🔧 Load CLIPSeg Models

No inputs, one job: Load CLIPSeg Models loads CLIPSeg, the model that lets you build a mask from a plain-language description instead of a color or a hand-drawn region - "the red hat," "the left hand," "the car" - rather than picking pixels by hue or clicking around a segmentation map. It's the same underlying idea that made the standalone biegert/ComfyUI-CLIPSeg node popular back in 2023 - feed an image and a text prompt, get a heatmap of where that phrase applies in the image - except here it's built into essentials as the loading step, feeding whatever downstream node in the pack consumes its output.

How it works

There's genuinely nothing to configure. No checkpoint path, no model picker - you drop the node down, and on first run it fetches the CLIPSeg weights and hands off a ready-to-use CLIP_SEG bundle. That bundle is a loader output, not something with a visible effect by itself; it only matters once it's wired into whatever text-prompted masking node in the pack actually consumes CLIP_SEG.

Inputs and outputs

No required or optional inputs. Output: CLIP_SEG.

Installing it

Via ComfyUI Manager, search "ComfyUI Essentials". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cubiq/ComfyUI_essentials

Restart. First time you actually run this node, expect a model download - give it a minute and a stable connection. The pack has been maintenance-only since April 2025, which mostly matters here in that CLIPSeg itself is a 2021-era model; don't expect this loader to get swapped for something newer.

Common issues & troubleshooting

Node runs fine but nothing seems to happen. That's expected - this is purely a loader. CLIP_SEG on its own doesn't produce a visible result; it only does anything once connected to the essentials node that actually consumes it downstream. If your graph ends here, you haven't wired it up yet.

First run hangs or errors out. Almost always the model download stalling or failing, not the node itself. Check your ComfyUI console for a stuck or failed fetch before assuming something's broken - a flaky connection on that first download is the usual cause.

Can't find the node that actually takes CLIP_SEG as input. It's not in this batch of nodes, but it exists elsewhere in the essentials pack - search your node list for anything mentioning CLIPSeg or segmentation-by-text once this loader's in place.