CLIPSeg Model Loader (Salt)

A one-job loader, and the job is worth knowing about even if the node itself is boring: CLIPSeg Model Loader (Salt) pulls in CLIPSeg, the model that lets you mask an image by describing what you want masked - "the dog," "the red shirt" - instead of drawing it by hand every single time. It's a genuinely different tool from the background-removal models you might already know (BiRefNet, InSPyReNet): those find "the salient subject" automatically, CLIPSeg finds whatever you tell it to, anywhere in the frame, including things that aren't the main subject at all.

How it works

CLIPSeg does zero-shot segmentation: it compares a CLIP text embedding of your prompt against the image and produces a heatmap of where that concept lives, with no fine-tuning or per-object training needed. It's been a known quantity in ComfyUI since 2023 - the standalone biegert/ComfyUI-CLIPSeg pack is where most people first met it, built for exactly the use case of skipping manual masking in batch inpainting jobs. Salt's version here is purely a loader; it hands off a CLIPSEG_MODEL object for whatever downstream node in the pack actually runs the segmentation and turns it into a mask.

The inputs and outputs that matter

The only input is model (STRING, default CIDAS/clipseg-rd64-refined) - a Hugging Face model ID. That default is the original, standard CLIPSeg checkpoint; leave it unless you specifically know of an alternate fine-tune you want to load instead.

Output is clipseg_model (CLIPSEG_MODEL), which feeds directly into whatever node downstream turns text + image into an actual mask.

How to install it

ComfyUI Manager: search SaltAI-Open-Resources or SaltAI, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/get-salt-AI/SaltAI

then restart. One thing worth flagging that the README doesn't spell out: the model weights aren't bundled with the pack. The first time this node actually runs, it downloads CIDAS/clipseg-rd64-refined from Hugging Face into your local cache - so the first run needs internet access, and if you're air-gapped or behind something that blocks huggingface.co, it'll fail right there rather than later at the masking step.

Common issues & troubleshooting

Import errors on load. CLIPSeg is built on Hugging Face's transformers package. Most ComfyUI installs already have it from other packs, but version conflicts are a known, recurring pain point in this ecosystem - custom node packs share one Python environment with no isolation, so one pack pinning transformers>=4.40 and another pinning transformers==4.30 can silently break each other. If this node fails to import right after installing something else, that's the first thing to check.

Download failure, not a ComfyUI error. If the model ID got typo'd, or you're offline on first run, you'll see a Hugging Face download failure in the traceback, not a Salt- or ComfyUI-specific message - read it, it usually says exactly what went wrong.

Slow first run, fast after. The download only happens once; subsequent runs load from your local HF cache and are fast. If every run feels slow, check that your cache directory is actually persisting between sessions (a container or ephemeral environment wiping the cache will make it re-download every time).