stylegan2-pytorch
Guides StyleGAN2 PyTorch image GAN training, checkpoint generation, interpolation, and programmatic sampling with the stylegan2_pytorch package.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- stylegan2-pytorch
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides StyleGAN2 PyTorch image GAN training, checkpoint generation, interpolation, and programmatic sampling with the stylegan2_pytorch package.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/repo-provenance.md1.8 KB
- 参考文档references/repo-routing-metadata.json337 B
- 参考文档references/troubleshooting.md4.1 KB
- 脚本scripts/check_install.py5.6 KB
03 SKILL.md 正文
StyleGAN2 PyTorch Repo Skill
Use this repo skill when a task involves the stylegan2_pytorch package, the
stylegan2_pytorch command, or StyleGAN2-style image GAN training/sampling in
this implementation. The package is CUDA-first: importing the main module
asserts that torch.cuda.is_available() is true, so do not plan CPU-only use.
Fast routing
- Use training for command-line training from an image folder, checkpoint resume/restart decisions, sample generation, interpolation GIFs, low-data augmentation, FID/logging options, transparent images, attention/vector-quantization knobs, and multi-GPU settings.
- Use programmatic-api when the user
wants Python code with
ModelLoader,Trainer, orStyleGAN2, especially loading a trained checkpoint and saving generated images from tensors. - Read references/troubleshooting.md when an install, import, CUDA, optional dependency, CLI parsing, checkpoint, or package-version failure blocks either workflow.
- Read references/repo-provenance.md before deciding whether this skill is stale for a checkout or should be refreshed.
Package facts to keep in mind
- Distribution name and import name:
stylegan2_pytorch. - Console command:
stylegan2_pytorch. - Public exports from the package root:
Trainer,StyleGAN2,ModelLoader, andNanException. - The core CLI is implemented with Python Fire around
train_from_folder, so Fire help displays underscore flag names even though the README commonly shows hyphenated flag names. - Training inputs are recursive image folders containing
.jpg,.jpeg, or.pngfiles. - Default outputs are
results/<name>/for generated images andmodels/<name>/model_<n>.ptplusmodels/<name>/.config.jsonfor checkpoints and model settings.
Install and environment check
The documented public install path is:
pip install stylegan2_pytorch
For a local clone, an editable install is useful while developing or verifying against that checkout:
pip install -e .
Use a CUDA-capable PyTorch/torchvision build before relying on the package. The
main module imports aim, einops, kornia, vector_quantize_pytorch,
torch, torchvision, and other setup dependencies at import time.
After installing, run the bundled checker from this skill directory if the user has an environment ready:
python scripts/check_install.py
It performs only import, CUDA, signature, and CLI-help checks; it does not train or download data.
Workflow starting points
- For a new training run, open training, then
use its data-layout and CLI references to choose
--data,--name, image size, batch/accumulation, output directories, and augmentation settings. - For generation after a run, keep the same
--name,--models_dir, and--results_dirlayout and use the training generation/interpolation routes. - For Python sampling, open
programmatic-api;
ModelLoaderexpects the defaultbase_dir/models/<name>/checkpoint layout. - If the request is about another image-generation stack such as Diffusers, ComfyUI, Stable Diffusion LoRA training, image classification, detection, or segmentation, this skill is probably not the right operating graph.
Verification status
This skill was constructed with a CUDA backend inspection environment and a
post-integration plan for CLI help, one-step training, and ModelLoader
sampling checks. The user requested not to import the skill into the managed
repo-skill library, so use the generated directory directly unless a later
request authorizes import.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).