(Down)Load CLIPModel

This little node exists to solve one specific annoyance: DynamiCrafter and ToonCrafter don't run on the same text encoder as your usual SDXL or Flux workflows, and if you feed them the wrong CLIP your prompts quietly do nothing useful. This node loads the Stable Diffusion 2.1 CLIP text encoder the models actually want, and - like the rest of Kijai's "(Down)Load" nodes - grabs the file for you the first time if you don't have it.

It's part of kijai's ComfyUI-DynamiCrafterWrapper. In the pack's "Update2" refactor, Kijai rebuilt the wrapper to integrate properly with ComfyUI's own conditioning system, and the note in the README is explicit: you now can and have to supply both a CLIP Vision model and a CLIP text model. This node is how you supply the text one without going to hunt for the right file.

How it works

DynamiCrafter's video-diffusion backbone is built on a Stable Diffusion 2.1-lineage model, so it speaks the SD 2.1 CLIP text embedding, not the SD 1.5 / SDXL / Flux ones. That's the whole reason this node ships instead of you just using a stock CLIP loader - it points you at the correct encoder and removes the "which CLIP?" guesswork. First run downloads it into your models folder; after that it just loads.

The inputs and outputs that matter

There's exactly one input:

• model - a two-item dropdown: stable-diffusion-2-1-clip-fp16.safetensors or the full-precision stable-diffusion-2-1-clip.safetensors. Take the fp16 one unless you have a concrete reason not to; it's half the size, loads faster, and the quality difference on a text encoder is nil in practice.

The output is clip (type CLIP), and it wires straight into a standard CLIP Text Encode node - one for your positive prompt, one for negative. Those two CONDITIONING outputs then feed the sampler (DynamiCrafterI2V, ToonCrafterInterpolation, DynamiCrafterBatchInterpolation). So the chain is: this node → CLIP Text Encode (×2) → the sampler's positive / negative inputs.

Installing it

It's bundled with the pack, so there's no separate install. Get the pack via ComfyUI Manager (search ComfyUI-DynamiCrafterWrapper) or cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-DynamiCrafterWrapper, run pip install -r requirements.txt, and restart. The CLIP file itself downloads on first use - nothing extra to place by hand.

Common issues

• "Why is my prompt being ignored?" Nine times out of ten it's because a DynamiCrafter/ToonCrafter graph got wired to the wrong text encoder - a leftover SDXL or Flux CLIP from another workflow. This node exists precisely to stop that; make sure your CLIP Text Encode nodes are fed from this output, not a stray one.
• First-run pause. It's downloading the encoder. Check the console; it's a one-time thing.
• Don't confuse it with CLIP Vision. The samplers take a separate clip_vision input (the image encoder that reads your input frame). That's a different model loaded through ComfyUI's normal CLIP Vision loader - this node is only the text side. Two encoders, two jobs.