Load CLIP (Quantized, Simple)

Same job as Load CLIP (Quantized), minus two controls you'd only touch if you were debugging. The "Simple" variant drops the quant_format and kernel_backend dropdowns and just loads the file - it runs the full loader with format set to auto and the backend set to pytorch, which the source does by literally calling the non-Simple class. That's the whole trick: it's a wrapper, not a separate implementation.

You reach for this when you grabbed a quantized text encoder from silveroxides' HuggingFace and just want it to work. The full loader's format dropdown earns its keep when you're mixing formats or diagnosing a file; for the common case - one quantized encoder, no idea what layout it uses, don't care - this is the cleaner node. Text encoders are increasingly the VRAM bottleneck on 12GB cards (the KB makes the point with Flux 2 Dev, whose encoder is 24B of a 56B total), so "quantize the encoder, keep the model precise" is the move, and this node is the low-friction way to do it.

The inputs that matter

• clip_name - the quantized encoder file from text_encoders.
• type - the one thing Simple doesn't auto-detect. It's the same two-dozen-option architecture list (stable_diffusion, sd3, ltxv, cosmos, wan, hidream, chroma, and so on), and you still need to pick the family that matches your model, because that's what builds the correct CLIPType.
• disable_dynamic and low_memory - the shared loading-path toggles. low_memory's tooltip: fast low-impact loading on, or ComfyUI's default when off.

Output is a single CLIP, straight into a CLIP Text Encode node.

Installing it

Standard pack install: ComfyUI Manager → search "ComfyUI-QuantOps", or clone and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-QuantOps

It pulls in unifiedefficientloader>=0.5.2 on first launch.

The catch

"Auto" isn't magic. It detects recognized quant layouts - int8, tensorwise int8, the fp8 variants, mxfp8, nvfp4 - and if the file isn't one of those, it won't pretend otherwise. So this node is only "no-thinking" for files that came from the convert_to_quant toolchain or silveroxides' own repos. And be aware the whole pack is deprecated: int8 ConvRot is native in ComfyUI now, the author has stopped maintaining, and new work on a recent ComfyUI is better served by the built-in loaders. This node is the easy path for legacy quantized encoder files, not a thing to build a fresh stack on.