This project contains information about the inswapper-512-live model and the application releases based on it. The first released application is a real-time face anonymization tool for macOS, which allows users to swap their face in the camera with a predefined template character in real time.
2025.06.21 : Release v0.1.2. No major updates, v0.1.1 may no longer be valid.
System requirements: Apple Silicon Macs(M series) and macOS >= 13.5.
- Download the latest version from the Release page of this project;
- Since the application has not yet passed Apple's notarization process (due to the new developer account and no previously released AppStore apps), users need to manually modify privacy settings to allow the application to execute correctly;
- Visit the Picsi.Ai profile page and log in with your Discord account to obtain a UToken, enter it into the application's login box, and it may take 2-3 seconds to complete the login. The login process requires an internet connection.
- The application is optimized for 30 FPS, but actual frame rates may vary based on camera hardware (some external cameras, such as Studio Display, may drop to 25 FPS). But using an external camera generally results in more stable model inference times;
- There is still significant room for improvement in pre-processing and post-processing steps, such as the real-time performance and stability of face tracking, as well as the strategy for pasting back the results;
- The output of the time cost statistics may be inaccurate.
- The inswapper-512-live model itself is not included in this release and is only available under strict licensing agreements;
- This application is not designed for real-time arbitrary identity modification but rather for face anonymization and controlled demonstrations;
- All source faces are pre-selected templates, ensuring no open-ended face-swapping functionality is available in this release.
inswapper-512-live is an ultra-lightweight face swapping model that provides native 512x512 resolution. Compared to inswapper_128, it requires less than 1/10 of the computational resources yet produces outputs with 16 times the pixel count(512x512 vs 128x128).
The numbers in the table may fluctuate as we continue to gather more diverse test data.
| Model | Output | Params | GFLOPs | Similarity | Realism | Attributes |
|---|---|---|---|---|---|---|
| INSwapper_128 | RGB 128x128 | 138.3M | 174.7G | 86.9 | 63.3 | 78.8 |
| INSwapper_Cyn | - | - | - | 92.8 | 75.9 | 80.6 |
| INSwapper_Dax | - | - | - | 92.5 | 86.0 | 81.5 |
| INSwapper_Dax(w Optimizer) | - | - | - | 93.4 | 90.2 | 87.7 |
| INSwapper_512_live_base | RGB 512x512 | 9.5M | 12.1G | 87.0 | 73.7 | 80.1 |
| INSwapper_512_live_mini | RGB 512x512 | 4.7M | 6.3G |
Params: Number of parameters in the network, with lower values indicating smaller model file size.
GFLOPs: Computational complexity measured in FP32(half if using FP16), with lower values indicating faster inference.
Evaluation Metric Explanation:
-
Similarity: The similarity between the generated face and the source face is evaluated using a third-party, high-precision face recognition model. The similarity score is calculated using the cosine similarity between the feature embeddings of the source and output images, scaled to a range of 0-100, where higher scores indicate greater similarity;
-
Realism: This metric evaluates the realism and clarity of the generated face. We employ a comprehensive evaluation framework that combines the FID score, custom-designed metrics for facial clarity, and human perceptual evaluation. The score ranges from 0 to 100, with higher scores indicating a more realistic and clear result. We used a test set consisting of hundreds of high-resolution facial images, and randomly resized between 160x160 and 640x640 pixels in testing time;
-
Attributes: This metric assesses the correlation between the attributes of the generated face and the target face, including factors such as pose, gaze direction, expression, etc. We employ third-party facial attribute models to extract feature representations of both the generated and target images. The similarities between these feature vectors are then computed to obtain the final metric score, which ranges from 0 to 100, with higher scores indicating better attribute alignment.
All resources in this project, including models, libraries and applications, are strictly for academic use or personal testing and cannot be used for any commercial purposes.
contact#picsi.ai or contact#insightface.ai