WINT8 CLIP Loader

On the 2026 generation of models, the text encoder is a second VRAM budget you have to plan around. Flux 2 Dev's encoder alone is roughly 18GB, and the community consensus is blunt about the split: quantize the encoder hard, keep precision in the diffusion model. But here's the thing nobody warns you about - ComfyUI's stock CLIP loader doesn't understand INT8 encoders. In the author's own words, it "either ignore[s] the quantization or fail[s] silently and load[s] garbage." WINT8 CLIP Loader is the fix for exactly that: the little brother to the WINT8 Diffuser Loader, doing for the text encoder what that node does for the diffusion model.

How it works

Instead of handing a file path to ComfyUI and hoping, this node loads the state dict itself, sniffs the tensor dtypes for int8, and injects its INT8 ops as custom_operations on the text-encoder loader. If the INT8 ops aren't available it falls back to standard loading gracefully rather than erroring. It also supports dual CLIP - two encoders loaded and merged in one call, which is what flux (clip-l + T5), SD3, and SDXL need - with per-encoder load mode control.

The inputs that matter

• clip_name_1 - a dropdown of the text encoders it finds in your clip, text_encoders, and clip_vision folders (plus unet_gguf, so GGUF encoders show up if you run ComfyUI-GGUF).
• clip_type_1 - the encoder architecture, defaulting to flux2; the full list covers the usual suspects (sd3, ltxv, wan, hidream, chroma, flux, and more).
• load_mode_1 - the one you'll actually touch. The author's own description: auto detects INT8 from the file and uses INT8 ops if found, int8 always forces INT8 ops, and standard skips them entirely - use that for fp16/bf16 encoders.
• dual_clip - the toggle that reveals a second clip_name/clip_type/load_mode row when you need a merged pair.

The single output is clip, which wires into a CLIP Text Encode node - or through the WINT8 Power LoRA Loader, which passes CLIP through untouched.

Install

Same pack, same story as the WINT8 Diffuser Loader: ComfyUI Manager (search "Winnougan") or

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/WINT8-ComfyUI

then restart. No pip requirements to fight with. One genuine dependency to know about: if you want to load GGUF text encoders (the T5-XXL / Qwen3 quantized encoder route), you still need ComfyUI-GGUF installed - this node can see those files and route them, but GGUF's own loader does the actual format work.

Common gotchas

• If the file dropdown shows only "none", your encoder isn't in one of the folders the node scans. For the 2026 models that's usually the text_encoders folder, and these files are separate downloads you grab alongside the checkpoint.
• Leave load_mode on auto until you have a reason to change it. Forcing int8 on an fp16 encoder without on-the-fly logic behind it is asking for nonsense output; standard is the right call for plain fp16/bf16 files.
• The failure mode this node exists to prevent - an INT8 encoder loading "fine" but producing garbage text - is silent by nature. If your prompt conditioning looks subtly wrong and your encoder is quantized, check which loader you're actually using. That's the whole reason this node is worth reaching for.