EmbeddingGemma 2 for ComfyUI
google/embeddinggemma-2, repackaged as one file for ComfyUI, for use with the ComfyUI-EmbeddingQuery node pack (semantic search over your input/output folders: text and images in one embedding space).
Install
Download text_encoders/embeddinggemma_2_bf16.safetensors into ComfyUI/models/text_encoders/.
What changed from the original
- Text and vision towers only: the audio tower is dropped (β440M parameters, 0.91 GB).
- The tokenizer (
tokenizer.json) is embedded in the file as a uint8tokenizer_jsontensor, the way ComfyUI's Gemma 4 files carry theirs. - Weights are unchanged: same bf16 tensors, same key names as Google's
model.safetensors.
Outputs match Google's model through transformers (text cosine 1.0000, images 0.9999).
License
Apache 2.0, the same as the original. EmbeddingGemma 2 is by Google; see the original model card for its intended use, limitations and citation. This repo only repackages the weights.
Model tree for nynxz/embeddinggemma-2
Base model
google/embeddinggemma-2