XB-BOX - 📦 模型加载大全V2 (GGUF)

The GGUF version of the pack's dual-model loader. XB_ModelLoaderV2_GGUF loads a high-noise and low-noise model from .gguf files - each with its own LoRA stack, SageAttention preset, and block-swap count - plus the shared CLIP and VAE, all from one node. If you're running the refiner/high-low pattern and you need quantization to fit, this is the node that does both in one box.

Everything from the safetensors V2 carries over: the model_type keyword filter, the four-LoRA-per-side layout with _on/_strength toggles, independent sage_high/sage_low and blockswap_high/blockswap_low, and the shared clip/clip_type/vae. The differences are the two things that make it a GGUF node: the model dropdowns list .gguf files, and there's no weight_dtype selector because the quantization lives in the file.

Outputs are Model High, Model Low, CLIP, VAE - same contract as V2, so it drops into an existing high/low workflow without rewiring the rest of the graph.

How it works

Same runtime borrow as the other GGUF variants: it needs the ComfyUI-GGUF node pack installed in custom_nodes, finds it by directory name, and calls its UnetLoaderGGUFAdvanced for both model_high and model_low. Missing that dependency is a hard error, not a warning - the node refuses to run until ComfyUI-GGUF is present.

The two models load independently, their LoRAs apply independently, and the Sage/block-swap hooks fire per side. CLIP loads through ComfyUI-GGUF's CLIPLoaderGGUF (quantized text encoders - the Q4/Q8 T5s and Qwen encoders are the usual suspects), and the VAE stays a plain safetensors VAELoader.

The inputs that matter

• model_high / model_low - two .gguf UNets. This is where the Q ladder does the heavy lifting: Q4_K_M high model + Q8 low model is a common pairing, letting the detail finish stay crisp while the structure pass stays small.
• lora_high_1 / lora_low_1 - both default to on, same as V2. Check the toggles before assuming a LoRA leak.
• blockswap_high / blockswap_low - per-model block offload. On a tight card, run the high side with blocks swapped and the low side at 0.
• clip_type - shared, must match both models' text encoder family.

Install

# ComfyUI-GGUF is the real dependency:
cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF.git

# then this pack:
cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git
# or: ComfyUI Manager → "XB_ToolBox"

Without step 1 you'll hit the "ComfyUI-GGUF 未安装" error as soon as the node executes.

Notes

Two things to weigh before reaching for this. First, dual GGUF models + LoRAs is the worst-case combo for the dequant tax - every LoRA on a quantized model forces dequantize-patch-requantize per layer, and you're doing it on two models. If your refiner pair is small enough to run at fp8, V2 might be faster than V2_GGUF. Second, loading two quantized models is still loading two models; the node saves VRAM per-file, not the second residency. If your goal is strictly "one model, fits on card," the V1_GGUF is the honest pick.