Load MotifVideo Text Encoder

Every MotifVideo workflow starts here, whether it looks like it or not. Before KSampler can denoise a single frame it needs conditioning, and conditioning means turning your words into vectors. That's this node's entire job, and it's not a CLIP. Motif-Video 2B reads prompts with a T5Gemma2 encoder - a T5-family model built on Gemma-2-style architecture - which is a different prompting animal than the CLIP you're used to from SD 1.5 or SDXL. No more tag soup and comma stacking; write prose, because the encoder is a language model and it understands sentences.

Motif Technologies ships this node as part of their official ComfyUI pack for the Motif-Video 2B text-to-video model. It's Apache 2.0, no API, no key, no server. Just weights and a graph.

What it does

MotifTextEncoderLoader loads motifvideo_t5gemma2.safetensors from your models/text_encoders/ folder and wraps it as a standard ComfyUI CLIP object. The node bundles its own tokenizer and config values, so one file is the whole encoder - there's nothing else to hunt down. The config lives inside the pack: 2560-wide hidden states, a 262,144-token vocabulary, and a Gemma tokenizer. It reads the safetensors with safetensors.torch.load_file, then hands it to comfy.sd.CLIP, and what comes out the other end is a plain CLIP that's compatible with any text-encode node, including the pack's own MotifTextEncode or the core CLIPTextEncode if you prefer that.

The inputs that matter

Only two required fields, and you'll touch one of them:

• clip_name - dropdown of files in models/text_encoders/. Pick motifvideo_t5gemma2.safetensors.
• dtype - bfloat16, float16, or float32. bfloat16 is the sane default and matches how the transformer wants to run; float16 saves a bit of VRAM, float32 is for when you're chasing exactness and have the card for it.
• device (optional, hidden under "advanced") - default or cpu. Load it onto the CPU if you're tight on VRAM; the encoder gets dispatched when needed instead of sitting in VRAM.

Output is one CLIP, and it wires straight into MotifVideo Text Encode.

Installing it

Install the whole pack first - this node doesn't exist standalone:

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

ComfyUI Manager search won't find this pack yet - the registry PR is planned but not submitted, and Manager's "Install via Git URL" path is gated by its security_level setting. The clone route above is the reliable one. Then drop the encoder file in place:

huggingface-cli download Motif-Technologies/Motif-Video-2B \
  text_encoder/model.safetensors --local-dir /tmp/motif-video-2b
# rename to motifvideo_t5gemma2.safetensors and move to models/text_encoders/

If you load the pack's example workflows on a recent ComfyUI (v0.18.0+), it'll offer to download all three weights for you automatically - that works too.

Where people get burned

• Empty dropdown. If clip_name lists nothing, the file isn't in models/text_encoders/. ComfyUI only shows what's already in that folder.
• VRAM. The full model package is ~16GB and this encoder is a real chunk of it. On a smaller card, run bfloat16 and set device to cpu; don't try to run everything in VRAM at once.
• Prompting habits. Since the encoder is LLM-class, verbose natural language beats a comma-separated tag list. That's a feature, but it means old SDXL muscle memory will get you mediocre results until you stop using it.

The heavyweight reputation of this pack is deserved - the recommended setup is H200-class hardware with --highvram, and the README's measured speed is ~30s/step at 1280×736, 121 frames. This node is the cheap part. It's also the part that decides whether KSampler has any idea what you asked for.