IP-Adapter for face locking is legit, way faster than training a LoRA.

Town Square / 小镇广场 General Chit-Chat / 日常闲聊
1

I used to always want consistent character faces, so back in the day I’d honestly train a LoRA—feed it dozens or even hundreds of images. Then I found out IP-Adapter is basically “single-image fine-tuning.” One image is enough, and the speed difference is night and day.

It was released by Tencent AI Lab in August 2023. In practice, you run it like a ControlNet. For face locking, grab the ip-adapter-plus-face_sd15 model and drop it into your ControlNet folder.

Usage is dead simple: write your normal prompt, drag your reference image into ControlNet’s image box, then pick the IP-Adapter preprocessor and the matching model. The key is the Starting Control Step—set it around 0.5. Lower value means the reference face is stronger but your model and prompt get weaker; higher value flips that. I usually fiddle between 0.4 and 0.6.

First run will download an encoder, so just wait a bit. Overall experience is fast—swap the reference image and boom, instant face change, no need to retrain.

2

“Single-image fine-tuning” — that’s a pretty spot-on way to put it.

3

That 0.5 step is honestly the secret sauce.

4

Is the era of training LoRA about to end?

5

The difference between plus-face and regular face is pretty huge tbh.