MotifVideo Text Encode

MotifTextEncode is the pack's answer to a minor annoyance: in a Motif-Video workflow you need two separate CONDITIONING objects - positive and negative - before KSampler will run, and the standard way to get them is two CLIPTextEncode nodes doing the same dance twice. This node does it in one step. Feed it a CLIP and two strings, and out come both conditionings, already named positive and negative so you don't have to remember which wire is which.

It's a convenience node, nothing more. If you'd rather use core CLIPTextEncode twice, you can - the CLIP from MotifTextEncoderLoader is fully compatible. But the convenience is real: one less node to place, and the negative prompt ships with a sensible default so you don't stare at a blank field wondering what to type.

What it does

Under the hood it's exactly what two CLIPTextEncode nodes would do. It tokenizes the positive string and the negative string, then runs each through encode_from_tokens_scheduled on the CLIP you gave it. The outputs are a positive CONDITIONING and a negative CONDITIONING, both compatible with KSampler.

The inputs

• clip - the CLIP from Load MotifVideo Text Encoder. This is the T5Gemma2 encoder, and it behaves differently from CLIP, so prompt accordingly (more below).
• text - your positive prompt, multiline. Default empty.
• negative_prompt - multiline, default blurry, low quality, distorted, artifacts.

Two outputs, positive and negative, wire into KSampler's positive and negative slots. In image-to-video mode they feed into MotifVideo Image Encode instead, which injects the start image and passes the pair through to the sampler.

Installing it

It's one of five nodes in the official Motif Technologies pack, so the install is pack-wide:

cd ComfyUI/custom_nodes
git clone https://github.com/MotifTechnologies/ComfyUI-MotifVideo2B.git
pip install -r ComfyUI-MotifVideo2B/requirements.txt

Manager search won't surface it yet (the registry PR is planned but not merged, and Manager's Git-URL install is blocked under the default security level), so the clone is the dependable route. The pack's example workflows trigger an automatic download of the three weight files on first load if your ComfyUI is current enough - v0.18.0 was the version they tested against.

Where people get burned

• Writing tags instead of prose. This is the one that actually matters. The encoder is a language model, not CLIP - T5Gemma2 understands full sentences, and tag-soup prompting wastes that. A woman in a red coat walks down a snowy street at dusk, snow falling, cinematic lighting will get you much closer to the intent than 1girl, red coat, snow, street, best quality. It's the same advice as every recent LLM-encoder model, and it's the difference between "the model ignored my prompt" and "the model did what I asked."
• The negative prompt exists here for a reason. Flow-matching models made negative prompts feel obsolete, but Motif-Video's recommended settings run a real CFG of 8.0, so the negative conditioning is doing work. Keep the shipped default unless you have a reason - it's already tuned for the common failure modes.
• Blank positive, weird results. The default text is empty. KSampler will happily run a mostly-unconditioned video and you'll wonder what went wrong. Type something first.

The honest take: this is the least interesting node in the pack and the most important to get prompting right on. It doesn't load models, it doesn't make sampling decisions - it just converts your words into the vectors the sampler actually obeys, twice, in one box.