op7418/guizang-yingzao-skill

Turn your photos—buildings, streets, shops, artifacts, and local foods—into truly art-directed editorial posters—not just applying filters to photos, nor sticking titles on blank spaces.

Yingzao first reads the composition and identity of the photo, then selects a compatible dominant reference to design the Chinese display text and the spatial relationship between the image and text, and finally hands over the original image, reference, and sparse layout background image to the image model for a complete redraw. Subject processing, active color field, regional materials, and text occlusion all occur within the same visual world.

Starting in 30 seconds

After installation, directly say in the Agent that supports Skills and image generation:

You can also send the following message to your agent and let it complete the installation itself:

What will it do?

• Identify whether the photo is a front elevation, downward, diagonal, framed view, partial component, or environmental narrative, and fix only one real composition issue affecting the image.
• Protect identity anchors such as roof slopes, flying eaves, plaques, contours, and asymmetrics, preventing the real building from being “optimized” into another building.
• Before looking at the reference images, first propose four domains: subject, background, interaction, and layout actions, to avoid more references and more mediocre results.
• Select compatibility directions from the built-in recipe and real reference images, without mechanically linking dozens of style terms into prompts.
• Design and display width, center of gravity, stroke contrast, finishing strokes, character style, rhythm, and surface media; Small characters can use another conventional family to establish hierarchy.
• Extract, shift, partition the main semantic on the material, and make the real contour occluded, coagulate, or meshed with the title and color.
• When multiple images are “merged” into one, extract each subject and reconstruct a scene in shared perspective, light direction, contact shadow, and edge language; Only when group photos are explicitly required is the rectangle retained.
• After finishing the poster, you can proceed with a 3×3 video storyboard and prompts that can be handed directly to the video model.

Usage Examples

If the location, temple name, or historical information is uncertain, you can have the Agent search first; Yingzao only includes verified information, visible photos, or information you have confirmed in the poster.

Working Methods

High-stylization tasks consistently use three types of visual input: the original image correction handles identity, the dominant reference handles the complete visual mechanism, and the typesetting backing covers text range, common axis, action area, and occlusion relationships. The selected recipe is first compiled into the current photo action across four domains: subject, background, font, and interaction; These actions write the final ImageGen prompt while using tags like S1 / B1 / T1 to point to specific areas in the backing image. The pre-call handoff checklist checks whether each enabled token actually enters the model input, then locks the three images together with the full prompt; reinterpret titles use character slots instead of large standard font outlines, avoiding mistakenly treating Song font as the final font. The image model fuses them instead of re-rendering them according to web layout.

Generating Pre-Gate Control

Yingzao only performs deterministic checks before calling the image model:

• Photo adaptation, scrolling, and safe patch areas
• The boundary between facts and words
• Compatibility between dominant reference and target geometry
• Three visible design domains: subject, background, and interaction
• Complete delivery of recipe tokens to model actions and pad image markings
• Font CMAP overlay and fallback
• True literal, common axis, collision, no-pass zones, vertical order, and subject occlusion contracts