📥 LTXAVTextEncoder Download

LTX-2 is Lightricks' audio-video model - the one that generates synchronized sound and video in a single pass, and found its killer app in lipsync and music videos. It's fast and genuinely novel, but it has a setup tax, and this node exists to pay part of it: it fetches the text encoder LTX-2 needs and hands you a CLIP output ready to plug into an LTX workflow.

Here's why that's not trivial. LTX-2 doesn't use a normal CLIP text encoder - it pairs its DiT with Gemma 3 12B, a full language model, as the prompt encoder. That thing is enormous (the full weights are ~22GB), and on launch it OOM'd nearly every consumer card until offloading patches landed. This node grabs a quantized build (gemma_3_12B_it_fp4_mixed) that's been squeezed down to fit, and gives you a CPU-offload option, so you can actually run LTX-2 without a datacenter GPU.

How it works

It's a download-and-load helper. It pulls the encoder file and loads it as a CLIP object - the type LTX-2's conditioning nodes consume. The fp4_mixed in the name is the important part: it's a mixed 4-bit quantization of Gemma 3 12B, which is what makes the 12-billion-parameter encoder loadable on a normal card instead of needing 24GB+ just for the text side. The device option lets you push it onto the CPU to spare VRAM for the actual video generation.

The inputs that matter

• CLIP_name - the text encoder to load. There's one option here, gemma_3_12B_it_fp4_mixed.safetensors - the quantized Gemma 3 that LTX-2 wants. This is the field that matters.
• device - default (put it where ComfyUI normally would, i.e. GPU) or cpu (offload it). On a tight VRAM budget, cpu is your friend: it keeps the big encoder off the GPU so there's room for LTX-2 itself. It's slower, but "slower" beats "out of memory."
• Ckpt_name - selects the paired model from the pack's checkpoint list.

The single output is CLIP - wire it into your LTX-2 text-encode/conditioning nodes in place of a normal CLIP loader.

Common issues

Out of memory when loading. This is the LTX-2 story. The Gemma encoder is huge even quantized. Set device to cpu to offload it - that's exactly what the option is for, and it's how most people run LTX-2 on sub-24GB cards. The KB notes that on release, offloading the text encoder was the difference between "runs" and "OOM on nearly every card."

Download fails on Windows/macOS. aria2c isn't installed by default there, and SDVN's download nodes need it. Install it manually per the README.

The CLIP won't connect to my LTX nodes / quality is off. This encoder is specific to the LTX-2 family. It's not a drop-in for SD, SDXL, or Flux text encoding, and vice versa - those want their own CLIP models. Make sure the rest of your graph is an LTX-2 pipeline (and remember LTX-2 punishes short prompts - it wants long, specific prose, so write accordingly).

Slow generation with cpu offload. Expected trade-off. CPU offload frees VRAM at the cost of speed; if you have the headroom, default is faster. Pick based on your card.

Installing SDVN_Comfy_node

ComfyUI Manager: search "SDVN_Comfy_node" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node

Then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from the ComfyUI root and restart. Install aria2c on Windows/macOS for the download. The node is under 📂 SDVN / 📥 Download.