GPU VRAM

≥ 24 GB (gemma4-Peak has not yet been measured)

≥ 24 GB; more for concurrent jobs or presets

The naming model is included in all presets gemma4 (~16 GB of weights); the actual peak has not yet been measured, so assume ≥ 24 GB until then. The old values (10–12 GB, visual_3d 9–11 GB, norm_exact 8–10 GB) apply only to opt-in images with gemma3. Token-free variant eva02: 6–8 GB; eva02 runs before naming, but sequentially—peak = max(eva02, gemma4).

CPU cores

4 cores

8 cores and more

Prefetch workers = min(6, cores − 2); can be overridden via RUNPOD_PREFETCH_WORKERS; 0 = serial operation. The workers also render the contact sheets for naming (~15 s per sheet) and thus affect both phases.

Random Access Memory

≥ 16 GB

≥ 16 GB base plus ~1 GB per prefetch worker

At most one mesh per worker in memory: six workers with 100-MB STLs ≈ 22 GB, plus ~0.5 GB per worker for contact sheet rendering.

Hard Drive

≥ 70 GB free

≥ 70 GB free space for the image and loading the tarball

Image ~28 GB (model weights ~22 GB, of which gemma4 accounts for ~16 GB); loading the tarball temporarily requires about twice that amount.

Share

5 times the size of the catalog on the hard drive

5 times the size of the catalog on the hard drive

Makes the data accessible to both servers involved.

Network

No Internet access required

No permanent connection

All models are included in Docker .