Load DualCLIP (Quantized)

Modern models don't use one text encoder. SDXL pairs CLIP-L with OpenCLIP-G, Flux pairs CLIP-L with T5-XXL, and the newer LLM-encoder models pair something big with something small (Qwen with Byt5, Gemma with a small CLIP). When both halves need to be quantized - and on a 12GB card they usually do - you don't want two loader nodes and two format pickers fighting over the same CLIP connection. This node loads both quantized encoders and merges them into a single CLIP object in one go.

It's the dual-encoder version of Load CLIP (Quantized): same quant formats, same kernel backend choice, same loading toggles, just with two file pickers on top. The node's own description even ships with a recipe list, which is a genuinely nice touch:

• sdxl - clip-l + clip-g
• sd3 - clip-l + clip-g, clip-l + t5, or clip-g + t5
• flux - clip-l + t5
• hidream - at least one of t5 or llama; the description recommends both
• hunyuan_image - qwen2.5vl 7b + byt5 small
• newbie - gemma-3-4b-it + jina clip v2

How it works

The node loads both state dicts, detects each file's per-layer quant format, and merges the two sets of quantization metadata before calling ComfyUI's load_text_encoder_state_dicts. One nice detail: either encoder can be quantized or not - the hybrid ops handle mixed pairs, so you can run a quantized T5 against a full-precision CLIP-L if that's what you have. Also note that text_encoder2 lists files from both text_encoders and checkpoints folders, and for the ltxv type it resolves from checkpoints specifically.

The inputs that matter

• text_encoder1 and text_encoder2 - the two files, in the order your type expects (clip-l then t5, etc.).
• type - twelve preset architectures: sdxl, sd3, flux, hunyuan_video, hidream, hunyuan_image, hunyuan_video_15, kandinsky5, kandinsky5_image, ltxv, newbie, ace. This drives which encoder combination ComfyUI builds.
• quant_format - auto unless you have a reason to force one.
• kernel_backend - pytorch or triton; triton is the fast int8 path but needs the setup below.
• disable_dynamic and low_memory - the pack's shared loading toggles.

Output is a single CLIP feeding a CLIP Text Encode node.

Installing it

ComfyUI Manager → search "ComfyUI-QuantOps", or:

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

Restart and let requirements.txt pull in unifiedefficientloader>=0.5.2. Encoder pairs live on silveroxides' HuggingFace.

The caveats

The fast int8 path isn't free: int8_tensorwise on an RTX 30-series card wants torch 2.10+cu130 or newer, the matching triton-windows wheel, silveroxides' pre-compiled comfy-kitchen int8 wheel, and the --enable-triton-backend launch flag. And the pack is deprecated - int8 ConvRot is native in ComfyUI (v0.27.0+) now, so the author stopped maintaining and says as much in the README. This node's real remaining job is loading the older quantized dual-encoder files already in the wild. If you're wiring up something brand new, check whether ComfyUI's own dual CLIP loader already covers the formats you have.