Load Falcon Perception Model

The smallest honest node in this pack does the thing every model loader secretly wishes it did: it points at one .safetensors file and loads it. No downloading a whole HuggingFace repo. No trust_remote_code=True. No cache directory quietly filling with .json and .py files you'll never read.

Load Falcon Perception Model is the companion loader for the ComfyUI-Falcon-Perception pack, and it's the reason this pack is as painless to set up as it is. You don't need it to understand the segmentation itself - you need it to get the model into your graph in one step instead of ten.

What it does

It's a two-input, one-output node, and it's nearly featureless by design.

• model_name_or_path - default falcon_perception/model.safetensors. This is resolved relative to your ComfyUI models/ directory, which is why the default works if you drop the weights where the README says. If you keep the file somewhere else, you can type an absolute path and the node will use that instead - it checks the models-relative location first, then falls back to whatever you gave it.
• device - auto, cuda, or cpu. auto lets ComfyUI's model management decide; cuda pins it to your current torch device.

The single output is a FALCON_PERCEPTION_MODEL handle, and there's exactly one thing it plugs into: Falcon Perception Generate. This is a loader, not a pipeline - it does the loading so the generate node stays focused on segmenting.

How it works

Most ComfyUI loaders pull the architecture and the weights from a HuggingFace repo. This one bundles the entire architecture inside the pack, in a falcon_core/ folder, and loads only the weights from your file. Before building the model it patches torch.nn.attention.flex_attention for PyTorch 2.7+ (where BlockMask and AuxRequest got renamed or removed), then it loads the state dict with ComfyUI's own load_torch_file, moves the model to your chosen device, and sets it to eval mode. Because it's a 0.6B model, the whole thing is small enough that loading feels snappy.

That "no remote code, no HF cache" design isn't just convenience - it's the pack's stated reason for existing. No version-mismatch surprises between the checkpoint and whatever HF decides to serve you today.

Getting the weights

Download model.safetensors from the tiiuae/Falcon-Perception HuggingFace repo, then:

ComfyUI/
└── models/
    └── falcon_perception/
        └── model.safetensors

It's a 0.6B-parameter file, so roughly a gigabyte or two depending on precision - modest by ComfyUI standards. That's the entire "model install," which is the whole selling point.

Installing the pack

cd ComfyUI/custom_nodes
git clone https://github.com/222222222l/ComfyUI-Falcon-Perception
cd ComfyUI-Falcon-Perception
pip install -r requirements.txt

Or search "Falcon Perception" in ComfyUI Manager and restart. The requirements are torch>=2.5.0, transformers, einops, pycocotools, pillow. The torch floor is the one that bites - the architecture needs flex_attention, which only exists in 2.5+. If you see an ImportError about torch.nn.attention.flex_attention, that's your environment, not the node.

Gotchas worth knowing

• The bundled example workflow disagrees with the README on the folder name. The shipped example-workflow.json points at Falcon-Perception\model.safetensors, while the README says falcon_perception/model.safetensors. Both are valid if you create the matching folder - just pick one and keep the path and the folder in sync, or you'll get a FileNotFoundError.
• On-screen silence is normal. The loader reports progress to the console, not to the graph; if the node throws, the traceback lands in the terminal too.
• The README's own clone URL is a your-username placeholder. Use the real repo link above when installing manually.

One honest caveat: this pack is very new - a single commit, and at the time of writing there's basically no community usage or feedback anywhere yet. So the loader works, the source is tidy, but you're an early adopter. For a node whose entire job is "load one file, no drama," that's a reasonable risk to take.