Load StyleGAN Directions

Discovering a direction is the fun part. Remembering which component on which seed produced that one good jawline edit is not. This node is the boring-but-necessary half of the workflow: it loads a saved bundle of StyleGAN directions off disk, so your edit stops being "rerun PCA and hope the ordering didn't shift".

It pairs with the pack's standalone scripts. discover_directions.py writes each component as its own named tensor in one .safetensors file (component_00, component_01, …) plus metadata recording the discovery order, so loading it back is exact rather than reconstructive:

python discover_directions.py model.safetensors --method sefa

By default that lands in ComfyUI/models/stylegan_directions/ - a sibling of models/stylegan/, deliberately not the same folder, so model checkpoints and direction bundles never show up mixed in one file picker.

What you set

stylegan_directions_file is a dropdown populated from ComfyUI/models/stylegan_directions/*.safetensors only. The pack registers that folder with a .safetensors-only extension list, so a .pkl or a stray .json won't appear, and a bundle dropped anywhere else won't either.

direction_name is optional and does two quite different things:

• Blank → you get the whole bundle back as a stacked tensor, in the order the file's metadata says (component_order), falling back to sorted key order. That's what you want for StyleGANDirectionSweep-style exploration, where you're poking component_index to see what exists.
• Filled in (e.g. component_03) → you get that one direction, wrapped as a single-element batch. Useful once you've identified the keeper, because then every ApplyStyleGANDirection downstream has component_index = 0 and nothing to misclick.

Output is one STYLEGAN_DIRECTIONS, wired into ApplyStyleGANDirection or StyleGANDirectionSweep. Once it's a file, this is deterministic: same bundle, same edit, next month, different machine.

Where the files come from

The in-graph discovery nodes (DiscoverGANSpaceDirections, DiscoverSeFaDirections) don't write anything - they hand back an in-memory bundle that dies with the session. So you have two options for permanence:

1. Keep the discovery node in the graph. Re-running costs seconds for SeFa and a handful for GANSpace, and the result is reproducible for a given seed.
2. Export offline with discover_directions.py and load the file here. This is the option that survives checkpoint swaps, model-folder reorganisation, and someone else opening your workflow. The script even renders a preview grid if you pass --sweep 0,1,2 --sweep-out sweep.png, so you can identify components before you ever open ComfyUI. Note --sweep renders images, so it needs the compiled StyleGAN ops working; discovery itself doesn't.

Install

ComfyUI Manager → search ComfyUI-StyleGan (publisher dflfd), or:

cd ComfyUI/custom_nodes
git clone https://github.com/dfl/ComfyUI-StyleGan

Restart. Models in ComfyUI/models/stylegan/, directions in ComfyUI/models/stylegan_directions/ (create it if the node's dropdown is empty - the folder is created lazily by the pack's scripts, not guaranteed to exist on a fresh clone). The pack's four pip deps are trivial; the actual pain is the runtime-compiled CUDA extensions needing CUDA Toolkit, ninja and MSVC/GCC, with a 30–60s plugin build the first time you render.

Because the offline script imports the pack directly, run it with the same Python environment ComfyUI uses if you can - it needs torch and safetensors at minimum, and it resolves dnnlib/torch_utils from the pack folder itself.

Where people get burned

• The file isn't in the dropdown. It's not in models/stylegan_directions/. The picker does not browse anywhere else.
• direction_name typo. You get a ValueError that also lists every key actually in the file, which is the friendliest failure in this whole pack. Read it instead of guessing again.
• Mismatched checkpoint. A direction bundle carries its method, component count and w_dim - but not which model it came from. Load a SeFa bundle from model A against model B and the node happily hands it over; the edits just come out as garbage or mush. Name your files with the model in them; the script's default filename already does.
• A direction that felt strong yesterday feels flat today. Different PCA run means different ordering and sign. Once you've found a keeper, export it. That's the entire reason this node exists.