PicPrompt · Image Prompt Decompiler
Decomposing reference images into editable visual fields and four generation formats
Multimodal AI Product Case / 17
PicPrompt decomposes a reference image into eight editable visual fields, then reformats them into ready-to-use prompts for different image models.
Users can drop, paste, or enter an image URL and use a chosen multimodal provider to extract subject, style, lighting, composition, camera, color, mood, and details, then format for Midjourney, Seedream, Nano Banana, or GPT-Image-2. The product has no proprietary backend; keys and archives stay in the browser.
The public site is live with single-image and up-to-ten batch extraction.
Supports OpenAI, Anthropic, Gemini, Doubao, and GLM providers.
Structured fields are editable, with conversational refinement and four target formats.
Keys, thumbnails, and results are stored locally; during extraction the image is sent directly from the browser to the selected model provider.
Reverse Prompting Should Not Be a Text Black Box
A single generated paragraph is difficult to edit. The product first outputs eight visual fields plus negative items so a user can change lighting, camera, or subject without disrupting the rest.
Different models need different language and parameters. A shared intermediate structure lets one extraction become Midjourney syntax, natural-language Seedream, or other model formats.
Local-first Does Not Mean No Provider Call
PicPrompt has no own backend. API keys remain in localStorage and archives in IndexedDB, reducing the platform's sensitive-data footprint.
During visual extraction, the image still goes directly to the chosen model API. Help content makes this explicit and advises against saving keys on others' devices and recommends regular exports.
PROCESS
Key Decisions
Structure first, format second
One intermediate object supports multiple model outputs and partial editing.
Bring your own provider and key
Avoid a central proxy backend, reducing the platform's custody of credentials and images.
Make the archive portable
Local IndexedDB supports privacy and speed; JSON backup mitigates browser-storage loss.
Role & Collaboration
I led the product loop, structured fields, provider and formatter architecture, privacy language, batch extraction, and local archive experience.
AI supported implementation and testing; I owned data boundaries, provider differences, credential guidance, and release acceptance.
Validation, Outcome & Reflection
The public site is live with the main upload, settings, extraction, editing, formatting, and archive interfaces.
Extraction quality and cost depend on the user's provider and key; the product does not promise exact reproduction for every image.
A multi-model product must separate model capability from product structure: providers may change, but user understanding and control must remain.
