Anymatix Dual CLIP Loader

Most of the models worth running in 2026 don't trust one text encoder to understand your prompt. SDXL pairs CLIP-L with CLIP-G and averages them. Flux pairs CLIP-L with a T5-XXL. SD3 mixes and matches CLIP-L, CLIP-G and T5 depending on the variant, and Wan, HiDream and the Hunyuan family each want their own two-encoder dance. AnymatixDualCLIPLoader is the twin of ComfyUI's stock DualCLIPLoader that handles exactly that: you point it at two text encoder files, tell it the recipe, and it stitches them into a single CLIP object your model can actually use.

The Anymatix difference is the same as everywhere else in this pack - the two encoder inputs are strings (paths or filenames) rather than dropdowns, so a workflow can reference encoders by URL through AnymatixFetcher and still load cleanly on a machine that didn't have them when the graph was written. Both names get stripped to basenames and resolved against your model folders, so a bare filename also works if the files are already sitting where ComfyUI expects them.

The type dropdown does the pairing for you

This is the part that saves you from a textbook failure. A CLIP-L file is just a pile of tensors; it doesn't know whether it's an SDXL left half or a Flux left half. The pairing lives in the type:

• sdxl → clip-l + clip-g
• sd3 → clip-l + clip-g, or clip-l + t5, or clip-g + t5
• flux → clip-l + t5
• hidream → at least one of t5 or llama (t5 + llama recommended)
• hunyuan_image → qwen2.5-vl 7b + byt5 small
• plus ltxv, kandinsky5, ace, newbie (gemma-3-4b-it + jina clip v2) and the rest of the dozen or so options

Get the pairing wrong and you'll see silent garbage embeddings or a runtime shape error rather than a friendly "you picked the wrong combination" - so treat the dropdown as the spec, not a suggestion. There's also an optional device input (default or cpu) for pushing the encode off the GPU when VRAM is tight.

Inputs and outputs

• clip_name1, clip_name2 (STRING) - the two text encoder files.
• type (enum) - the architecture pairing above.
• device (optional) - default or cpu.
• Output: CLIP → straight into CLIPTextEncode. Note this node does not need a separate AnymatixCLIPLoader for the second encoder - that's what dual means; a common beginner mistake is stacking two single loaders and wondering why the graph looks wrong.

Install

Same pack, same routine:

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

then restart, or use ComfyUI Manager (search "anymatix-comfy-nodes").

Issues to expect

Encoders are big - Flux's t5-xxl alone is ~9-10 GB in fp16 - so the "why is my run OOMing at the very start" answer is often this node. Force device: cpu if the encode is the memory bottleneck. And remember: this node loads files; the files have to exist. Pair it with an AnymatixFetcher if you want the download handled for you, and don't hand it a URL expecting it to fetch.