Live
Image Prompt Studio
Keep what works. Record where.
Image Prompt Studio treats prompting as both a data-management problem and a repetitive production task. It lets a user keep reusable descriptions and references for identity, objects, locations, lighting, image grading and photographic or painterly style, together with their own observations about language that helped or failed with a particular model. A deterministic composer then assembles the selected material into a structured, reviewable prompt for use in an external GenAI image tool. Some image generation tools now separate characters, poses, objects and backgrounds as reference images for the same reason: consistency depends on keeping those parts distinct.
This is a working example of an approach, not a packaged product to transplant unchanged. Another organisation would start from its own workflow, constraints and evidence.
The observed problem
Designed for People doing careful, repeated image work with GenAI tools — restorers, archivists, families and creative technologists
Careful image work depends on iteration. A prompt may need a more precise preservation instruction, a different keyword or sometimes less language before it reaches a useful balance. What helps one model may fail in another, so the references, wording, model and result need to be recorded together rather than presented as universal prompt engineering.
Working demonstration
- ✓Descriptor library: identity, locations, poses, objects, outfits, lighting, grading and visual styles
- ✓Personal playbook: language, outcomes and quirks recorded per model
- ✓Guided composition with conflict detection and deterministic compilation
- ✓Three worked examples with cited research and structural child-safety rules
Still to explore
- 01Gather feedback on the descriptor and playbook method from real use
- 02Grow the reference library beyond the three worked examples
- 03Evaluate a provider-execution round-trip demo with recorded provenance
Demonstration in brief
“Image Prompt Studio began with a hard problem from my photo-restoration research: generative tools can polish an image while quietly changing the person.
The answer wasn't one perfect prompt or a definitive database of good prompt engineering. It was a way for anyone to keep their own reusable references, styles and observations about what a particular model appeared to understand.
A model can get you to a plausible image quickly. Keeping the same character, light and style across a set of images is where human direction comes in, and that direction has to be kept somewhere it can be reused.
Small wording changes matter. Too little direction can lose identity or intent; too much can over-constrain the image. A clumsy instruction for an anthropomorphic mouse once produced both the character performing the required action and an unwanted computer mouse on the table — a useful record of how that model interpreted the word.
That also turns a repetitive task into a controlled workflow. The composer combines selected reference data and model-specific language, surfaces conflicts, and deterministically produces one structured prompt for a person to review and use in an external GenAI image tool.
The demo does not generate an image or promise fidelity. It keeps the inputs and constraints visible so the human can refine the prompt and judge the result. It runs entirely in your browser, and nothing uploads.”