InstantID can generate customized avatar photos with just one reference image, without the need to train any models.
Instant ID controls facial features during the diffusion process by combining ControlNet and IP-Adapter. A unique design of Instant ID is that it passes the facial embedding from IP-Adapter as cross-attention input to ControlNet's UNet.
Online experience: https://huggingface.co/spaces/InstantX/InstantID
I tried the online demo and the model has 8 preset styles.
Test 1:
Upload a still of Baobao (Hu Ge) from "Blooming Flowers" and enter the prompt below, choose the Jungle style.
Test 2:
Keep the facial image unchanged, upload the reference pose (shown below), enter the prompt, and choose the watercolor style.
Common deployments
Next, we will introduce several commonly used deployment methods:
WebUI
Tutorial: https://github.com/Mikubill/sd-webui-controlnet/discussions/2589
Notes:
InstantID uses 2 models on WebUI. Always set the IP-Adapter model as the first model because the ControlNet model retrieves output from the IP-Adapter model.
The models should be placed in the specified directory: {A1111_root}/models/ControlNet
Rename the models separately to ensure extension functionality recognition: ip-adapter_instant_id_sdxl and control_instant_id_sdxl.
ComfyUI
Tutorial: https://github.com/ZHO-ZHO-ZHO/ComfyUI-InstantID
Build with Gradio yourself
For those who are proficient in coding, you can also implement it using Python scripts without using WebUI or ComfyUI. The official repository provides examples:
https://github.com/InstantID/InstantID?tab=readme-ov-file#download