Load LTXV Audio Text Encoder

If you're trying to get LTX-2 to generate video with sound - a talking head, a lipsync, a music video - this is the node your prompt flows through first. It's LTX-2's text encoder loader, and the encoder it loads is not the little CLIP you're used to. It's Gemma 3 12B, a full-size language model from Google, the thing that lets LTX-2 parse long natural-language scene descriptions instead of tag soup. That's also, in one sentence, why LTX-2's first weeks on the subreddit were full of OOM errors.

How it works

Two model inputs, one output:

• text_encoder - the Gemma 3 12B weights, from ComfyUI/models/text_encoders. The node description is blunt about the recipe: ltxav: gemma 3 12B.
• ckpt_name - the LTX-2 checkpoint, from ComfyUI/models/checkpoints. It's needed because the LTX-2 text path isn't just Gemma alone; the checkpoint carries the connector that projects Gemma's output into the video model's conditioning space.
• device - default or cpu. This one is a lifeline: on a 16GB card, running the 22GB fp16 Gemma on CPU while the 19B transformer works the GPU is a legitimate way to fit the whole thing.

The output is a CLIP object - same wire type as any other text encoder - which goes into CLIP Text Encode (Prompt) and from there into conditioning, just like an SD workflow. Mechanically the loader builds an LTXV-typed CLIP from those two files and hands it over.

The Gemma problem is the whole story

The community's LTX-2 experience in January 2026 was dominated by this one component. Gemma 3 12B is enormous: about 22.7GB in fp16, larger than many video models people had ever run. It was the leading cause of launch-day OOMs on almost every card under 48GB, and it was slow to load on top of that. The workarounds that stuck:

• Run the encoder on CPU (device=cpu), which trades load time and speed for the ability to fit at all.
• Use a quantized Gemma - fp8 repacks and Unsloth's 4-bit builds (roughly 7GB) were the community favorites.
• Bypass the built-in prompt enhancer entirely. LTX-2 pipes your prompt through Gemma to "enhance" it before encoding, and the community's verdict was that the enhancer writes embarrassingly basic prompts. A common trick was to show what Gemma was generating, sigh, and write your own prompt instead - you can still feed that to this loader; you're just not using the enhancer's rewrite.

Two other habits worth stealing. LTX is famously prompt-sensitive: short prompts reliably underperform, so write long, specific scene descriptions (or have an LLM write them - yes, the irony is not lost on anyone). And for audio specifically, keep LoRA strengths disciplined; people report audio going "very low" once the LoRA strength passes a certain point, and that's a model quirk, not your loader.

Where it fits

It ships with ComfyUI core, added alongside the rest of the LTX-2 audio nodes when native support landed in January 2026. The full LTX-2 audio-video graph is: Load LTXV Audio Text Encoder → CLIP Text Encode → conditioning, Load LTXV Audio VAE on the audio decode side, and the model itself loaded separately. It's a newer, bigger, more temperamental cousin of Load CLIP - same job, more VRAM, much better at understanding a real sentence.