Running overnight batch 3D generation: a practical setup guide

The overnight batch feature on Production plans is straightforward to configure but benefits from a specific setup routine. This article describes the workflow that produces the most consistent results: how to structure the prompt list, how to configure the queue, and what to check when you open the results folder in the morning.

Structuring the prompt list ¶

A batch queue works best when the prompts are grouped by object category rather than mixed randomly. The agent processes each prompt independently, but grouping similar objects means that any systematic issue with a particular category shows up as a cluster in the results rather than scattered across the folder, which makes it easier to identify and re-queue.

Each prompt in the batch should include the same level of detail you would use for a single-asset generation: object category, material descriptors, scale reference, and any stylistic modifiers. Prompts that work well for single assets work equally well in batch. Prompts that are vague for single assets produce inconsistent results at scale.

Configuring the queue ¶

In the batch settings panel, set the naming prefix to match your project's asset naming convention. Set the start time to 22:00 JST if you want the queue to begin after business hours. The overnight scheduling mode will hold the queue until that time and begin processing automatically.

Set the export format before submitting the queue. Changing the export format after a batch has run requires re-exporting, which does not re-run the generation but does take additional time. If your pipeline needs both FBX and GLTF, select both formats before submission.

Morning review workflow ¶

When you open the results folder, the first check is the completion report, which lists any prompts that produced a flagged output. Flagged outputs are assets where the manifold check found an issue or where the polygon count deviated significantly from the target. These are worth reviewing before importing into your pipeline.

The second check is a quick visual scan of the thumbnail grid. The agent generates a thumbnail render for each asset at a standard three-quarter view. Scanning the grid takes about two minutes for a fifty-asset batch and catches any obvious category errors before you spend time importing.

The overnight queue is most useful when you have a well-defined asset list and consistent prompt quality. It is less useful as a way to explore what the agent can do with vague prompts. For exploration, single-asset generation with immediate feedback is the better tool.