Anymatix LTX Audio Text Encoder Loader

LTX found its real niche when it stopped being just another video model and became one that generates the audio to go with the video - dialogue that matches the mouth movements, music that fits the motion. The LTX-2 family's whole pitch is synchronized audio-video, and that means the text conditioning side of the model isn't just one encoder. It's a two-part load: a text encoder for the language side and the audio-video checkpoint itself. AnymatixLTXAVTextEncoderLoader handles the first part - it builds the CLIP object that tells an LTX audio-video model what's being said - and it does it with the pack's signature string inputs instead of dropdowns.

The node mirrors ComfyUI's own LTX audio-video text encoder loader (it literally delegates to the upstream node's execute), which keeps the behavior identical to stock while changing where the names come from. Both text_encoder and ckpt_name are strings, resolved by basename against your model folders, so files already present load from plain filenames and URL-provisioned files load via AnymatixFetcher. There's also a device selector (default or cpu) - genuinely useful, because the text encoders these models want can be chunky and you don't always want them eating GPU memory during the encode.

What's what

• text_encoder (STRING) - the language model file (LTX's audio pipeline uses a T5-family encoder).
• ckpt_name (STRING) - the audio-video checkpoint this text encoder conditions with; the loader needs to know which model it's talking to.
• device - default or cpu.
• Output: CLIP → the conditioning side of the LTX audio-video sampler chain.

When you'd use it

If you're generating talking-head or music-video style clips on an LTX-2-class model where audio comes out of the same pass as the video, this loader is on the path - wire its CLIP into the model's prompt conditioning alongside the usual text prompts. It sits next to the LTX audio VAE loader in the same workflow; the text encoder feeds meaning, the VAE handles the audio waveform. Pure text-to-video LTX without audio? You don't need this node.

Install

The pack install is the usual one:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

restart, or use ComfyUI Manager (search "anymatix-comfy-nodes"). LTX files are large - budget for the checkpoint plus encoder weights, and let AnymatixFetcher provision them if you want the workflow to self-install on a fresh machine.

Issues to expect

The mismatch trap is real here: LTX released several audio-video versions, and a text encoder or checkpoint from the wrong generation will refuse to cooperate in confusing ways. Keep the encoder and checkpoint from the same model card. And remember LTX's hard rule that frame counts have to land on its 8n+1 grid - the loader is happy, but your empty-latent node will reject an off-grid clip count before sampling ever starts.