Getting the Redux + PuLiD Workflow Running
I’ve been playing with the Redux and PuLiD workflow in ComfyUI lately, mostly to see if it could keep faces consistent across different styles. Here’s what worked—and what didn’t—when I tried setting it up.
First, the “Import Failed” error popped up, which usually means something’s missing in the custom nodes. I went into the pulid_flux folder inside my ComfyUI directory and ran pip install -r requirements.txt from the command line. That fixed it most of the time, but if you’re still stuck, double-check the folder paths.
Then there was the “No module named insideface” error. Turns out this one’s picky about Python versions—it needs 3.12. If you’re on an older version, you’ll need to grab the right insightface module from hanamizuki-ai’s Hugging Face repo. I downloaded the 3.12 wheel file, dropped it into my ComfyUI root folder, and installed it manually with:
python.exe -m pip install F:\2\ComfyUI_windows_portable\insightface-0.7.3-cp312-cp312-win_amd64.whl
After a restart, ComfyUI stopped complaining.
The Clip Loader error was trickier. It kept yelling about shape mismatches (torch.Size([1152, 3, 14, 14]) vs. torch.Size([1024, 3, 14, 14])). Updating ComfyUI to the latest version fixed it, which makes sense—older versions sometimes don’t play nice with newer model architectures.
For the actual models, here’s where everything goes:
pulid_flux_v0.9.0.safetensors→ComfyUI/models/pulid/- Flux Redux model →
comfyui/models/style_models(get it from Black Forest Labs’ repo) - Clip Vision model → SigLIP’s Hugging Face page
It’s a bit of a puzzle, but once everything’s in place, the workflow actually holds faces together surprisingly well—even when you throw wild styles at it. I’ll dig into how that part works in the next section.
