NVIDIA FastGen Configure for Claude Code: Review a Diffusion Experiment Setup

FastGen Configure helps align a training or distillation config with the task's network, dataset, supervision, precision, and expected output geometry.

中文

Setup for Claude Code

Claude Code and a FastGen checkout with its existing experiment configs. Static review uses only source and synthetic requirements. Executing a config uses the relevant FastGen environment.

Copy the full pinned skills/fastgen-configure folder into .claude/skills/fastgen-configure/ with its reference files.
Copy skills/fastgen-runtime from the same NVIDIA commit into .claude/skills/fastgen-runtime/. Open Claude Code at the separate FastGen project root.
Invoke /fastgen-configure, explain which task you are planning, and give the closest existing config.

Example task

Describe a text-to-image experiment trained on permitted synthetic geometric images, with no class labels. Give a similar local config that contains a class-conditioning preset. Ask for a settings review and a minimal patch proposal.

Expected result

The review identifies the mismatch, traces the relevant conditioning and loader fields, and gives a config comparison against a compatible local example.

Upstream documentation

Original SKILL.md on GitHub · CC-BY-4.0 license · Source reviewed 2026-11-15

Frequently asked question

Can I reuse settings from any network with the same model family?

Check the exact task mode and compatible geometry, presets, and requirements before copying values from another experiment.