XB-BOX - 📦 模型加载大全V1

A normal ComfyUI loading setup is five nodes - UNET loader, CLIP loader, VAE loader, and a LoRA loader per LoRA. XB_ModelLoaderV1 is the pack's "model loading encyclopedia": one node that loads your UNet, CLIP, up to eight LoRAs, and VAE, then optionally applies SageAttention and block-swap VRAM tricks, and hands you all three (MODEL, CLIP, VAE) from a single spot.

The pitch is graph cleanup. If you're assembling a video model like Wan or LTX, the loader tangle eats a third of your canvas. This collapses it into one tall node and - the genuinely useful part - adds a keyword filter so you don't scroll through two hundred checkpoint names to find the one in your wan/ subfolder.

How it works

Underneath it's a thin shell over ComfyUI's official UNETLoader, CLIPLoader, VAELoader, and LoraLoader. The shell adds three things:

1. model_type keyword filtering. Type wan and every model/CLIP/VAE/LoRA dropdown only lists files inside a subfolder whose path contains wan. The dropdowns literally can't show anything until you type a keyword.
2. Eight LoRA slots, each with an _on boolean and a _strength float, applied in sequence to the loaded model and CLIP.
3. SageAttention + block swap. sage_preset patches attention with SageAttention (the KJNodes-derived accelerator - auto mode or the built-in M/N/GROUP/WAVE/WARP configs), and blocks_to_swap offloads the first N transformer blocks to system RAM to claw back VRAM.

The inputs that matter

• model_type - the keyword gate. Everything else is gated behind it, and that's the first thing that confuses people.
• model / clip / vae - dropdowns, each filtered by keyword.
• clip_type - 28 options from the official CLIPLoader (wan, ltxv, stable_diffusion, sd3…). This is the one to get right - a wrong clip_type is the classic "black output / garbage conditioning" failure.
• clip_device - default or cpu. Pushing CLIP to CPU frees VRAM at the cost of slower encoding; useful on cramped cards.
• model_weight_dtype - default or the fp8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2). fp8 is how 12B+ models fit.
• lora_1..8 with _on / _strength toggles.
• sage_preset - 关闭 (off) is the default and the right place to start.
• blocks_to_swap - 0 = off. Each block offloaded costs speed.

Outputs: MODEL, CLIP, VAE.

The folder gotcha

The keyword filter only lists models that live in a subfolder matching the keyword. A model sitting directly in models/diffusion_models/ - no subfolder - will never appear, no matter what you type. So the practical rule: keep your Wan models in models/diffusion_models/wan/, your LTX in ltx/, etc. If a dropdown says "无匹配模型" and you know the file exists, it's almost always a folder-structure problem, not a scan problem. Model files are picked from diffusion_models, CLIPs from text_encoders, and LoRAs from loras, all filtered the same way.

Install & troubleshooting

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

No pip deps for the loader itself. Two real-world warnings: SageAttention needs the sageattention package installed for 自动/built-in modes - if it's missing, the pack silently falls back to passing the model through, so you think you're getting a speedup you aren't. And block swap on a model whose blocks are already offloaded elsewhere (e.g. a GGUF loader with its own swap) can double-pay. Keep blocks_to_swap at 0 until you actually OOM, then raise it one block at a time - every block you swap is time you'll never get back.