OAD — an open document format for AI-assisted image editing

A creative document should outlast its application

Within ARtBase, I am developing the idea for OAD: an open, application-independent document format for future image-editing workflows, with AI able to work throughout the document rather than only on a final flattened image.

OAD is a working name. The specification and file extension are not final. The initial concept focuses on layered bitmap artwork; it is not a released standard or a finished editor.

Preserve the image and what makes it editable

Long-term preservation requires more than instructions that another application might render approximately. The artist-approved appearance needs to survive alongside the information needed for further editing.

The concept brings together:

• The approved visual result.
• Editable structure, including raster layers, groups, masks and compositing settings.
• Sources and transformation or effect recipes for derived content.
• Artist intent, semantic descriptions and relevant AI-generation records.
• Provenance explaining how content was created and changed.

These are proposed document-model elements, not a list of implemented features.

AI throughout the document

The goal is to make AI useful across the document’s structure: understanding individual layers and masks, relating derived elements to their sources, interpreting editing intent and assisting with changes in context.

Instead of treating an image document as an opaque file, an agent could work with a structured representation of its parts and relationships. Intensive AI assistance should still respect the artist’s decisions and preserve the accepted result.

Source, recipe and approved result

For derived content, the concept preserves three complementary things: the original source, the recipe describing what was done and the approved rendered result.

That matters because future software may implement an effect, blend mode or colour transformation differently. Keeping the approved pixels provides a visual reference even when a recipe cannot be reproduced exactly.

Declared facts and AI interpretations are different

Artist-declared intent, information extracted from a source document and AI-inferred descriptions should remain distinguishable. An AI interpretation can be useful without becoming an authoritative fact about the artwork.

This is part of making documents understandable to AI while preserving their provenance and the artist’s authorship.

Research direction

The early scope is layered raster artwork, with extensibility considered for later work. A format specification, migration tooling and an ARtBase reference implementation are possible components of the project.

The open questions include visual fidelity, editability, storage overhead, extraction coverage and the quality of semantic enrichment. Those need implementation and evaluation before stronger claims can be made.

Why I am exploring this

In My art should outlast the software I use to make it, I explain the personal motivation behind OAD: preserving my artwork, its editable structure and the freedom to continue with different tools. The article describes a preliminary design, with practical validation still ahead.