wenyachen/ai-wedding-photo-skill/ai-wedding-photo

AI Wedding Photo

Guide the user from style selection to model-ready prompts and optional image generation.

Workflow

1. Check whether the user has already supplied clear portrait references for both people and any pet that must appear.
2. Present all nine styles from references/styles.md. Show assets/style-previews/style-menu.png first, then show individual previews when the user wants a closer comparison. Use Markdown image syntax. If the interface cannot display local images, provide the style names and concise descriptions.
3. Ask the user to choose one style or request a recommendation. If recommending, ask only for the desired setting, mood, season, and whether pets or a story timeline should appear.
4. After the style is selected, ask which image model or platform will be used. Offer: OpenAI/ChatGPT Images, Midjourney, Nano Banana, 即梦, 豆包, or another named model.
5. Read the selected style entry in references/styles.md, the directing rules in references/shot-direction.md, and the model rules in references/model-adapters.md.
6. Collect only missing essentials:

• Reference portraits for identity preservation
• Number and identity of people or pets
• Preferred aspect ratio or output count
• Any clothing, location, pose, or cultural requirements

1. Produce:

• A concise creative brief
• One master prompt
• A directing table for 6-9 distinct frames
• One complete model-adapted prompt per frame
• A negative prompt only when the chosen model supports it

1. Run the diversity check in references/shot-direction.md. Rewrite duplicated frames before generating.
2. If running in Codex and an image-generation tool is available, ask for missing reference images, then generate directly. Generate every frame as a separate image request using its own prompt; never ask one request to create the whole set or a contact sheet. Generate one representative frame first unless the user explicitly requests a full set. Preserve identity and iterate from user feedback.
3. If direct generation is unavailable, give the frame prompts in separate copy-ready code blocks and concise usage notes for the selected model.

Interaction Rules

• Show the style menu before asking about the model unless the user already chose a style.
• Do not ask every possible preference at once. Ask the minimum needed for the next step.
• Treat uploaded faces as identity references, not as style references.
• Preserve ethnicity, age range, facial structure, glasses, hairstyle, and distinguishing features unless the user requests changes.
• Keep skin texture natural. Avoid automatic face slimming, whitening, age regression, or body reshaping.
• For multi-image sets, treat the reference collage as a directing target: reproduce its variety of interaction, expression, camera distance, composition, and environmental storytelling without copying a single image literally.
• Do not accept superficial variation such as changing hand position while keeping the same standing pose, camera distance, facial expression, and background.
• Use no more than two formal side-by-side standing portraits in one set.
• Keep subjects looking directly at the camera in no more than two frames. Favor mutual eye contact, laughter, movement, quiet glances, and absorbed interaction.
• Give each frame one clear emotional beat and one physical action. Avoid vague directions such as "pose naturally" or "look intimate."
• Do not promise exact identity reproduction. State briefly that results depend on the selected model and reference quality.
• Never identify a real person from an image.

Asset Map

Preview paths and style IDs are listed in references/styles.md.

Use scripts/make_contact_sheet.py when a style is supplied as nine separate images and needs a square 3x3 preview:

python3 scripts/make_contact_sheet.py INPUT_DIR OUTPUT.png