Technology in the AI field moves very fast. The standalone “ReActor-UI” is now outdated. Today, I highly recommend using the ComfyUI-ReActor extension. It is more powerful, flexible, and integrated into the ComfyUI workflow.
Contents
Why ComfyUI-ReActor?
The biggest advantage of the ComfyUI version is its advanced control. It now includes features like:
- Detection Control: You can precisely control how the AI finds faces.
- Gender Filtering: You can set the tool to only swap male or female faces.
- Source Selection: It is easier to manage which source face goes to which target.
Improved Face Detection
There are more face detection models available now than before. These models help the AI find faces even in difficult angles or lighting.
The Best Model: InsightFace inswapper_128
When it comes to the swapping itself, there is an ongoing discussion about model size. Even though larger models are being developed, InsightFace’s inswapper_128 is still the best. While its 128×128 resolution might seem small, no other model has surpassed it in terms of accuracy and likeness. It remains the gold standard for quality results.
https://github.com/Gourieff/ComfyUI-ReActor
Technical Note: The Resolution and Quality
One common question is about the resolution of the inswapper_128 model. Since it only works at 128×128 pixels, you might think the result will look blurry on a high-definition (HD) screen.
Why 128px is still the best
Even though the resolution is low, this model is famous for its “similarity.” It captures the human face’s features more accurately than many newer, larger models. In face swapping, looking like the target person is the most important factor.
The Solution: Post-Processing with Face Restoration
To overcome the size limit, we use a two-step process. First, the AI swaps the face using inswapper_128. Second, we apply a Face Restoration model like GFPGAN or CodeFormer.
These restoration tools act like a “smart upscaler.” They take the 128px swapped face and rebuild the details—like eyes, skin texture, and hair—into a high-resolution image. By combining these two technologies, you get a result that is both highly accurate and crystal clear.
Current Verdict
Until a 256px or 512px model can match the accuracy of inswapper_128, this “Swap + Restore” workflow remains the professional choice for the best quality.
Troubleshooting: Library Conflicts and Re-installation
Sometimes, ReActor stops working after you install or update other custom nodes. This usually happens because of a Python library conflict.
The Cause: “Dependency Hell”
ReActor requires specific versions of Python libraries to run correctly. The most common culprit is NumPy.
The NumPy Problem: ReActor and its core components (like InsightFace) often need NumPy version 1.26.x. However, many new AI nodes now require NumPy 2.0 or higher. The Conflict: When you install a new node, ComfyUI might automatically upgrade NumPy to 2.0. This “breaks” ReActor because its internal code is not compatible with the new version. Other libraries like OpenCV or Protobuf can also cause similar issues.
The Easy Fix: Reinstall ReActor
If ReActor disappears or shows an “Import Error,” the best solution is to re-install it. You don’t need to delete the folder. Simply running the install script again will fix the environment.
Open your terminal/command prompt.
Go to the ComfyUI/custom_nodes/ComfyUI-ReActor folder.
Run install.bat (Windows) or install.py (Linux/Mac).
Why does this work?
The ReActor installation script is designed to “pin” (force) the library versions. When you run it, it detects the incompatible NumPy 2.0 and replaces it with the correct NumPy 1.26.4 (and other necessary versions) that ReActor needs. This restores the balance and makes the node work again.
