Anymatix CLIP Loader

Every modern model family has its own text encoder, and ComfyUI's stock CLIPLoader exists precisely because "which text encoder" stopped being a question with one answer years ago. SD 1.5 wants clip-l. SDXL wants clip-l and clip-g. Flux wants clip-l plus a T5. LTX wants its own T5 variant, Wan wants umt5, and the newer video and image models each drag their own favorite along. AnymatixCLIPLoader is a path-taking twin of that stock node: same type dropdown that encodes all those recipes, same CLIP output - but instead of a dropdown of files in models/clip, the clip_name input is a plain string path.

That makes it the natural partner for AnymatixFetcher. A workflow that ships a text encoder by URL can download it on first run and feed the path straight in. Hand-typing a filename works too, because unlike the checkpoint twin in this pack, this loader strips the path down to the basename and lets ComfyUI resolve it against your model folders - so a file already sitting in models/clip or models/text_encoders loads whether you paste the full path or just the name.

The type input is the whole game

This is where beginners get lost, because you don't just pick a file - you tell the node what kind of clip architecture it is, and the dropdown lists the recipes right in the node description:

• stable_diffusion → clip-l
• stable_cascade → clip-g
• sd3 → t5-xxl / clip-g / clip-l
• flux, mochi, cogvideox, cosmos → the t5-xxl family (cogvideox wants the 226-token padding variant)
• wan → umt5-xxl, lumina2 → gemma-2-2b, hidream → llama-3.1 or t5, and so on through two dozen entries including newer Qwen-based encoders for omnigen2, joyimage and minimax.

Pick the wrong type and the loader will happily try - and you'll get a shape mismatch or a garbage embed rather than an error that tells you what you did wrong. The node description is the cheat sheet; use it.

There's also an optional device input (default, or cpu) if you want to force the encoder off the GPU, which can matter when the text encoder is huge relative to your VRAM budget.

Inputs and outputs

• clip_name (STRING) - path or filename of the text encoder weights.
• type (enum) - the clip architecture from the recipes above.
• device (optional) - default or cpu.
• Output: CLIP, which wires into whatever your model needs - a CLIPTextEncode for prompting, or for dual-encoder setups you chain this with its sibling loaders.

Install

Same pack, same recipe as the other loaders here:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

…then restart, or install via ComfyUI Manager by searching "anymatix-comfy-nodes". No extra dependencies beyond what the pack pulls in.

Issues to expect

The realistic failure modes are (1) feeding it a path that doesn't exist yet because you skipped the fetcher, and (2) choosing a type that doesn't match the weights you loaded. Both give confusing downstream errors. And one honest note: if you're not using URL-addressed workflows, the stock CLIPLoader gives you the same thing with a friendlier file picker - reach for this twin when you need the path-based behavior, not out of habit.