Choose processing for your hardware

You don’t need a GPU to get started. Install with the CPU setup, try a small sample, and add hardware or Cloud processing as your library grows. The hardware guide explains device access and supported runtimes, and Ask Frameleaf can help you narrow your choices.

Match the task to the hardware

Machine learning and video transcoding have their own settings, and a GPU that handles video encoding well may still be too small for a description model. Smart search, faces and text recognition run on your server or a worker on your own network, and Studio’s final export stays at home too. For heavier tasks, Cloud jobs are there when you want them, with a clear destination and a cost estimate before anything runs.

For model advice, give the exact GPU model, total and currently free dedicated memory, host platform and runtime. Shared or unified memory also serves the operating system and other apps, so count it separately from dedicated VRAM, and leave headroom for jobs running at the same time.

Choose a description model

The description model table lists supported model families with memory estimates, and the adviser adds working-memory headroom on top. Its local recommendations cover reviewed NVIDIA Ampere and Ada configurations using CUDA; for other hardware it points you to the right guide and asks for the details it needs.

Florence captions are a lightweight fallback and don’t follow custom prompts. Qwen and Phi models follow the supported description settings. Stick to the supported model families: others can fail to load even with plenty of memory.

Compare a sample first

Use Preview a description to try a small sample before you change the whole library. Changing a search model can mean rebuilding its embeddings. See Search models and AI descriptions.

Cloud GPU lets you run supported heavy tasks without buying a bigger GPU. Your server shows which media will be sent, the model and the estimate before you confirm. Turnaround depends on upload time, queue time and the size of the task. See Cloud GPU processing.

If a model fails to load

1. Check the model family is supported, then check the runtime, download source and free memory.
2. Read the machine learning logs. A failed download and an out-of-memory error need different fixes.
3. Test one item before you retry a full queue.

See Hardware acceleration and Troubleshooting.