ComfyUI_Face_similarity

A custom ComfyUI node that compares face identity between two images using InsightFace embeddings. Useful for verifying that a generated image preserves the identity of a reference person.

By phobod·Created 2 months ago·Updated 2 months ago· 0

phobod/ComfyUI_Face_similarity

Nodes—

On cloudLocal install

Stars0

Updated2 months ago

Readme

ComfyUI Face Similarity (InsightFace)

A custom ComfyUI node that compares face identity between two images using InsightFace embeddings.
Useful for verifying that a generated image preserves the identity of a reference person.

How it works

InsightFace extracts a 512-dimensional face embedding from each image.
The node computes the cosine similarity between the two embeddings and returns a score in the range [-1, 1], where 1 means identical and 0 means unrelated.

Installation

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

CPU users: edit requirements.txt and swap onnxruntime-gpu for onnxruntime before running pip install.

The InsightFace model weights are downloaded automatically on first use to ~/.insightface/models/.

Node inputs

| Input | Type | Description | |---|---|---| | reference_image | IMAGE | Source photo of the person | | generated_image | IMAGE | Output image to verify | | model | dropdown | InsightFace model to use (see below) | | detection_size | dropdown | Face detector input resolution: 640, 320, 160 | | threshold | float 0–1 | Minimum score to consider the same person |

Node outputs

| Output | Type | Description | |---|---|---| | image | IMAGE | Pass-through of generated_image | | similarity_score | FLOAT | Cosine similarity score | | verdict | STRING | Human-readable result string | | is_same_person | BOOLEAN | True if score ≥ threshold |

Available models

| Model | Size | Notes | |---|---|---| | buffalo_l | ~500 MB | Most accurate, recommended default | | buffalo_m | ~200 MB | Good balance of speed and accuracy | | buffalo_s | ~100 MB | Faster, slightly less accurate | | buffalo_sc | ~100 MB | Small, single-channel variant | | antelopev2 | ~300 MB | High quality alternative |

All models are downloaded automatically from the InsightFace model zoo.

Score reference

| Score | Interpretation | |---|---| | > 0.6 | Confidently the same person | | 0.4 – 0.6 | Likely the same, worth checking | | < 0.4 | Identity has drifted |

A threshold of 0.6 is a good starting point. Adjust based on your use case.

Example workflow

[Reference Image] ──┐
                    ├──► [Face Similarity] ──► similarity_score
[Generated Image] ──┘         │               verdict
                          is_same_person ──► (route / filter / re-sample)

Connect is_same_person (BOOLEAN) to a routing node such as ImpactConditional to automatically re-generate if identity drifts.

Requirements

Python 3.9+
ComfyUI (any recent version)
insightface>=0.7.3
onnxruntime>=1.16.0 (CPU) or onnxruntime-gpu>=1.16.0 (CUDA)
opencv-python>=4.8.0
numpy>=1.24.0

License

This repository's code is licensed under the MIT License.

Note on model licenses:
The InsightFace models downloaded at runtime (buffalo_*, antelopev2)
are licensed under CC BY-NC 4.0
and are not permitted for commercial use.
For commercial projects, verify licensing with the
InsightFace authors.