CLIP Loader

The node's own description is the best summary: load 1–4 external CLIP models. For baked CLIP from checkpoints, use Model Loader. That second sentence is the whole job description - this is the loader you reach for when your model file doesn't ship its own text encoder.

That happens more than you'd think. Flux, SD3, and AuraFlow are usually distributed as diffusion-only UNets, so the CLIP has to come from somewhere else. The modern models also split their text encoder into multiple files: Flux wants two (clip_l.safetensors plus the big t5xxl encoder), SD3 wants three (CLIP-L, CLIP-G, T5). Wiring three separate stock CLIP loaders into one graph is where the tangle starts; this node loads them all and merges them into a single clip output.

The inputs

• clip_count - 1 to 4. Set it to how many modules your architecture needs, and the matching number of clip_name1…clip_name4 slots light up. Only the ones you actually fill get loaded.
• clip_nameN - pick from your models/clip/ and models/text_encoders/ folders (it scans both, deduplicated). .safetensors and .gguf both work, though GGUF CLIP needs the ComfyUI-GGUF extension installed.
• clip_type - the one that decides how the loaded modules get wrapped and combined. The dropdown has 30+ architectures: flux, flux2, sd3, sdxl, wan, mochi, ltxv, hunyuan_video, cosmos, krea2, and plenty more. Get this wrong and your encoder won't match the model - shape errors, or conditioning that's simply nonsense.

Output is one clip socket feeding any CLIP Text Encode node (including the pack's own CLIP Text Encode [Eclipse]).

A typical Flux setup

Model Loader [Eclipse] (model_type: UNet)     CLIP Loader [Eclipse]
├─ model ──────────────────────────────────► KSampler
                                             clip_count: 2, clip_type: flux
                                             clip_name1: clip_l.safetensors
                                             clip_name2: t5xxl_fp16.safetensors
                                             └─ clip ──► CLIP Text Encode (pos/neg)

That's the whole reason this node exists - two text encoders, one socket, no noodle.

Installing

Pack install as usual:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_SmartModelLoader.git
cd ComfyUI_SmartModelLoader
python -m pip install -r requirements.txt

or ComfyUI Manager → ComfyUI Smart Model Loader, then restart.

Common issues

• "Checkpoint does not contain a valid clip" - you pointed a baked-CLIP expectation at an external file, or you used Model Loader with a UNet. Use this node for the CLIP and Model Loader for the diffusion model.
• Blank or gibberish prompts - almost always the wrong clip_type for your model, or a missing T5 module in a Flux ensemble. Flux without T5 is a parrot with no vocabulary.
• GGUF CLIP not showing up - you need ComfyUI-GGUF; the loader lists GGUF files only when that extension is present.