Faceswap Training Best Practices: A Complete Guide to Better Deepfake Results

Training is the most intensive and often confusing stage of working with Faceswap. Because it currently relies more on experimentation than established science, a great deal of conflicting advice circulates online. This guide consolidates the community's best practices into a clear, ordered set of recommendations to help you achieve the strongest possible results from your Faceswap models.

The most critical rule is to train on multiple sources for both faces. Relying on a single video for the Original (A) and Swap (B) data does not give the AI enough information to perform a proper swap. Limited data leads to poor color matching, insufficient coverage of facial expressions, and a substantially lower-quality final output. Varied footage across different scenes, lighting conditions, and angles gives the model the robustness it needs.

The most reliable way to decide whether training is complete is to inspect the preview images. The third column of each preview shows the swapped face, which you can compare against the original. Ideally the swap matches the original in detail and quality, though some blur and loss of detail is unavoidable. If one face reconstructs noticeably worse than the other, this signals a data problem that requires higher-quality or more varied material for the lagging face. Loss values, by contrast, only measure reconstruction of A-to-A and B-to-B and cannot assess swap quality. Waiting for loss to level out can also lead to overtraining, which degrades results over time.

YouTube videos are typically heavily compressed and shot under poor lighting, making them a weak data source unless no alternative exists. Ripping Blu-ray discs, or carefully selected DVDs, yields far cleaner data. If you film your own footage, vary the lighting, expressions, and poses, minimize motion blur and video noise, use ample lighting, a fast shutter speed, and a tripod. Do not discard every slightly blurred or imperfect frame; the model must learn to handle such imperfections. Remove only images that are severely misaligned or show no usable face.

Choose faces with similar shape and proportions. Swapping a thin, angular face onto a wide, rounded one, or making extreme changes to face shape, will produce poor results that often require manual correction. Skin tone should also be kept similar. The AI can adjust minor tonal differences, but the further the skin tone is from the original, the worse the outcome. Varied lighting in the training data helps the model learn these adjustments, and Color Augmentation should remain enabled when tones differ. Finally, avoid beards extending beyond the face area: such hair is only swapped within the face region and will not align properly, so sources with full beards are best avoided. Mustaches and subtle goatees are acceptable.

Fit Training involves first training on data the model will not see in the final output, such as footage from other films, then running a short final training phase on the actual target video. This reduces the risk of overtraining and produces clearer results. It requires multiple source footage, as training from a single source is not feasible. The fit phase should be much shorter than the original training, typically under 10% of its length. Changing the Swap data for Fit Training is not required but is recommended to further mitigate overtraining. Fit Training also works for reused models across multiple scenes, and keeping backups before each session lets you retain the best-performing version.

Pretraining or reusing a model across different face pairs can speed up training, but it causes identity leakage, in which the output fails to match the target face and takes on traits of the earlier identity. Because the model is incentivized to add to its existing knowledge, retraining on different faces mixes identities and degrades quality. A safer method exists: start a new model matching the pretrained one and replace its encoder with the pretrained encoder. This is more involved and still slower than a fully pretrained model, but it avoids identity leakage. The Load Weights option now officially supports this approach.

HDR videos use a non-linear colorspace, making colors and lighting unreliable for the model to learn facial features. Most, but not all, 4K films are HDR, and there is no dependable way to convert HDR footage for Faceswap without manually regrading the entire source. Where possible, use a non-4K Blu-ray, which generally does not include HDR.

Successful Faceswap training depends on diverse, high-quality data, careful face selection, visual evaluation of previews rather than loss, and disciplined use of Fit Training. By avoiding pretraining shortcuts and HDR sources, and by following these practices in order of importance, you can substantially improve the quality and realism of your swaps.