Your asset library, searched by meaning and reused by the agent

Every working artist has a folder of assets they have already made, and a naming scheme that made sense at the time. Six months later, finding the right chair in it takes longer than modelling a new one, which is how most asset libraries quietly stop being libraries.

Mixar now treats that library as something the AI agent for Blender can actually use. Train it on the Blender asset libraries you choose, and three things change: search works by meaning instead of filename, your assets become browsable cards in chat, and the agent reuses them when it builds, instead of remodelling things you already own.

Training: opt-in, incremental, and in the background#

Training happens in the Assets workspace. Tick the libraries you want indexed, press Train Model, and Mixar renders a preview of each asset and embeds it into a visual index. Nothing is indexed until you enroll it, and unenrolling a library removes its entries.

The first pass over a large library takes a while, so it runs in the background with live progress and a cancel button, and you can keep working through it. Later runs are incremental: only new or changed assets are re-indexed, and enrolling or unenrolling retrains on its own.

Because the index is built from rendered previews, search understands appearance. A query for "globe" finds the globe whose datablock is named after a UUID, and a reference image finds the assets that look like it.

Three ways to reach the same index#

Search, in the Assets workspace. Type a prompt, attach a reference image, or both. Hits come back ranked, and each one can point the asset browser at the matching asset so it is one drag away.

Library Mode, in chat. Switch the chat to Library and your enrolled assets appear as clickable thumbnail cards. An empty message browses, typed text runs the same semantic search, and clicking a card places the asset at the 3D cursor. Browsing and placing work offline.

The agent, while it builds. Ask Mixie to use your library and it searches the index before modelling each object. A clear match gets placed. Ambiguous matches become a question, with each candidate shown as a thumbnail next to a model-from-scratch option, so the close calls stay yours.

A match threshold slider controls how eager that reuse is: raise it and the agent only takes near-certain matches, lower it and it leans on the library harder. The agent can also save objects it just built into a library you name, so the next scene starts with more to reuse.

Generations become a library on their own#

Completed Image to 3D and Model Gen generations are archived automatically into a library called Mixar Generations, with a rendered thumbnail, and indexed alongside everything else. A model you generated last week is findable by describing it, and the agent can place it instead of spending another generation on the same object.

Where this fits in the larger picture: the same agent runs the mechanical passes covered in the Blender MCP guide, and reuse is one more reason a scene brief resolves into placement work rather than modelling work. For what the generation side does when the library has no match, text to 3D covers the options.

Frequently asked questions

How does Mixar search a Blender asset library by meaning?

Training renders a preview of each asset in the libraries you enroll and embeds those previews into a visual index. Queries, whether text or a reference image, are matched against that index, so results depend on what assets look like rather than what their files are called.

Can the agent build a scene using only my own assets?

Yes. Say "only use assets from my library" and the agent places matches and skips objects it cannot find, instead of modelling substitutes. Without that instruction it prefers your assets where they match and builds or generates the rest.

Do my generated models end up in the library too?

Completed Image to 3D and Model Gen generations are archived into a Mixar Generations library automatically, with thumbnails, and indexed for search. It enrolls itself the first time it is created and can be unenrolled like any other library.

Is my asset library uploaded when I train?

Training uploads the rendered preview images and asset metadata, such as names, descriptions, tags and which file each asset lives in, to build the search index. The blend files themselves stay on your machine, and a delete control wipes the server-side index whenever you want to start clean.