Instant ID is a relatively new model for controlnet, primarily used to fix facial features of characters. It can flawlessly swap faces using just one image, potentially making it the strongest face transfer model currently available.
Instant ID technology has demonstrated its unique advantages in various fields, mainly in the following aspects:
1. As an innovative method for preserving identity, Instant ID effectively resolves the contradiction between training efficiency and identity fidelity.
2. It is plug-and-play compatible with current text-to-image foundational models, LoRAs, ControlNets, etc., within the community, allowing for zero-cost maintenance of character identity attributes during inference. Additionally, Instant ID maintains excellent text editing capabilities, enabling smooth embedding of identities into various styles. Whether adding decorations, changing hair colors, or swapping outfits, it can handle it with ease.
3. Experimental results show that Instant ID not only surpasses methods based on single-image features for embedding (such as IP-Adapter-FaceID) but also holds its own against methods like ROOP and LoRAs in specific scenarios, with lower costs.
For a more detailed explanation of the principles, you can read the Instant ID team's analysis on GitHub: https://github.com/InstantID/InstantID
Step 1: Pre-use Preparation
If your controlnet does not have Instant ID, you need to update the extension using the Akebi startup package.
Additionally, you need to download the controlnet model to your local controlnet model folder.
Model download links:
ipadapter model
ControlNet model
If the network is poor, you can scan the QR code on the right to join the group for access, or experience it on the esheep online SD website.
After downloading, rename them to ip-adapter_instant_id_sdxl and control_instant_id_sdxl, respectively, and then place them in the folder.
Step 2: Image-to-Image Setup
[Large Model] SDXL model must be used, as Instant ID currently only supports SDXL.
I am using DreamShaper XL V2.1 Turbo here.
This model can be downloaded from the C-Station; if you cannot access it, you can scan the QR code on the right to join the group for access.
Or experience it on the Esheep online SD website.
[Sampling Method] DreamShaper XL V2.1 Turbo model must use DPM++ SDE Karras.
[Sampling Steps] 8
[Size] 768*1024
[CFG Prompt Relevance] The large model author recommends setting CFG to 2 for the best results.
[Redraw Intensity] 0.6
Refer to the image below for specific parameters; local redrawing should cover the face.
Step 3: ControlNet Setup
Your first unit must be a combination of embedding and ip-adapter.
The first unit is the model for capturing facial features.
Your second unit must be a combination of keypoints and control-instant-id-sdxl.
This unit serves as the replacement template; in image-to-image, you can choose not to upload an image, and SD will automatically use the image you uploaded above as the template.
Step 4: Click to Generate
After setting the above parameters, click to generate. You can see that the facial features are restored very well, except for some flaws at the edges, which can be improved by adjusting the mask edge blur.
After adjusting the redraw area and mask edge blur:
Notes
Parameters are not fixed and need to be adjusted according to your large model prompts; for the DreamShaper XL V2.1 Turbo large model I used, the above parameters were set.
Choosing high-definition facial materials can achieve better face-swapping effects.
------------------------------------------------------------------------------------------
Zhanzhangsucai AI Tutorials is the AI drawing tutorial platform under Zhanzhangzhijia
A wealth of AI free tutorials, continuously updating dry content
To learn more in-depth AI drawing tutorials, please visit Zhanzhangsucai AI Tutorials:
