LongCat Avatar Text Encode

LongCat_Video_SM_Encode is the pack's text-conditioning node - it takes your prompt plus negative prompt and turns them into the te_cond conditioning the sampler consumes. It exists because Avatar 1.5 uses the LongCat-Video UMT5 text encoder, and unless you hand it a standard ComfyUI CLIP model it will happily load (or auto-download) that official text encoder itself. For most people the default "just works" path is the right one: don't connect clip, let the node handle the encoder, and spend your effort on the prompt instead.

Why the prompt matters more than usual

LongCat is a video model, so the text is doing double duty - it sets the character and the scene, and the reference image is only a starting anchor. The most common failure people hit ("the reference blinks / doesn't carry through / the face drifts") usually comes down to a vague prompt. Write for the person in your reference: appearance, clothing, what they're doing while speaking, the setting. The node's default prompt is a good model:

A western man stands on stage under dramatic lighting, holding a microphone close to their mouth. Wearing a vibrant red jacket with gold embroidery, the singer is speaking while smoke swirls around them, creating a dynamic and atmospheric scene.

The default negative prompt is already a solid quality-badge blocklist ("static, blurred details, extra fingers, poorly drawn hands…") - keep it.

The inputs

• prompt / negative_prompt: the multiline fields you'll actually edit.
• clip (optional): connect a ComfyUI Load CLIP (with umt5_xxl_fp8_e4m3fn_scaled.safetensors in models/clip/) to force the single-file fallback path. Leave it empty to use the native official UMT5 encoder - the README's preferred path, because it matches the official pipeline's conditioning exactly.
• auto_download_missing_text_encoder (default true): when on and no clip is connected, the node downloads only the official tokenizer/ and text_encoder/ assets from meituan-longcat/LongCat-Video into ComfyUI/models/longcat/LongCat-Video/. It's a bounded manifest - no arbitrary URLs, no token box. Turn it off if you've placed the files yourself and want zero network activity.
• text_encoder_root (default LongCat-Video): subdirectory under models/longcat/ where the official encoder lives.
• offload_device: cpu (default) keeps the big UMT5 encoder out of VRAM; cuda is faster but needs the headroom. Ignored when clip is connected.

Output is a single te_cond (CONDITIONING) - wire it to the sampler's te_cond socket.

Install

This node is part of the ComfyUI-LongCat-Avatar pack, so install once and all nine nodes appear:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-LongCat-Avatar
cd ComfyUI-LongCat-Avatar
pip install -r requirements.txt

Or search ComfyUI-LongCat-Avatar in ComfyUI Manager, then restart. CUDA GPU required - the whole pack is CUDA-only.

Where people get burned

If the node stays on the native path, the text encoder is bf16 and large; leave offload_device on cpu unless you've got VRAM to spare. And if you toggle auto_download_missing_text_encoder off before placing the files, the node will error on a missing layout - that toggle is a convenience, not a requirement. One more: if you do connect a clip that isn't the UMT5 model, conditioning silently takes the fallback path and results may look off. Use the UMT5 repackage, or trust the native path.