AI models and the GPU
Which nodes use an AI model pack, how the packs download, and how the tools share and free graphics memory.
Model packs
The AI nodes run models that are downloaded once, separately from the tools. The simulations, the comp, convert and creative nodes need no download.
- Tracker models, about 1.2 GB: AI Point Tracker, AI Planar Tracker, AI Vector Tracker.
- Scene Solver models, about 4.6 GB: Scene Solver.
- Masking models, about 3.3 GB: Matte Creator, AI Matte Propagator, AI Matte Perfector, Edge Detail Extension.
- Depth & Normals models, about 1.5 GB: Depth and Normal Creator. The other depth and normals nodes work on what it makes, without a model of their own.
- Upscaler models, SeedVR2 included, about 6.9 GB: AI Upscaler - Matcher.
The AI engine the models run on (about 2.3 GB) comes with the first AI pack you choose, once.
Downloading
Tick the packs you want in the installer, or none: the first time a node needs a pack it asks, and says how big it is. Add or remove packs later in ANR Update.
Some models are published by their authors for non-commercial research only. They are not part of ANR Tools, not included and not licensed by us; options that would use one say "research only, not included". See research models.
The graphics card
The AI nodes and the simulations need an NVIDIA RTX 20, 30, 40 or 50 series card (or RTX professional) with driver 580 or newer. 8 GB of graphics memory or more is best: models and simulations live in it while you work.
Graphics memory, freed by itself
An AI node keeps its model on the card while you use it, so the next frame does not wait for it to load. A simulation keeps its engine there, so the next frame carries on from the last one. When you move on to other work, the tools give that memory back by themselves:
- An AI model nobody has used for 10 minutes is unloaded. If the card is running short of memory (less than a fifth of it free, or less than 2 GB), a model idle for 30 seconds is unloaded instead. The next time a node needs it, it loads again.
- A 3D simulation (the fire nodes, the muzzle flash, sparks and the blood nodes) whose engine nobody has viewed for 10 minutes releases it; when the card is short of memory, after 1 minute. Viewing the node again simulates again from its start, or plays its disk cache if it has one.