I’m looking at nvidia/tao/fast-foundationstereo on NGC. I can see versions v1.0 and v1.2, and the model card lists ~14.6M parameters (a distilled/pruned real-time model).
Could you confirm:
Which of these versions are trainable checkpoints intended for fine-tuning via TAO (as opposed to deployable/inference-only)?
Is Fast-FoundationStereo offered in only one size (~14.6M), or are there multiple encoder/size configurations available for training?
Is the TAO fine-tuning workflow (data format, train/export) the same as FoundationStereo, and is a single RTX 4090 (24GB) sufficient to fine-tune it?
Any guidance on how its fine-tuning accuracy compares to fine-tuning the full FoundationStereo (vitl) for a domain-specific case?