Krea 2 CLIP (CRT AutoDL)

Krea 2 became the default open base model for stylized work and training in 2026, and running it locally means sourcing three files: the 12B transformer, its VAE, and a text encoder. This node is the third one, the part everyone forgets until they hit the error. It's the CLIP loader half of the CRT AutoDL family, and its whole job is to download the right encoder and hand you a working CLIP output so you never have to hunt a HuggingFace URL again.

The name is accurate but worth unpacking: "Krea 2 CLIP" here is not CLIP in the SD1.5 sense. Krea 2 uses Qwen3-VL as its language model - Krea deliberately picked a vision-language model so the same stack could later do editing - and CRT's node pulls the 4B qwen3vl_4b_fp8_scaled repack straight from Comfy-Org's Krea-2 repo. It loads it with ComfyUI's native Krea2 CLIP type, so the tensors come out shaped exactly like the sampler expects.

What you actually do with it

The node has zero inputs - there's nothing to configure. Wire the CLIP output into a CLIP Text Encode (Prompt) node (or whatever encoder your Krea workflow uses), then drop CRTAutoDLKrea2TurboModel (or Raw) and CRTAutoDLKrea2VAE next to it and you have a complete local Krea 2 stack with three nodes.

The first time you queue the graph, it downloads the file into ComfyUI/models/text_encoders/. Watch the console, not the canvas: CRT's loader prints a progress bar, and a several-gigabyte file can make it look like ComfyUI froze. Subsequent runs are instant because the file already exists and the loader skips straight to comfy.sd.load_clip.

Where people get burned

• Missing the CLIP entirely. People drop in the model and VAE, hit an expected Qwen3-VL style error, and assume the model is bad. It isn't - the encoder was never loaded.
• First-run silence. A 4GB+ download with no visible UI spinner reads as a hang. Check the terminal for CRT AutoDL progress before you kill it.
• Mixed-up encoders. This CLIP is only for Krea 2 Turbo and Raw; it won't work with a Flux 2 or ERNIE stack even though they look similar at the socket level.

Installing CRT-Nodes

Install the pack through ComfyUI Manager by searching for CRT-Nodes, or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/plugcrypt/CRT-Nodes.git
pip install -r requirements.txt

Then restart ComfyUI. The AutoDL nodes have no heavy optional dependencies - they use the built-in download utility and huggingface_hub - so this is one of the rare CRT families that installs clean on every machine. The only cost is disk space, and the Krea 2 encoder is the smallest file in the stack.