woctordho/

comfyui-gemma-3-4b-it

Single file version of gemma-3-4b-it, used by NewBie-Image. It can be directly loaded in ComfyUI using the Load CLIP node with type lumina2.

Its format is similar to https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/text_encoders/gemma_2_2b_fp16.safetensors . Only language_model in the original gemma3-4b-it included here, while multi_modal_projector and vision_tower are not. The official NewBie-Image actually has the ability to encode images as prompts, and I'm leaving it for future work.

The tokenizer is included as the tensor spiece_model.