An API restores layers to flattened advertisements

BRIA launches Delayer Agent, an API that breaks down a flattened advertisement into structured layers to modify its text, products, logos, and regulatory elements.

The source file is gone, the agency has shut down, or the campaign was archived as a simple JPEG. The ad remains visible, but every change requires rebuilding it. Delayer Agent aims to automate that process by separating a flattened image into editable elements.

BRIA presents its new tool as a way to put previously produced creative assets back into circulation. An imported ad is analyzed, broken apart, and described as a structured composition. The result distinguishes the background, products, logos, text, calls to action, and regulatory notices.

The name varies depending on where the feature is accessed. BRIA calls it Delayer Agent in its announcement and “De-layer” in Bria Create, while the version available on fal is named Ad Delayer. The principle remains the same: convert a finished visual into a document made of independent layers.

The tool does not recover the original design file. It attempts to rebuild a usable representation from the final rendering alone. The layer order, boundaries, purpose, and content are therefore inferred from the image. This distinction matters when BRIA claims that losing the PSD no longer means losing the ad.

The API returns a JSON document containing the canvas dimensions and a list of layers. Each layer includes an identifier, type, subtype, stacking order, and bounding box. Extracted visual assets are linked to an address from which they can be retrieved separately.

In the example published on fal, a jewelry ad is broken down into nine elements. The document identifies the background color, a background image, two products, and five text blocks. Each object can then be moved, replaced, edited, or downloaded.

The extracted images retain their position and fitting mode within the composition. A product can therefore be identified as an image placed above the background, complete with its coordinates, width, height, and display order. This structure provides the information required to reconstruct the visual in a compatible interface.

Text receives special treatment. Delayer restores it in two forms: a genuinely editable text block and an SVG version that reproduces its original appearance. The first can be used to change a price, headline, date, or notice. The second preserves a visual reference when the detected typeface does not exactly match the initial rendering.

The editable block contains more than a transcription. The schema shown on fal also includes the estimated font family, weight, size, color, letter spacing, alignment, line height, capitalization, underlining, and italics. External style sheets can be referenced to load the associated fonts.

This dual output avoids a common trade-off in automated conversion. Vectorized text can remain faithful to the image but be difficult to rewrite. Recognized text is easy to edit, but its appearance may change if the exact font is unavailable or if the layout has been interpreted incorrectly. BRIA supplies both representations so the client application can choose the most appropriate one.

The presence of an SVG does not guarantee that every character will be recognized correctly. Hand-drawn lettering, curved text, proprietary typefaces, or very small notices can complicate reconstruction. Users will need to compare the result with the original ad before producing a new version.

The functional classification of elements is intended to support automation. A layer is not merely described as an image or text. It can be identified as a logo, product, primary headline, secondary copy, call-to-action button, or compliance notice. An application can then apply different rules to each category.

A localization tool could translate the text while leaving the logo and product untouched. A versioning system could reposition the call to action, shorten the headline, and adapt the background to another format. Regulatory notices could be isolated and replaced with versions specific to each country.

Delayer does not necessarily perform all these changes by itself. The new direct access mainly supplies the structure on which an editor, agent, or production pipeline can operate. BRIA already offers features in Bria Create for generating, resizing, localizing, and animating ads.

This separation could interest brands with years of campaign archives but incomplete collections of working files. An approved ad can serve as the starting point for new sizes, languages, or offers without being entirely rebuilt by a design team.

It could also reduce dependence on the software originally used for production. The response is not presented as a native PSD file, but as a structured document designed for integration. A platform still needs to interpret the JSON, download the assets, and provide its own composition engine to display and edit the layers.

The format is therefore an advantage for developers, but not necessarily a deliverable that someone working only in Photoshop can use immediately. An additional conversion step may be required to create a PSD, a file for another editor, or a template compatible with an internal tool.

BRIA says the visual retains the full resolution of the uploaded file. This means the composition is not intentionally reduced to a simple preview. It does not recreate information missing from the source. A compressed, enlarged, or low-quality image will remain limited by the file provided.

Reconstruction faces another difficulty: a flattened image does not show what was located behind its objects. If a product obscures part of the background, no hidden layer is available to reveal that area. The system can separate visible elements, but moving them significantly may expose missing sections that require additional editing.

The result is also not guaranteed to match the organization of the original document. Two objects grouped together in the PSD may be separated by Delayer, while distinct elements could be merged. Names, groups, masks, effects, and settings that are not visible in the final image cannot be recovered