Nunchaku Text Encoder Loader (Deprecated)

Quick heads-up before you wire this in: NunchakuTextEncoderLoader is deprecated. The current node is NunchakuTextEncoderLoaderV2, and new graphs should use that. This page is here because old workflows and tutorials still reference the original, and it helps to know what it did and what changed.

What it does is load Flux's text encoders - CLIP-L and T5-XXL, including Nunchaku's 4-bit T5 - and output a CLIP for your prompt-encode nodes. The encoder is a separate file you can quantize independently of the diffusion model, and on tight VRAM the encoder quant sometimes matters more than the model quant, which is the whole reason a dedicated 4-bit T5 loader exists. Nunchaku's upgraded 4-bit T5 was billed as matching fp8 T5 in quality.

How it works

Same core job as the V2: it takes two encoder files and packages them into the single CLIP object ComfyUI expects. The difference is that this original version exposed the 4-bit machinery as manual toggles, where the V2 hides that complexity and just does the right thing. If you're reading a guide that flips use_4bit_t5 on, this is the node that guide predates.

The inputs and outputs that matter

• model_type - fixed at flux here (the V2 lists it as flux.1).
• text_encoder1 / text_encoder2 - your CLIP-L and T5 files.
• t5_min_length (default 512, range 256–1024, step 128) - T5 padding length; leave at 512 for standard Flux.
• use_4bit_t5 (disable / enable) - the manual switch for loading Nunchaku's 4-bit T5. This is the toggle the V2 removed by handling it automatically.
• int4_model - selects the INT4 T5 build when use_4bit_t5 is on. Reads none until you have one available.

Output is a CLIP for your CLIP Text Encode nodes.

How to install it

It comes with the ComfyUI-nunchaku pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. You'll also want the backend wheel (via NunchakuWheelInstaller) for the 4-bit T5, plus the encoder files in your text_encoders folder.

Common issues & troubleshooting

Just use the V2. The single best move here is to swap this node for NunchakuTextEncoderLoaderV2, which is cleaner and actively maintained. There's no real reason to build a new graph on the deprecated one.

use_4bit_t5 on but nothing loads. You enabled 4-bit T5 without an INT4 T5 file selected in int4_model. Either download Nunchaku's 4-bit T5, or leave use_4bit_t5 off and load a normal fp8/fp16 T5.

Empty dropdowns. No encoder files present. Grab CLIP-L and a T5 into models/text_encoders.

Prompts feel off after a version update. An old node against a newer engine wheel can drift. That mismatch is exactly the kind of thing the V2 rewrite was meant to end - another nudge to migrate.