Face_Anon_Simple_LoadModel

Face_Anon_Simple_LoadModel is the pack's entry point, and the name undersells it. It's not pulling one checkpoint off the shelf - it's assembling a four-part SD 1.5 diffusion stack that totals roughly ten gigabytes, most of which downloads on your first run. It's also the node you touch least: set it up once, wire the model output into the sampler, forget it.

Why it looks so different from a normal ComfyUI loader

Almost everything you load in ComfyUI is a single safetensors file. This pack is different because it's a port of face_anon_simple, a research project built on Hugging Face diffusers rather than the native ComfyUI stack. So this node does the equivalent of from_pretrained() three times in a row: once for a standard SD 1.5 UNet, once for a ReferenceNet, once for a second "conditioning" ReferenceNet - each its own ~3.2GB file plus a config.json, all pulled from hkung/face-anon-simple on Hugging Face. It then bolts on a VAE from your ComfyUI models/vae folder and a DDPM scheduler bundled inside the pack. The result is a StableDiffusionReferenceNetPipeline - the thing the sampler actually runs.

That architecture has consequences. First, first-run patience: three multi-gigabyte downloads, which Hugging Face caches to disk (the README calls out that this lands on the C: drive on Windows). Second, you need a real SD 1.5 VAE sitting in ComfyUI/models/vae - the README names vae-ft-mse-840000-ema-pruned.safetensors, and that's the standard choice.

The inputs that matter

• repo (STRING, default hkung/face-anon-simple): leave it as the HF repo id to load straight from the Hugging Face cache. Clear it to an empty string and the node loads from ComfyUI/models/Face_anon_simple instead - auto-downloading into that folder if it's missing. That's the "I want to know where my 10GB lives" option.
• vae (dropdown): pick your SD 1.5 VAE. The default none is a trap - the code raises an error if you run it like that.
• lowvram (BOOLEAN, default True): enables model CPU offload so the whole pipeline doesn't need to sit in VRAM. Leave it on unless you're benchmarking.

Output: a single model (FACEANON_PIPE), which wires straight into Face_Anon_Simple_Sampler. There's nothing else to connect.

Installing

Via ComfyUI Manager, search "Face_Anon_Simple". Or, per the README:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_Face_Anon_Simple.git
pip install -r requirements.txt

The requirements file is a single line - face_alignment. That's the only hard Python dependency; torch, diffusers and transformers should already be in your ComfyUI environment. Restart ComfyUI after installing.

The gotchas

• First run is slow and heavy. Three 3.2GB files plus the VAE you still have to supply. Budget disk space and a few minutes.
• none VAE is an instant error. The default dropdown selection isn't a valid choice - it's the pack telling you to pick one.
• This is a research port, and it shows. smthemex ships model ports to ComfyUI fast (StoryDiffusion, Sonic, LucidFlux, TwinFlow, SenseNova U1, ...), and users occasionally hit example workflows that lag behind the current code. If the node errors at load, check you're on the latest version and the models actually finished downloading - that's where most of the pain lives in this family of packs.