FC StyleLoraLoad

FC StyleLoraLoad is how ComfyUI-FaceChain solves the "which model and which style" problem for you. Where most ComfyUI graphs start with a plain CheckpointLoaderSimple plus your own hand-picked LoRAs, this node bundles both steps into one dropdown: pick a style name and it loads the matching base checkpoint, merges in the style LoRA at its intended strength, and hands you the prompt text to go with it.

That's a genuinely nice bit of UX, and it's the reason the pack's workflow files just work when you drag them in. You get MODEL, CLIP, VAE, and style_prompt (STRING) out. The first three wire into the sampler exactly like a checkpoint loader's outputs; the style_prompt string goes into a CLIP Text Encode node. The README is explicit about the intent: "load the checkpoints and style Lora used by facechain, download them first, and then merge them, providing relevant prompts."

How it works

The one input - style_name - is an enum with ~75 choices, split across two base models that ship in the pack's styles/ directory:

• leosamsMoonfilm_filmGrain20 - a film-grain, photographic base (styles like Hanfu, Barbie Doll, Working Suit, Cybernetics punk)
• MajicmixRealistic_v6 - the well-known SD 1.5 realistic checkpoint (styles like Summer Ocean Vibe, Men's Suit, Retro Style)

Each style entry is a JSON file that describes its model_id, the LoRA revision, the LoRA file, and the multiplier strengths for style vs. human. The node then does the ComfyUI thing: loads the base checkpoint (guessing config from the filename), loads the style LoRA with load_lora_for_models, and returns the merged model plus the style's canonical prompt. Because this is SD 1.5 territory, a style here means "a real photo with a specific look," not an SDXL or Flux style.

The gotcha: everything downloads from ModelScope

Here's what will surprise you. The base checkpoints are not in your models/checkpoints folder, and this node does not look on Hugging Face. On first use of a given base model it runs snapshot_download from ModelScope (Alibaba's hub) - ultimatech/leosamsMoonfilm_filmGrain20 or ultimatech/majicmixRealistic_v6 - and if the checkpoint isn't already in ComfyUI's checkpoints path, it downloads it there. The style LoRAs download from ModelScope the same way. So the first time you touch this node you're looking at a multi-gigabyte download plus a wait while it merges, and if your network can't reach modelscope.cn the node just stalls.

It also means the node is doing checkpoint loading and LoRA loading in one shot, so it's slower to construct than a plain loader. That's fine in a graph where you only need one style per run. If you want to switch styles mid-workflow, the JS that ships with the pack even forces this node to a fixed width in the UI so the long dropdown list stays readable.

Troubleshooting

• Node "hangs" on first use - it's downloading a checkpoint or LoRA from ModelScope. Watch the console for snapshot_download lines.
• Model already in your checkpoints folder but ignored - the node matches by the exact filenames leosamsMoonfilm_filmGrain20.safetensors / majicmixRealistic_v6.safetensors. Name matters.
• Style dropdown feels like a wall of Chinese names - that's expected; every style is bilingual (汉服风(Hanfu)), so the English is in the parentheses.

Install is the standard pack routine - ComfyUI Manager (search "ComfyUI-FaceChain") or git clone https://github.com/THtianhao/ComfyUI-FaceChain into custom_nodes/, then restart. Just budget for the ModelScope downloads and the pack's heavy startup dependency install (onnxruntime-gpu, mmdet, mmcv, insightface, etc.) before the dropdown even renders.