gpt-image-2

• Pass the original image as the edit target to gpt-image-2, using high input fidelity and high-quality square image output.
• Let the model reorganize the line art, color blocks, halftone dots, and paper relationship itself; do not append grayscale separations, duotone mapping, halftone angles, noise, or misregistration filters after generation.
• Save non-destructively; it is recommended to use "original file name-riso-portrait-v1.png".
• In the first round, generate only one main version; accept it first, then decide whether local corrections are needed.

Step 5: Inspect at three scales

Check according to references/quality-gate.md :

1. Full image: whether the person, pose, clothing, and visual center of gravity are consistent with the original image;
2. Face close-up: whether the eyes, nose, lips, jaw, hairline, and expression still look like the person;
3. Chat list and circular cropping: whether it is still recognizable when shrunk, and whether the top of the hair, chin, and gaze are eaten by the edges.

Also count the visible fingers one by one, and check high-risk details such as earrings, glasses, necklaces, and microphones. A beautiful overall atmosphere cannot replace factual checking.

Step 6: Correct one fact at a time

• Identity drift: re-emphasize the face shape, facial feature proportions, hairstyle, and expression, and reduce stylistic freedom.
• Too much like a filter: require reorganizing the line art and color blocks, and prohibit retaining photographic continuous light and shadow.
• Incorrect fingers, accessories, or objects: fix only one clear fact, and repeat "all other person features, actions, and style remain unchanged".
• Common-sense errors: return to the source image and real-world facts to determine the correct count, structure, and spatial relationships, then make local corrections; do not fully trust the generative model, and do not redo an already accepted whole image because of one local error.
• Cropping unsuitable for avatar: adjust only the shot size and safe zone, do not redo the person.

Solve only one problem per round. After correction, re-run the three-scale check to avoid fixing a finger but breaking the face.

Step 7: Deliver

• Display the final PNG directly when runtime support allows, and report the save path.
• Briefly explain the color scheme, cropping, and intentional corrections used; at the same time confirm that identity, pose, and key objects have been preserved.
• When the user requests it, additionally create original/result comparison or iteration process images, but do not bake labels, watermarks, or marketing text into the avatar.
• Process images should only show stages that actually occurred and are visually distinct; remove duplicate results, do not fabricate deformed frames, and make local issues clearly visible in uniformly sized small images.

Completion Criteria

• Still recognizable as the same person when scaled down to avatar size.
• The image is a redesigned Riso illustration, not a grainy photo or a two-color filter.
• No factual errors in facial features, visible fingers, accessories, clothing, pose, or gaze.
• Circular cropping will not cut off the top of the hair, the chin, or the gaze direction.
• The original image remains unchanged, and the final PNG path and model boundaries have been reported.

Dependencies

• An Agent runtime that can view local or uploaded images
• Can perform image editing gpt-image-2
• The generation service may incur model invocation costs; the Skill itself is free