LTXV2 Text Encoder Loader (MultiGPU)

If you've OOMed on LTX-2, odds are the culprit was the text encoder, not the video model. LTX-2 pairs its 19B video DiT with a Gemma 3 12B text encoder that weighs in around 22GB in fp16 - bigger than most people's entire card. The first weeks after the January 2026 open-weights release, this thing was responsible for the majority of OOM reports on every GPU under 48GB, including 4090 and 5090 rigs. This node is the direct answer: it loads Gemma and lets you tell ComfyUI to keep it on the CPU while the video model hogs the GPU.

It's the headline node of the ComfyUI-LTX2-MultiGPU pack. Author nathandreamfast built the pack after OOMing a 4090+5090 box, and his writeup is explicit that the text encoder was "the main issue" - the moment he could push Gemma to CPU, his 10-second 1080p generations started working.

How it works

It's a thin wrapper around ComfyUI's core LTXAVTextEncoderLoader with a device dropdown bolted on. But this node gets a different kind of dropdown than the pack's other loaders: it routes through the text-encoder half of ComfyUI's model management, so the choice controls where the Gemma weights land and encode, independent of the video model. Load it on cpu and your system RAM holds the encoder; load on cuda:1 and it takes a slice of your second card's VRAM.

That "decent CPU" caveat in the README is real. The encoder lives in system RAM, so 32GB+ is the practical floor. This is the pack's whole strategy: Gemma on CPU, video model on one GPU, VAE decode on the other.

The inputs and outputs that matter

• text_encoder - the Gemma 3 encoder file from your models/text_encoders folder. The fp16 one is huge; a quantized fp8 build is worth hunting down if RAM is tight.
• ckpt_name - yes, it wants a checkpoint too, even though it's loading an encoder. That's where the text encoder's config comes from, so point it at your LTX-2 checkpoint.
• device - the dropdown. cpu is the move if your GPUs are OOMing.

One output, CLIP (the CLIP type ComfyUI uses to stand in for the text encoder) → straight into CLIPTextEncode for your prompt.

Installing it

Same pack, same one-time install: ComfyUI Manager → search "ComfyUI-LTX2-MultiGPU" → install → restart. Or:

cd ComfyUI/custom_nodes/
git clone https://github.com/dreamfast/ComfyUI-LTX2-MultiGPU

Restart ComfyUI, done. No requirements.txt - ComfyUI's own deps cover everything. You need a ComfyUI with native LTX-2 support and the Gemma file downloaded into text_encoders. The bundled workflow also ships an ollama prompt enhancer, which is apt: LTX punishes short prompts, and the community standard is to let an LLM write them - either the built-in Gemma enhancer or an external one.

Common issues & troubleshooting

• OOM during encode - move device to cpu. That's literally the node's reason to exist. The standard LTX-2 launch-day mitigations still help too: ComfyUI's --reserve-vram 4 flag and previews disabled.
• Nothing in the text_encoder dropdown - the Gemma file isn't there. Download it into ComfyUI/models/text_encoders/; it doesn't come with the pack.
• It refuses to load without a checkpoint - by design. The encoder needs the checkpoint's config, and a lot of people are surprised by the second required input. Point it at your LTX-2 ckpt.

One honest caveat: this node only helps if you have the RAM for a 22GB encoder. If system RAM is tight too, reach for a quantized Gemma build instead - same node, smaller file. This is the single highest-leverage node in the pack, because it attacks the one component that actually broke people's rigs.