Gemma Loader

Google's Gemma2 2B, doing text-encoder duty instead of the usual CLIP or T5, is the somewhat unusual choice Nvidia's Sana models made for their text conditioning - and this node is what loads it. It's a small model as language models go, which is part of the point: Sana was built around the idea that you don't need a giant encoder to get good prompt following, echoing the same "small but capable" philosophy that made PixArt stand out before it.

Convenient bonus: unlike most model-loading nodes in this pack, you don't have to hunt down and place a file yourself first. Point this node at a model name and, per the pack's own README, "it should automatically download the requested model from Huggingface" the first time you run it.

Inputs. Three, all required. model_name picks which Gemma checkpoint to pull, with three options: google/gemma-2-2b-it (the vanilla Google release), Efficient-Large-Model/gemma-2-2b-it (the Sana team's own mirror), and unsloth/gemma-2-2b-it-bnb-4bit - a pre-quantized 4-bit version. The README's own recommendation is worth repeating verbatim: "Recommended to use the 4bit quantized model on CPU when low on memory," which points squarely at that third option if VRAM or RAM is tight. device sets where it loads (auto, cpu, or cuda - default cpu), and dtype controls precision, with default, auto (comfy), BF16, FP32, FP16, FP8 E4M3, and FP8 E5M2 on offer.

Output. GEMMA - a purpose-built type, not CLIP or T5. Wire it into Gemma Text Encode, the only node in this pack built to consume it.

A word of caution before you start. This node is part of Sana support in this pack, and the pack's own README puts a caution box on the whole Sana section: a lot of people have run into trouble with it, and the maintainer's own suggestion is to try the fork maintained by the Sana team instead, since it auto-downloads models more reliably. If you're setting up Sana specifically (rather than adding it onto an existing PixArt/DiT install where switching to the fork is inconvenient), it's worth reading that caution before you sink time here.

Installing the pack. ComfyUI Manager: search "Extra Models for ComfyUI." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI_ExtraModels

then pip install -r requirements.txt in your venv and restart - this pulls in transformers and bitsandbytes, both of which this node needs (the latter specifically for the 4-bit variant).

Troubleshooting. First-run downloads can be slow or stall on a flaky connection since HuggingFace is being hit live rather than from a file you already placed - if the node seems to hang the first time, that's likely what's happening, not a crash. If you pick the unsloth 4-bit option and hit errors, double-check bitsandbytes actually installed correctly for your platform; Windows in particular sometimes needs a newer build than what ships by default (python -m pip install bitsandbytes). And if cuda loading fails oddly, falling back to cpu - slower, but the more battle-tested path per the README's own recommendation - is a reasonable first troubleshooting step rather than a last resort.