🅛🅣🅧 Gemma 3 Model Loader

This is the node that loads LTX-2's brain for reading prompts. LTX-2 dropped the old T5 text encoder for Gemma 3 12B, and this loader is what pulls it in and hands ComfyUI a CLIP object your conditioning nodes can use. If you're setting up an LTX-2 or LTX-2.3 workflow from scratch, you'll wire this in near the start - nothing downstream understands your prompt without it.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. Fair warning up front: the Gemma text encoder is the single most cursed part of the LTX-2 launch. It's huge (22GB-plus at full precision), and getting it to load and offload cleanly caused the majority of the out-of-memory errors people hit in the first weeks. So this humble-looking loader is where a lot of setups either succeed or fall over.

How it works

Despite the name, "CLIP" here is ComfyUI's generic label for a text-encoder slot - the actual model is Gemma 3, a language model, not OpenAI CLIP. The node reads the Gemma weights from disk plus a matching LTX component and wraps them into the object the rest of the graph expects. It also caps how much prompt it will process via a token limit, which keeps memory and speed sane on long prompts.

The inputs that matter

There are only three, and two of them are file pickers:

• gemma_path - points at your local Gemma 3 folder. The README is specific: download every file from google/gemma-3-12b-it-qat-q4_0-unquantized into models/text_encoders/gemma-3-12b-it-qat-q4_0-unquantized. If the dropdown is empty, that folder is missing or misnamed.
• ltxv_path - the matching LTX component the loader pairs Gemma with.
• max_length - the token cap (default 1024). Longer prompts need a higher value, but every extra token costs memory and time.

The single output is a CLIP, which feeds your text-encode / conditioning nodes exactly where a normal CLIP would.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. The node lives under lightricks/LTXV. The node code is tiny; the model is the download - grab the full Gemma 3 folder into the exact text_encoders path above before the picker will show anything.

Where people get burned

• Empty dropdown. By far the most common issue: the Gemma folder isn't where ComfyUI looks. It must be all of the files, in models/text_encoders/gemma-3-12b-it-qat-q4_0-unquantized. Partial downloads and renamed folders both produce a blank picker.
• OOM on load. The full-precision Gemma encoder is what pushes cards over the edge. The community's standard fixes: launch ComfyUI with --reserve-vram 4 (or higher), disable previews, and use a quantized Gemma (people run 4-bit or fp8 repacks - around 7GB instead of 22GB) if your card is tight.
• System RAM, not VRAM. For sub-32GB cards the encoder gets offloaded to system memory, so 64GB RAM is the comfortable floor. On 32GB you may hit page-file swapping, which is slow and hard on your disk.
• Reloading it every run. Constantly loading and unloading a 22GB encoder is painful. That's exactly the pain the pack's Gemma text-encoding/caching nodes were built to relieve - save the encoding once and reuse it, or run Gemma through Lightricks' API while the diffusion model stays local.