Control_网格加载

Don't let the name fool you: Control_网格加载 ("grid loading") isn't loading a ControlNet. It's loading the base diffusion model - the thing ControlNet steers. Open this node in ComfyUI and you're looking at a straight rename of core's Load Diffusion Model (the UNETLoader) with a Chinese label and a parking spot under the ♐ControlNet submenu.

So why would you reach for it? Two cases. First, you're working with a standalone diffusion-model file - a .safetensors sitting in models/unet that isn't part of a merged checkpoint. In the DiT era that's increasingly common: newer workflows load the base model, the text encoder, and the VAE as three separate files instead of one big checkpoint, and this node is the "base model" leg of that setup. Second, you've just installed this pack, want the whole graph in its Chinese labels, and this is the loader that belongs next to the pack's Control_控制器加载 (its ControlNet model loader). One loads the guidance, this one loads what gets guided.

How it works

The mechanism is boring, and that's a compliment. The node lists every file in ComfyUI/models/unet, you pick one, and it runs comfy.sd.load_unet() on it, returning a MODEL. It's the same code path core uses - the pack just re-exports it. The one input that matters:

• unet_name - a dropdown of whatever diffusion model files you've dropped into models/unet (subfolders work too).

Output is a single MODEL socket. Wire it into anything that takes a model - a KSampler, the pack's ControlNet apply nodes, an upscale-model-free sampler - and it behaves exactly like a checkpoint's model half.

The trap beginners hit

A bare MODEL is not a complete generation rig. It has no CLIP text encoder and no VAE. Wire this alone into a sampler and you'll get errors about a missing clip or missing VAE the moment you try to encode a prompt. You need the other two pieces - a text encoder like the pack's Nodepad_文本模型加载 (its CLIP loader) or the text-encoder half of a checkpoint, plus any VAE loader. If your model ships as a normal merged checkpoint, this node is pointless for you: the standard Load Checkpoint gives you model, CLIP, and VAE in one go, and that's the tool you want.

Also worth knowing: if your file doesn't show up in the dropdown, it isn't in models/unet. Drop it there (or a subfolder) and refresh the node list.

Installing the pack

The README is a Chinese-language intro with screenshots and no install section, so the standard path is:

• ComfyUI Manager: search "ComfyUI_AC_FUN" and install.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/A719689614/ComfyUI_AC_FUN-V1.0

Restart ComfyUI. The pack is by Cc啊程, a Bilibili creator who's an architect/landscape designer by day and a Grasshopper-style visual programmer for fun - and it shows in how much this pack is wrappers around core nodes with Chinese names. There are no model downloads and no heavy dependencies, so install is quick. One quirk to expect: every node label is Chinese, so on an English UI these all sit under 🌌AC_FUNV1.0 with names like "网格加载" until you get used to them.

Bottom line

It's a faithful, renamed copy of a core node. Use it when a workflow is built around standalone diffusion-model files or when you're committed to the pack's Chinese-labeled graph. Otherwise, Load Checkpoint does more with less confusion. Not a trap, just a rename - and knowing that saves you the five minutes I spent wondering what a "grid" had to do with ControlNet.