Ghostwritten BiRefNet Loader

Ghostwritten BiRefNet Loader is the boring half of the two-node Ghostwritten Studios BiRefNet pack, and boring is the job. It doesn't remove anything. It loads a BiRefNet segmentation model once, hands you a BIREFNET_MODEL handle, and lets you share that handle across as many Remove Background nodes as you like. The actual cutout happens in Ghostwritten Remove Background; this is the optional loader in front of it.

So why does the loader exist at all? Mostly organization. The Remove Background node is perfectly happy to lazy-load a model itself - if you never touch this node, the workflow still runs. The loader buys you a single knob for model and precision when you've got several remove nodes in one graph, say because you're comparing the general weights against the portrait or HR variants on the same image. Switch once at the top instead of hunting through every downstream node. That's a small win, but when you're iterating on a cutout workflow it's the kind of small win you stop noticing only after you've got it.

How it works

Under the hood it's a thin wrapper. The node maps your friendly model name to a Hugging Face repo ID from a hardcoded registry (ZhengPeng7/BiRefNet, BiRefNet_lite, BiRefNet-portrait, BiRefNet_HR), picks a device - CUDA, then MPS, then CPU - and loads the weights through transformers.AutoModelForImageSegmentation with trust_remote_code=True. Loaded models go into a global cache keyed on (repo, device, dtype), so repeated runs reuse the weights instead of re-reading them off disk. The output handle is just that loaded model plus its device, dtype, and input size.

Two things follow from that. First, the weights are pulled straight from Hugging Face on first use and cached under ~/.cache/huggingface - this pack does not read ComfyUI's models/background_removal/ folder, so even if you already have the BiRefNet safetensors that ComfyUI core ships, the first run still downloads its own copy. Second, precision only does anything on CUDA. The source falls back to fp32 on MPS and CPU, silently. fp16 is the default and the right call on a GPU - it costs essentially no accuracy - but don't pick fp32 on a Mac and expect it to matter, because you were getting fp32 anyway.

The inputs that matter

• model_name - BiRefNet (general) (the default, best all-rounder), BiRefNet_lite (faster) (Swin-Tiny, less accurate, lighter), BiRefNet-portrait (people, trained on P3M-10k), or BiRefNet HR (1536) (for bigger inputs, at a VRAM cost).
• precision - fp16 default, fp32 if you must. On CUDA, fp16 is the one to use.

One honest caveat about that HR entry: the pack runs it at 1536px internally, even though the HR weights were trained at 2048. Feed it a 4K render and it will still downscale - fine for most sources, but if your whole reason for reaching for HR is very large inputs, you're leaving some of it on the table.

Install and gotchas

Install is the same for the whole pack, so once it's in you get the Remove Background node free. Easiest is ComfyUI Manager: search "comfyui-ghostwritten-BiRefNet" (or just "Ghostwritten") and install. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/GhostwrittenStudios/comfyui-ghostwritten-BiRefNet

Restart ComfyUI so the nodes register. The pack needs timm, einops, and kornia on top of the transformers and torch that already ship with ComfyUI - Manager usually handles those, but if the loader throws an import error, that's what's missing. The first load will look like it's hung: it's downloading a few hundred MB of weights over the network. Let it finish once and it's cached forever. If you wired a loader into your graph and a Remove Background node still re-loads, check that the loader's birefnet_model output is actually connected to that node's input - that's the whole deal here.