ComfyUI-YOLOE26

ComfyUI-YOLOE26

English | 简体中文

Open-vocabulary prompt segmentation nodes for ComfyUI powered by Ultralytics YOLOE-26.

Features

• Load YOLOE-26 models and reuse them across your workflow
• Segment objects using text prompts like person, car, red apple
• Get merged masks, per-instance masks, or per-class masks
• Structured JSON metadata for downstream automation
• Mask refinement without re-running detection
• Select the best instance from multiple detections

Installation

Via ComfyUI Manager (Recommended)

Search for YOLOE-26 in ComfyUI Manager and click Install.

Manual Installation

Clone this repository into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-YOLOE26.git
pip install -r ComfyUI-YOLOE26/requirements.txt

Restart ComfyUI and look for YOLOE-26 nodes in the node menu.

Quick Start

1. Add YOLOE-26 Load Model node and select a model
2. Add YOLOE-26 Prompt Segment node
3. Connect an image and enter a prompt like person
4. Run the workflow

The model will auto-download on first use. Recommended settings for your first test:

| Parameter | Value | |-----------|-------| | model_name | yoloe-26s-seg.pt | | device | auto | | prompt | person | | conf | 0.1 | | iou | 0.7 |

Nodes

YOLOE-26 Load Model

Load and validate a YOLOE-26 model.

Inputs:

• model_name — Model filename (dropdown shows official models plus local .pt files)
• device — auto, cpu, cuda, cuda:N, or mps
• auto_download — Automatically download missing models (default: true)
• offload_to_cpu — Move the model to CPU after each segmentation node finishes, freeing VRAM for diffusion models (default: false)

Outputs:

• model — YOLOE_MODEL for use with other nodes

Notes:

• Auto-downloaded weights are verified against pinned SHA256 digests before loading.
• The first text-prompt inference downloads the MobileCLIP text encoder (~250 MB) through Ultralytics and may install its CLIP dependency, so the first run needs network access (and git).
• Workflows saved with older versions of this pack that carry (local)/(downloadable) suffixes in model_name keep working.

YOLOE-26 Prompt Segment

Run prompt-based segmentation and get an annotated preview image.

Inputs:

• model — YOLOE_MODEL from Load Model
• image — Input image(s)
• prompt — Comma-separated class names (e.g., person, car, dog)

Optional:

• conf — Confidence threshold (default: 0.1)
• iou — IoU threshold (default: 0.7)
• max_det — Max detections (default: 300)
• mask_threshold — Mask binarization threshold (default: 0.5)
• imgsz — Inference size (default: 640)

Outputs:

• annotated_image — Preview image with detections drawn
• mask — Merged binary mask
• detection_count — Number of detections

YOLOE-26 Instance Masks

Get one mask per detected instance.

Outputs:

• instance_masks — MASK batch with one mask per instance
• instance_metadata_json — JSON with detection details
• count — Number of instances

YOLOE-26 Class Masks

Get one merged mask per prompt class.

Outputs:

• class_masks — MASK batch with one mask per class
• class_metadata_json — JSON with class-to-mask mapping
• output_mask_count — Number of class masks

YOLOE-26 Detection Metadata

Get structured detection data without generating images.

Outputs:

• metadata_json — JSON with boxes, scores, mask areas, and class names
• detection_count — Number of detections

YOLOE-26 Refine Mask

Post-process masks without re-running detection.

Methods:

• threshold — Binary thresholding
• open — Morphological opening
• close — Morphological closing
• dilate — Expand masks
• erode — Shrink masks
• largest_component — Keep only largest connected component
• fill_holes — Fill holes in masks

YOLOE-26 Select Best Instance

Select a single best mask from instance outputs.

Selection Modes:

• highest_confidence — Highest detection confidence
• largest_area — Largest mask area
• confidence_then_area — Confidence first, then area as tiebreaker

Model Locations

Place .pt files in one of these directories:

• ComfyUI/models/ultralytics/segm/
• ComfyUI/models/ultralytics/bbox/
• ComfyUI/models/ultralytics/
• ComfyUI/models/yoloe/

Directories configured through extra_model_paths.yaml under the ultralytics, ultralytics_segm, ultralytics_bbox, or yoloe keys are also picked up. After adding files, press r in ComfyUI to refresh the model list.

Auto-download supports: yoloe-26n-seg.pt, yoloe-26s-seg.pt, yoloe-26m-seg.pt, yoloe-26l-seg.pt, yoloe-26x-seg.pt (downloaded directly into models/ultralytics/segm/ and SHA256-verified)

Example Workflows

See examples/ for workflow JSON files:

| File | Description | |------|-------------| | basic_api_workflow.json | Minimal quick-start workflow | | all_nodes_showcase_api.json | All 7 nodes demonstrated | | practical_prompt_segment_api.json | Basic segmentation for inpainting | | practical_best_instance_api.json | Best instance selection | | practical_class_masks_api.json | Per-class mask routing | | practical_refine_mask_api.json | Mask post-processing | | practical_detection_metadata_api.json | Structured detection metadata | | practical_batch_multi_class_api.json | Batch input with multi-class prompts |

Prompt Format

Use comma-separated class names:

person
person, car, dog
red apple, green bottle

Requirements

• ComfyUI
• Python 3.10+
• PyTorch
• ultralytics >= 8.3.200, < 9.0.0

Tested With

Validated end-to-end on 2026-07-04:

• ComfyUI 0.26.2 (Windows 11, NVIDIA RTX 3090)
• Python 3.11.9
• PyTorch 2.8.0+cu128
• ultralytics 8.4.26 (unit suite also run against 8.4.41)

Behavior Notes

• Masks are produced at the original image resolution (retina_masks), so they align with detection boxes and downstream inpainting/compositing nodes.
• Recent Ultralytics releases return already-binarized masks; in that case mask_threshold values below 1.0 have no additional effect.
• Select Best Instance picks a single best mask across the whole batch when the metadata spans multiple input images.

Troubleshooting

Model not loading?

• Check the .pt file is in a supported model directory
• If using auto_download, verify network access

Download fails SHA256 verification?

• The file is removed automatically; retry once. If it fails for every model, Ultralytics may have republished the release assets — update this node pack or download the weights manually into a supported model directory.

First text-prompt run is slow or fails offline?

• The first prompt inference downloads the MobileCLIP text encoder (~250 MB) and may install the Ultralytics CLIP dependency (requires git). Run it once with network access.

Inference failing?

• Try reducing imgsz if running out of GPU memory
• Enable offload_to_cpu on the Load Model node when sharing VRAM with diffusion models
• Ensure your ultralytics version supports from ultralytics import YOLOE

No detections?

• Lower the conf threshold
• Check that your prompt matches objects in the image

License

MIT License for this repository's original code and documentation.

Ultralytics and YOLOE model weights are separate third-party dependencies with their own licenses.

Credits

This project references implementation ideas from spawner1145's prompt_segment.py, used with permission.