Nodes/FunUI/FunUI CLIP Loader
ComfyUI Node

FunUI CLIP Loader

Four encoder slots and a family dropdown core won't give you

By digital-garbage·Created about a year ago·Updated about 14 hours ago· 21
FunUI CLIP Loader
    • clip
    • status
    ◄clip_name1▾►
    ◄typeltxv►
    ◄clip_name2►
    ◄clip_name3►
    ◄clip_name4►
    ◄devicedefault►

    ComfyUI core gives you CLIPLoader, DualCLIPLoader and TripleCLIPLoader, and that's usually fine - except the modern video families don't fit the naming: MiniMax H3 wants one encoder of a specific family, LTX-2.3 wants two that are assembled into one, and the family has to be named correctly or the load fails in ways that don't look like a loading problem. FunUI's CLIP Loader is the version that spells the family out and gives you four slots.

    What it is

    One node, up to four text-encoder files, one CLIP out. Unlike a LoRA loader this can't be a chain, because the files are combined into a single encoder rather than stacked on a model - they have to arrive together, in the right order.

    The four slots are numbered combos, clip_name1 through clip_name4, which is the same shape core's dual and triple loaders use. Slots two through four are optional and default to an empty string, meaning "no file here" - an empty slot isn't a validation error, because most families only need one. Four is a ceiling above anything currently shipping; LTX-2.3 needs two.

    The inputs you actually set

    • clip_name1 - the encoder itself. Required, and order matters.
    • type - the family. ltxv for LTX-2, minimax for H3. The list is built from whatever your ComfyUI install knows, so if a family is missing from the dropdown, your ComfyUI is older than the model you downloaded.
    • clip_name2…4 - fill in only what the family needs.
    • device - default or cpu. On cpu the encoder lives off the GPU: prompts get slower to encode, and the diffusion model gets the VRAM.

    Out comes clip (wire it into your text encoding and into FunUI Apply Prompt Markup, which needs the encoder to do its blending) and a status string.

    Why it refuses instead of guessing

    Here's the design choice worth knowing about. If the family you asked for doesn't exist in this ComfyUI, the node throws - "this ComfyUI has no encoder family 'x'. Known families: …" - instead of falling back to the default family. Falling back would load the encoder against the wrong architecture and report success, and the resulting failure shows up much later as something that looks unrelated: garbage output, a crash in the sampler, a phantom fault. Refusing and printing the list of known families is the right trade, even if it's annoying the one time you hit it.

    You'll hit it when you open a workflow someone saved on a newer ComfyUI, or after a CLIPType member gets renamed and your saved graph still names the old one.

    GGUF, and what's in requirements.txt

    The pack's requirements.txt pulls in gguf, a pure-Python + numpy reader for .gguf containers. That doesn't mean text encoders in GGUF just work: this loader only lists .gguf encoders when the ComfyUI-GGUF node pack is installed, and FunUI deliberately doesn't install it for you. The author's own note in the setup guide is blunt - Gemma 4 as .gguf is refused, use the .safetensors of the same model. A .gguf diffusion model is a different story: that loads without the extra pack, but gets expanded to full size in memory, which defeats the point.

    Installing it

    cd ComfyUI/custom_nodes
    git clone https://github.com/digital-garbage/FunUI
    pip install -r requirements.txt
    

    Or ComfyUI-Manager: search the pack title, FunUI. comfy node install FunUI also works. Restart afterwards and check Settings ▸ About FunUI for the version you actually got - the registry copy and the git copy drift.

    One honest caveat: Manager installs a node pack's dependencies but can't reconcile them for you. FunUI's requirements.txt is written defensively for exactly this reason - it pins comfy-kitchen>=0.2.32 because the sparse-attention kernel needs it, and it never forces a torch, numpy or transformers upgrade. If pip offers to move one of those, read the diff first.

    Where people get burned

    Wrong folder. Encoders go in ComfyUI/models/text_encoders. A file sitting in models/clip (the old location) shows up for some loaders and not others.

    Encoding on CPU "for VRAM" and then wondering why the UI hangs. It's not broken, it's just doing a prompt encode on the CPU. With one short prompt you won't notice; with a Composer story split into twenty scenes you will.

    Family mismatch. H3's encoder is not interchangeable with an LTX one. If you're pairing a clip file with a model it wasn't made for, nothing in this node will save you, and the failure will land somewhere else.

    CategoryFunUI/Loaders

    Inputs (6)

    NameTypeDefaultDescription
    clip_name1COMBOText encoder file. The first slot is the encoder itself.
    typeCOMBOltxvWhich family the encoder is for. LTX-2 is 'ltxv'; MiniMax H3 is 'minimax'.
    clip_name2optCOMBOAnother encoder file, loaded after the ones above. Leave empty if the family needs no more.
    clip_name3optCOMBOAnother encoder file, loaded after the ones above. Leave empty if the family needs no more.
    clip_name4optCOMBOAnother encoder file, loaded after the ones above. Leave empty if the family needs no more.
    deviceoptCOMBOdefault'cpu' keeps the encoder off the GPU: slower prompts, more VRAM for the diffusion model.

    Outputs (2)

    NameTypeDescription
    clipCLIP—
    statusSTRING—