Nodes/ComfyUI_Anytext/AnyText Loader(Test)
ComfyUI Node

AnyText Loader(Test)

Load AnyText from your own MODEL/CLIP/VAE

By zmwv823·Created 2 years ago·Updated about a year ago· 103
AnyText Loader(Test)
  • c_model
  • clip
  • vae
  • model
  • vae
  • ckpt_name
ckpt_name
control_net_name
miaobi_clip
weight_dtypeauto

This is the experimental AnyText loader - the "Test" in the name is the author's own, not a value judgment, but do treat it as the less-settled option. What it adds over the regular loaders is the ability to hand AnyText a MODEL, CLIP and VAE you've already loaded with native ComfyUI nodes, instead of loading everything from a checkpoint inside the node.

That's genuinely useful if you want to reuse a checkpoint that's already resident, or drive AnyText off a model you've patched elsewhere in the graph. It's also where the LoRA behavior gets a caveat worth reading.

How it works

Same two-option core as the other loaders - full AnyText checkpoint via ckpt_name, or SD 1.5 base plus AnyText control model - but with three optional inputs bolted on: c_model (MODEL), clip (CLIP) and vae (VAE). The node's description states the rule plainly: "If not input c_model, then load ckpt_name model." So the optional inputs override the checkpoint path when present, and fall back to ckpt_name when they're not connected. It also flags a LoRA limitation: "Only LoRA without trigger word works" - so style LoRAs are in, trigger-word character LoRAs are out.

The inputs and outputs that matter

  • ckpt_name - the fallback checkpoint, used when c_model isn't connected. Full AnyText checkpoint (Option 1) or SD 1.5 base (Option 2).
  • control_net_name - None for a full checkpoint, the AnyText control model for the SD 1.5 path.
  • c_model / clip / vae (optional) - feed these from native loaders to bypass loading from ckpt_name. This is the whole reason to pick this node over the plain loaders.
  • miaobi_clip - optional CLIP for Chinese prompts without a translator (offers google-flan-t5-large here).
  • weight_dtype (default auto) - precision.

Outputs: model (AnyText_Model), vae (VAE), ckpt_name (STRING).

How to install it

ComfyUI Manager, search ComfyUI_Anytext, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zmwv823/ComfyUI_Anytext

then restart. Same model requirements as the rest of the AnyText family - a checkpoint (or SD 1.5 base + control model) and a base font in models/fonts.

Common issues & troubleshooting

Nothing happens with my connected model. Remember the fallback rule: if c_model is connected it's used; if not, the node loads ckpt_name. If you meant to drive it off your own model, make sure c_model is actually wired.

My character LoRA has no effect. Expected. The node explicitly only supports LoRAs without a trigger word. Trigger-word LoRAs don't fire through AnyText's pipeline here - use a style LoRA, or apply LoRAs to the model before you pass it in via c_model.

Should I use this or a regular loader? For a normal AnyText run, the plain UL_AnyTextLoader / UL_AnyText_Loader is the calmer choice. Reach for the Test loader specifically when you need to feed AnyText a model/clip/vae from elsewhere in your graph. If it acts up, fall back to the standard loader - that's what "Test" is telling you.

CategoryUL Group/Image Generation

Inputs (7)

NameTypeDefaultDescription
ckpt_nameCOMBO0 options:
control_net_nameCOMBO1 options: None
miaobi_clipCOMBO2 options: None, google-flan-t5-large/model.safetensors
weight_dtypeCOMBOautoOnly fp16 and fp32 works. 仅支持fp16和fp32。
c_modeloptMODEL
clipoptCLIP
vaeoptVAE

Outputs (3)

NameTypeDescription
modelAnyText_Model
vaeVAE
ckpt_nameSTRING