XB-BOX - 📦 模型加载大全V3 (GGUF)

LTX 2.3 is 22B. On anything short of a 24GB card, running it at full fp16 is a prayer, and even fp8 is tight on a 16GB. XB_ModelLoaderV3_GGUF is the pack's answer: the same dual-CLIP, dual-VAE loader as V3, but the UNet and text encoders load from .gguf quantized files so the whole LTX 2.3 stack has a chance of fitting a consumer card.

Same contract as V3, same job: one node that loads model, two CLIPs (merged into one output), and two VAEs - the video VAE and the audio VAE that LTX 2.3 needs for its synchronized audio-video generation. If you're wiring the pack's XB_LTX23_InfiniteStreamer, this is the loader to pair with it when VRAM is the constraint.

How it works

The GGUF variant of V3, which means:

• model - .gguf UNet via ComfyUI-GGUF's UnetLoaderGGUFAdvanced.
• clip1 + clip2 - both .gguf, merged through ComfyUI-GGUF's DualCLIPLoaderGGUF into a single CLIP output, mirroring how the safetensors V3 uses the official DualCLIPLoader. Same contract, quantized encoders.
• vae1 + vae2 - plain safetensors VAE loaders. The video and audio VAEs stay full precision - this node quantizes the big DiT and the text encoders, not the VAEs.
• LoRA stack - eight slots (lora_1..8 with _on/_strength).
• sage_preset + blocks_to_swap - attention acceleration and block offload, same hooks as the safetensors V3.

Outputs: MODEL, CLIP, VAE1, VAE2. No model_weight_dtype dropdown - the quant is in the file.

The inputs that matter

• model - pick your Q. Q4_K_M is the "make it fit" pick for 12–16GB; Q8 if you have headroom and want the quality back.
• clip_type - default ltxv, and keep it there; the dual encoder pairing is LTX-specific.
• clip1 / clip2 - the quantized encoders. LTX 2.3's text encoder is big, so a Q4/Q5 T5-style encoder is usually the difference between fitting and not.
• vae1 / vae2 - video and audio VAE. Watch the order, same as V3: getting them swapped produces nonsense audio latents.

Install

# ComfyUI-GGUF first - this node errors without it:
cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF.git

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

The hard "ComfyUI-GGUF 未安装" error when it can't find the pack is the #1 support complaint, and it's the one the README's "no dependencies" line doesn't cover.

Notes

The quality math is friendly here: the VAEs stay full precision (the lossiest part of most pipelines), and LTX 2.3's DiT tolerates quantization about as well as any transformer. Q8 for the model is basically indistinguishable from fp8 at half the size; Q4_K_M is where the tradeoff starts showing. And the dequant tax on LoRAs applies to the GGUF UNet - if you're stacking several character LoRAs onto a Q4 model, expect a real speed hit and consider whether V3 at fp8 fits instead.