AP ClipSEG Light
A lightweight ComfyUI custom node pack that generates segmentation masks from plain-text prompts using CLIPSeg.
Features
- Text-driven masking — describe what you want masked in plain language (e.g.
face,sky,person) - Multi-prompt averaging — separate prompts with
|to blend multiple heatmaps:face | head | eyes - Soft or hard masks — output a feathered sigmoid heatmap or a crisp binary mask
- numpy 2.x safe — all image I/O goes through torch/PIL, no numpy array ops
- Auto model caching — models are kept in GPU/CPU memory after the first load; no reload cost on subsequent runs
- numpy 2 / TF crash fix — automatically stubs broken TensorFlow imports caused by transformers ≥ 4.50 on numpy 2.x systems
- Mask Trim utility — trim any mask by percentage along X/Y axes with feathered edges
Requirements
| Package | Minimum version |
|---|---|
torch |
any recent |
torchvision |
≥ 0.15 |
transformers |
any recent |
Pillow |
any recent |
Models are downloaded automatically from the HuggingFace Hub on first use:
| Model | Size | Quality |
|---|---|---|
clipseg_rd64 (default) |
~350 MB | Higher |
clipseg_rd16 |
~100 MB | Faster |
Installation
# Option A – ComfyUI Manager (search "AP ClipSEG Light")
# Option B – manual
git clone https://github.com/adampolczynski/ComfyUI_AP_ClipSEG_Light \
ComfyUI/custom_nodes/AP_ClipSEG_Light
pip install transformers Pillow
Node: AP CLIPSeg Text Mask
| Input | Type | Description |
|---|---|---|
image |
IMAGE | Input image batch |
prompt |
STRING | Text description of the region to mask. Use | to separate multiple prompts |
threshold |
FLOAT (0.01–0.99) | Heatmap threshold. Lower → larger mask, higher → tighter mask |
smooth_radius |
INT (0–32) | Gaussian blur radius applied before thresholding |
soft_mask |
BOOLEAN | True = feathered sigmoid output, False = binary mask |
invert |
BOOLEAN | Invert the output mask |
model |
ENUM | clipseg_rd64 or clipseg_rd16 |
device |
ENUM | auto / cuda / cpu |
unload_after_run |
BOOLEAN | Free model VRAM after each run |
Output: MASK — float32 tensor [B, H, W] in [0, 1]
Node: AP Mask Trim
Trims a mask by percentage along the X and/or Y axis. The output canvas size is unchanged — trimmed pixels fade to black.
| Input | Type | Description |
|---|---|---|
mask |
MASK | Input mask |
crop_x |
INT (-100–100) | Trim from right (positive) or left (negative). Example: 30 removes rightmost 30 % |
crop_y |
INT (-100–100) | Trim from bottom (positive) or top (negative). Example: -40 removes top 40 % |
auto_align |
BOOLEAN | When True, trim axes rotate to match the mask's principal orientation (PCA). crop_y trims the mask's own "bottom" regardless of rotation |
feather |
INT (0–200) | Soft-edge width in pixels at the cut boundary |
mask_dilate |
INT (0–64) | Expand the trimmed mask outward by this many pixels |
mask_blur |
INT (0–64) | Gaussian blur applied to the final mask |
Output: MASK — trimmed float32 tensor, same dimensions as input
Tip: chain AP CLIPSeg Text Mask (prompt=hair) → AP Mask Trim (crop_y=-50, feather=20) to isolate the top of the head.
License
MIT — see LICENSE.