What does this skill do, and when should you use it?
This skill guides NVIDIA TAO Toolkit workflows for FastFoundationStereo (FFS), a stereo depth estimation model. It covers training, evaluation, export, TensorRT engine generation, and PyT or TensorRT inference through the unified TAO depth_net CLI. FFS is the bp2 commercial distilled variant of FoundationStereo, using pruned per-layer widths and a serialized forward path for approximately 10× lower latency than the full model. Docker, nvidia-container-toolkit, container-accessible data, and a matching bp2 commercial checkpoint are required for representative deployment workflows.
Reads stereo image pairs, optional ground-truth disparity, optional occlusion masks, and per-line annotation files with two to four columns. Configures TAO depth_net with FastFoundationStereo, a dataset class, the required 15 bp2 model-width overrides, and action-specific training or deployment settings. Runs train, evaluate, inference, export, and gen_trt_engine actions, producing checkpoints, KPI metrics, disparity samples, static FP32 ONNX, or FP16 TensorRT engines. It also provides checks for container exit status, status.json KPI output, training loss, and PyT-versus-deployment drift.
- A vision engineer fine-tuning FFS on a proprietary stereo dataset and taking the result through TensorRT deployment.
- A team with model_best_bp2_serialize.pth that wants to evaluate or deploy the commercial bp2 model without training.
- A researcher training or evaluating on Middlebury, KITTI, ETH3D, FSD, IsaacReal, or Crestereo data.
- A developer running disparity inference on stereo pairs without ground-truth labels and inspecting the generated samples.
- A deployment engineer exporting FFS to static ONNX and generating an FP16 TensorRT engine.
How do you install this skill?
- Before running, explicitly confirm Docker, GPU, mount paths, and the scope of writable output directories; the skill provides no automatic rollback or cleanup strategy.
- Provide and validate the bp2-compatible model_best_bp2_serialize.pth checkpoint. Missing width overrides, max_disparity=192, or gwc_feature_normalize=true can cause shape errors or invalid disparity.
- Dynamic deployment inputs that violate stride constraints may show success in status.json while producing no artifacts; inspect actual output files rather than relying only on exit status.
- Chinese-language usability and mainland-China reachability are not evidenced; S3, container images, and TAO dependencies may create deployment-access risks.
- Shell / CLI
- Local filesystem
Dockernvidia-container-toolkit
Install the individual skill with the skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill tao-train-fast-foundation-stereo --yes
For Codex-specific installation:
npx skills add nvidia/skills --skill tao-train-fast-foundation-stereo --agent codex --yes
Execution also requires Docker and nvidia-container-toolkit. The source does not document a separate package version or standalone installer for this skill.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-train-fast-foundation-stereo" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- train fast stereo,
- real-time stereo disparity,
- distilled stereo depth.
Ask the Agent for a task involving “train fast stereo,” “real-time stereo disparity,” “FastFoundationStereo,” or “distilled stereo depth.” Provide container-accessible stereo images, optional ground truth, an annotation file, and the bp2 checkpoint path. Training uses AutoML by default; explicitly request “disable AutoML” or set automl_policy: off to disable it for that run. For deployment actions, follow the referenced TAO Deploy instructions and use the train, evaluate, inference, export, gen_trt_engine, inference, evaluate sequence; raw-bp2 deployment can skip training.
What are this skill's strengths and limitations?
- Covers the full path from training and validation to ONNX export, TensorRT generation, and inference.
- Documents the FFS-specific bp2 width invariants, max_disparity requirement, and deployment constraints.
- Supports several registered datasets plus GenericDataset for non-canonical layouts.
- Includes practical guidance for KPI verification, error patterns, and checkpoint handoff.
- Requires Docker, nvidia-container-toolkit, and correct mounts for data, outputs, and checkpoints.
- Raw deployment and representative fine-tuning depend on a pretrained bp2 commercial checkpoint that the source does not provide.
- Without that checkpoint, the workflow can be validated but the resulting metrics are not representative of the bp2 model.
- The source does not provide a concrete GPU compatibility matrix or a standalone test suite.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| TAO FastFoundationStereo Real-Time Stereo Depth this page ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO FoundationStereo Stereo Depth ✓ NVIDIA · Official | 52 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Depth Anything v2 ✓ NVIDIA · Official | 50 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO CenterPose Pose Estimation Skill ✓ NVIDIA · Official | 49 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO Grounding DINO Open-Vocabulary Detection ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
Compared with full FoundationStereo, FFS is its bp2 commercial distilled variant with pruned per-layer widths and a serialized forward path, targeting approximately 10× lower latency. The entrypoint, action verbs, dataset classes, and deployment chain remain the same.
How did FollowSkills review this skill?
The skill limits declared tools to Read and Bash and requires Docker, NVIDIA Container Toolkit, read-only data mounts, and user-supplied checkpoints. It discloses data, output, and checkpoint paths and describes the training, evaluation, export, and deployment chain. Deductions apply because Bash and Docker permit substantial external computation, file writes, and GPU-resource effects, while the documentation lacks explicit stepwise user confirmation, strong isolation, sensitive-data handling, dependency-integrity checks, and rollback procedures. The license and NVIDIA provenance improve traceability but do not replace safety controls.
The documentation gives concrete guidance on model widths, data formats, checkpoint resolution, KPI differences, common errors, and silent failures, and it specifies AutoML versus direct-training routing. Deductions apply because static review cannot reproduce the key paths; the container, TAO Core, tao-skill-bank, TensorRT, and referenced templates are not fully available in the supplied evidence, and some failures still shift diagnosis to the user. Under static calibration, the score is capped at 10.
Trigger phrases, supported actions, raw bp2 deployment and fine-tuning boundaries, dataset classes, inference input formats, and checkpoint prerequisites are described. Deductions apply because the skill requires an NVIDIA GPU, Docker, NVIDIA Container Toolkit, a commercial bp2 checkpoint, and reachable data paths. Chinese-language interaction and mainland-China network reachability are not documented, while examples depend on S3 and container infrastructure, so environment-fit evidence is incomplete.
Metadata provides a name, version 0.1.0, author, license, compatibility requirement, and tags; the documentation is layered across prerequisites, workflow, parameters, error patterns, and deployment references. Deductions apply because the benchmark report identifies missing recommended Instructions and Examples sections, a nonconforming author format, and unexpected folder depth. No clear changelog, maintenance owner, support lifecycle, or update path is provided, and several key details depend on referenced files not fully shown in the evidence.
The skill covers training, evaluation, inference, export, TensorRT engine generation, and deployment, with directly reusable YAML/Python parameter blocks, verification metrics, and a seven-action workflow. Deductions apply because nothing was executed for this review; output quality and the roughly 10× latency claim were not independently verified, and users must still provide the container, GPU, dataset, bp2 checkpoint, and correct paths and configurations. Under static calibration, the score is capped at 7.
The evidence includes fixed revision metadata, parameter templates, error-pattern documentation, an evaluation task, and an NVSkills-Eval report covering security, correctness, discoverability, and static findings. Deductions apply because the benchmark has only one task and no negative tasks, its reported results do not replace independent reproduction here, and there is no multi-source corroboration; key model behavior and performance claims remain primarily documentation assertions. Under static calibration, the score is capped at 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →