Flux2Klein CLIP (CRT AutoDL)

No widgets, no file picker - run this node and it fetches the text encoder Flux 2 Klein needs and hands you a CLIP output. It's part of CRT-Nodes' auto-download family: instead of hunting down the right encoder file on HuggingFace and dropping it in the correct models/text_encoders (or clip/) folder yourself, this node does the fetch-and-place for you the first time you run it.

Why Klein's "CLIP" isn't really CLIP

Worth knowing before you wire this up: Flux 2 Klein doesn't use a CLIP model in the classic sense. Its text conditioning runs through a Qwen3 LLM (Qwen3-4B for the 4B checkpoints, Qwen3-8B for the 9B ones) - the prompt gets wrapped in a chat template and hidden states from specific layers get stacked into the conditioning ComfyUI passes to the sampler. ComfyUI's node graph still calls this a CLIP type because that's the socket family text encoders plug into, but functionally you're loading an LLM, not a CLIP-style contrastive encoder. That has a real practical consequence worth knowing up front: Comfy's prompt-weighting syntax ((word:1.4)) does nothing here, because the encoder isn't reading tags - it's reading your prompt as an instruction. Write plain, specific sentences instead of a tag-soup prompt; Klein's own documentation and community testing both land on that same conclusion.

The inputs and outputs that matter

• No required or optional inputs - this loader has nothing to configure.
• Output: a single CLIP.

Since there's no widget for it, the node's exact size/precision choice (which Qwen3 variant, what quantization) isn't exposed in the schema - you get whatever CRYPT_EXE wired the node to fetch. If you need a specific size or a quantized build (fp8, Q-something) to fit your VRAM, this auto-download loader may not be the flexible option; a manual CLIPLoader pointed at your own downloaded file gives you that control.

Installing it

Ships with the rest of CRT-Nodes; no separate install for this node:

ComfyUI Manager - search CRT-Nodes, install, restart.

Manual:

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

Restart ComfyUI. The encoder itself downloads on first use of the node, not at install time - expect a real pause and a real download (LLM-based text encoders run several gigabytes) the first time it runs.

One thing worth flagging plainly: Flux 2 Klein's weights sit under Black Forest Labs' licensing (Apache 2.0 for the 4B checkpoints, a non-commercial license for the 9B), and HuggingFace repos under a non-default license are sometimes gated behind accepting terms on the HuggingFace side. This node's own README entry doesn't document how - or whether - it handles that; if the download stalls or fails outright on a fresh install, a gated-repo/authentication issue is a reasonable first thing to check before assuming the node is broken.

Common issues

• Download fails or hangs on first run - check whether the target HuggingFace repo requires accepting a license first; a bare HTTP fetch can trip on that even when the repo itself is reachable.
• Prompt weighting syntax has no effect - not a bug. Klein's text encoder is an LLM-based one and doesn't support Comfy's (word:weight) emphasis syntax at all; write direct, descriptive sentences instead.
• Wrong CLIP for a mixed workflow - if you're combining Klein with other Flux 2 variants or another model family in the same graph, make sure this output is only feeding the Klein-specific sampler path; text encoders are not interchangeable across model families even within the same broader Flux 2 line.
• Need a smaller/quantized encoder for VRAM reasons - this auto-loader doesn't expose that choice; switch to a manual CLIPLoader node pointed at a specific quantized file you've downloaded yourself if you need that control.