The asset library
Automatic metadata
Every asset is described, tagged and colour-read on arrival, which is what makes the library searchable at all.
An asset arrives already described. Nobody types a filename, nobody picks tags, and nothing sits in an untitled folder waiting for a tidy-up that never happens. ## What gets written | Written on arrival | What it holds | |---|---| | Name | `Model MM-DD HH:MI`, so the list is sortable and readable without being a sentence | | Kind and format | Image, video, 3D, audio or vector, with the mime type and byte size | | Description | A written sentence describing what is actually in the picture | | Tags | The objects, materials and subjects found in it | | Colours | The dominant colours, as values you can filter by | | Style | How the image reads | | Prompt | The exact words that produced it | | Model and tool | Which engine ran, and which card you ran it from | | Settings | The full parameter set the run used | | Where and when | The board, the workspace, who made it, the timestamp | The description, tags, colours and style come from a vision model reading the asset. That is the "AI tagging" part, and it is what turns a library into something you can search by memory rather than by filename. ## Name versus description The vision model writes a **description**, and the description goes in its own field. It never overwrites the name. A library where every row reads "A red ceramic vase on a white surface, soft window light from the left" is unusable, because every row looks like every other row until you read a full sentence. `Seedream 08-07 14:22` is scannable, and the sentence is one click away when you want it. :::note[Why this matters more later than now] On the day you make something you remember everything about it. The metadata is written for the version of you that comes back in November looking for the one that worked, and that person has no idea what it was called. ::: ## What it powers Three things read the record: - **Search**, which ranks across name, tags, colours, style, description and prompt at once - **Filters**, by kind and by colour - **The provenance line** on every asset, covered in [provenance](/docs/using-plnty/the-asset-library/provenance/) ## Uploads get it too Drag in an image from your desktop and it is read the same way. It will not have a prompt or a model, because there was not one, and it will have a description, tags, colours and style. That means your own reference material becomes searchable alongside what you generated, using the same half-remembered description you would use for either. :::tip[Rename when it earns it] The generated name is a good default and a bad label for the four assets a project actually ships. Rename those. The name follows the asset everywhere, including into the download, so a renamed asset arrives on a client's desktop already saying what it is. :::An asset arrives already described. Nobody types a filename, nobody picks tags, and nothing sits in an untitled folder waiting for a tidy-up that never happens.
What gets written
Section titled “What gets written”| Written on arrival | What it holds |
|---|---|
| Name | Model MM-DD HH:MI, so the list is sortable and readable without being a sentence |
| Kind and format | Image, video, 3D, audio or vector, with the mime type and byte size |
| Description | A written sentence describing what is actually in the picture |
| Tags | The objects, materials and subjects found in it |
| Colours | The dominant colours, as values you can filter by |
| Style | How the image reads |
| Prompt | The exact words that produced it |
| Model and tool | Which engine ran, and which card you ran it from |
| Settings | The full parameter set the run used |
| Where and when | The board, the workspace, who made it, the timestamp |
The description, tags, colours and style come from a vision model reading the asset. That is the “AI tagging” part, and it is what turns a library into something you can search by memory rather than by filename.
Name versus description
Section titled “Name versus description”The vision model writes a description, and the description goes in its own field. It never overwrites the name.
A library where every row reads “A red ceramic vase on a white surface, soft window
light from the left” is unusable, because every row looks like every other row until
you read a full sentence. Seedream 08-07 14:22 is scannable, and the sentence is one
click away when you want it.
What it powers
Section titled “What it powers”Three things read the record:
- Search, which ranks across name, tags, colours, style, description and prompt at once
- Filters, by kind and by colour
- The provenance line on every asset, covered in provenance
Uploads get it too
Section titled “Uploads get it too”Drag in an image from your desktop and it is read the same way. It will not have a prompt or a model, because there was not one, and it will have a description, tags, colours and style.
That means your own reference material becomes searchable alongside what you generated, using the same half-remembered description you would use for either.