ComfyUI Face Swap Guide: Fast AI Face Replacement with ReActor

AI face swapping is transforming digital image processing, enabling creators to transfer one person's facial features onto another image in minutes. Applications span film post-production, creative art, and entertainment. ComfyUI ReActor is a deep-learning face swap extension that brings this capability into a visual node-based workflow, making professional-grade replacement accessible to beginners and experienced creators alike. This guide walks through installation, core features, practical workflows, optimization, and responsible use.

Among AI face swap tools, ReActor stands out for speed, accuracy, and ease of use. Built on the InsightFace framework and integrating RetinaFace and YOLOv5 face detection, it reports face recognition accuracy above 95% and completes single-image swaps in about one second on capable hardware. Compared with traditional standalone tools, which often achieve 70–85% accuracy and take 5–10 seconds, ReActor works as a native ComfyUI node, supports multiple swap models such as Inswapper and ReSwapper, and integrates face restoration with GFPGAN and CodeFormer out of the box.

Before installing, confirm your system meets the requirements: Python 3.8 to 3.11 (3.10 recommended), an existing ComfyUI installation, an NVIDIA GPU with CUDA support (recommended) or another hardware accelerator, and at least 4 GB of available VRAM. Step 1: Clone the repository with git clone https://gitcode.com/gh_mirrors/co/comfyui-reactor-node and move into the project folder. Step 2: Run the automated installer with python install.py. The script installs the required Python dependencies, downloads the InsightFace face recognition model, configures the ONNX inference environment, and sets up the OpenCV image processing library. Step 3: Restart ComfyUI. Search for "ReActor" in the node search box to find the newly added face swap nodes.

ReActor provides several node families. The basic face swap nodes include ReActorFaceSwap, the core single-image replacement node; ReActorFaceSwapOpt, which exposes advanced options; and ReActorOptions, which supplies detailed configuration to the FaceSwapOpt node. Face model management nodes include ReActorSaveFaceModel for saving facial features as reusable model files, ReActorLoadFaceModel for loading them, ReActorBuildFaceModel for blending multiple facial features into one model, and ReActorMakeFaceModelBatch for creating models in bulk. Utility nodes include ReActorRestoreFace for improving output quality after swapping, ReActorImageDublicator for batch image duplication, and ImageRGBA2RGB for converting image formats to ensure compatibility. For a deeper understanding, the core logic lives in nodes.py (node implementations), reactor_utils.py (image processing helpers), install.py (installation automation), and scripts/reactor_faceswap.py (the core swap algorithm).

Start with a simple single-image workflow. First, load the source image containing the face you want to transfer and the target image receiving it, using ComfyUI's Load Image nodes. Next, add a ReActorFaceSwap node. Connect the source image to the source_image input and the target image to the target_image input. Set face_model to inswapper_128.onnx, and choose GFPGAN or CodeFormer as the face_restore_model. Set face_restore_visibility between 0.8 and 1.0. Finally, click Queue Prompt and review the result after a few seconds.

Detection settings strongly affect accuracy. det_size controls detection resolution: larger values are more precise but slower, with 640 recommended as a starting point. det_thresh sets how strict detection is, with 0.5 recommended. max_num limits the number of detected faces, which is useful in multi-person scenes. Restoration settings balance fidelity and enhancement. For high-quality source images, use face_restore_visibility of 1.0. To preserve more of the original target's characteristics, lower it to around 0.7. For CodeFormer, a codeformer_weight of 0.5 offers a balance between fidelity and restoration. For multi-person images, use source_faces_index and faces_index to select specific faces, and enable gender filtering to improve matching precision. Process crowded scenes one face at a time when necessary.

For batch jobs, configure ReActorImageDublicator with an appropriate batch size. Suggested settings are batch size 1 to 2 for GPUs under 8 GB, 4 to 8 for 8 to 12 GB, and 8 to 16 for GPUs above 12 GB. Use GFPGAN for high-quality restoration and det_size of 640, then save all outputs with Save Image. Video face swapping follows a pipeline: video file, frame extraction, face detection, ReActorFaceSwap, frame recombination, and output video. Use an appropriate interpolation setting to keep transitions between frames smooth, enable face_boost_enabled to improve facial detail, and apply ReActorMaskHelper to precisely control the replacement region.

Optimal configuration depends on hardware. NVIDIA GPUs benefit most from CUDA and TensorRT. AMD GPUs can use ROCm, Apple Silicon can use CoreML, and CPU-only systems offer broad compatibility at slower speeds. Memory management matters for longer sessions. The ReActorUnloadModels node releases models once processing finishes, and clearing the GPU cache helps prevent out-of-memory errors. In benchmarks, an RTX 4090 processes a single image in about 0.3 seconds and eight images in about 1.8 seconds, an RTX 3080 takes about 0.5 and 3.2 seconds, an RTX 3060 takes about 0.8 and 5.1 seconds, and a CPU (i9-13900K) takes about 4.2 and 28.5 seconds.

If model download fails, check your network connection or manually place the model files in ComfyUI/models/insightface, ensuring they are complete and uncorrupted. If you encounter CUDA out-of-memory errors, reduce the batch size, lower input resolution, use ReActorUnloadModels to free memory, and update your GPU driver. If face detection fails, adjust det_size (try 320, 640, or 1280), verify the image is sharp and the face is clearly visible, lower det_thresh, or improve lighting conditions and apply image enhancement.

In film and post-production, ReActor supports historical drama casting, stunt doubles whose faces are replaced with those of principal actors, and localized multi-language versions. Digital artists can create virtual idols by blending facial features, produce unique characters from mixed identities, or convert realistic faces into artistic styles. In education, creators can produce multi-language versions of lessons from a single instructor, build virtual teaching avatars, and develop personalized learning materials.

Responsible use is essential. Respect portrait rights and privacy, obtain necessary permissions, and clearly label AI-generated content. Do not use face swapping to deceive or mislead people, create fake news, or produce harmful material. Respect cultural differences and personal dignity. Note the model licensing limits: InsightFace pretrained models are restricted to non-commercial research use. Commercial projects should train their own models or obtain a commercial license, and all users should comply with open-source license requirements and applicable local laws and platform policies.

ComfyUI ReActor combines performance, usability, and rich functionality into a practical AI face swap toolkit for filmmakers, digital artists, and AI enthusiasts alike. A suggested learning path begins with single-image swaps to build familiarity, moves through batch and video processing and face model blending, then refines workflows for your hardware, and finally applies the techniques to creative projects. Technology is only a tool; creativity and responsibility determine its value. Start experimenting, and share your successful projects with the community.