StyleGAN Sampler

StyleGAN workflows in ComfyUI are weirdly inverted compared to everything else you use. There's no prompt, no KSampler, no denoising loop. You load a model, you make a latent, and this node - StyleGAN Sampler - is where pixels finally happen. If you've googled this node, you're probably one step away from "wait, that's it?" and then 64 images a second, which is roughly the point of the whole pack.

What it actually does

The StyleGAN generator is two networks bolted together: a mapping network that turns random z noise into a w style code, and a synthesis network that draws the image from that w. The sampler is the synthesis half. Give it a loaded STYLEGAN model plus a STYLEGAN_LATENT, and it runs the synthesis network and hands you an IMAGE out the other side, already converted from the model's internal [-1, 1] range to the [0, 1] range every other ComfyUI node expects.

It's a batch-friendly node, too. If your latent holds 8 vectors, it renders all 8 and concatenates them into one image batch. Wire the IMAGE output straight into a Preview or Save node.

The inputs that matter

Four inputs, and only two of them are decisions:

• stylegan_model - the STYLEGAN output from Load StyleGAN Model.
• stylegan_latent - from Generate StyleGAN Latent (or the inversion node, if you're reconstructing a real photo).
• noise_mode - const or random. This is the interesting one. Every StyleGAN layer has a small stochastic noise input that isn't part of the latent at all. const pins it to a fixed pattern, so the render is clean and reproducible. random lets each layer's noise roll fresh, which shifts subtle texture and detail from run to run - the same latent, slightly different skin grain. For stills you want const. For generating animation frames where a little per-frame flicker is desirable, random.
• seed - seeds the RNG, which mostly matters when noise_mode is random. In const mode the render is deterministic regardless.

Installation

Install the pack once and every node in it comes along. From ComfyUI Manager, search "ComfyUI-StyleGan", or:

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

Then restart ComfyUI and drop a model into ComfyUI/models/stylegan/ (create the folder - it doesn't exist by default). That part is easy. The hard part is the custom CUDA extensions.

Where people get burned

StyleGAN ships CUDA kernels (bias_act_plugin, upfirdn2d_plugin, filtered_lrelu_plugin) that compile at runtime - the first time you run a workflow you'll see "Setting up PyTorch plugin..." in the console, and it takes 30–60 seconds. That build is where almost every failure lives. The README is blunt about it: you need the CUDA Toolkit, ninja, and a C++ toolchain (GCC on Linux, Visual Studio on Windows), plus the right CUDA_HOME / PATH / LD_LIBRARY_PATH set. ComfyUI portable users often have to copy the Include/ and libs/ folders from a full system Python 3.10 install into python_embeded.

A common symptom from real users: GLIBCXX_3.4.32 not found when bias_act_plugin fails to load. That's your conda/SwarmUI environment shipping a stale libstdc++.so.6 while the freshly built plugin needs a newer one - upgrade the system/conda libstdc++, or move the pack to a Python environment that isn't pinned to an old toolchain. Note the pack's requirements.txt is just numpy, pickle, ninja - no fake dependencies hiding in there. The pain is the compiler, not the pip line.

Once built, it's fast. The README clocks 64 images/sec on a 3090 with a large batch, though the preview encode throttles realtime use to around 8 fps. That's still silly fast next to a 30-step diffusion pass, and it's the whole reason to bother with GANs in 2026.