HANIS Marketing Image Design Expert

Problem

Marketing image generation often fails on small but important details: Chinese copy may be missing or corrupted, styles drift during edits, reference-image models repaint people and text, and different image models return URLs or base64 in different ways. This skill turns visual marketing into a concrete workflow: choose the right model, size, style, and prompt structure before generating.

How it works

• Model routing: selects among custom-textmodel-gt, gpt-image-2, gpt-image-1, grok-imagine, and other models based on intent, such as text-to-image, image editing, transparent background, fast preview, or high-quality final output.
• Scene and style matching: includes 20 marketing scenarios and 15 design styles. Product marketing defaults to minimalism, city tourism to magazine editorial, recruitment posters to tech futuristic, and so on, with size suggestions such as 1024x1024, 1536x1024, and 1024x1536.
• Prompt engineering: builds structured English prompts with subject elements, background, color palette, style positioning, and explicit Chinese copy labels. It appends no watermark, no logo, clean image for cleaner delivery.
• Edit constraints: separates style transfer with content preservation from creative edits such as background replacement, adding elements, or removing watermarks. The documentation notes that custom-imagemodel-gt may regenerate all content, so gpt-image-2 is preferred when people, text, and layout must stay unchanged.
• Delivery checks: reviews Chinese copy, composition, color, and watermark absence, then saves the result whether returned as a URL or base64.

Limitations

• It defaults to Chinese copy, making it best for domestic marketing assets, official account covers, e-commerce images, and campaign posters.
• Transparent backgrounds are supported only by gpt-image-1; Tencent custom models support only limited preset sizes, so larger or custom aspect ratios may require another model.
• Reference images can be provided as URLs, base64 data URLs, or mask images; local large files are better passed as base64 to avoid download timeouts.
• Some models return base64 and require decoding, while URL responses expire, so files should be downloaded promptly.