Emergent Mind


Numerous emerging deep-learning techniques have had a substantial impact on computer graphics. Among the most promising breakthroughs are the recent rise of Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS). NeRFs encode the object's shape and color in neural network weights using a handful of images with known camera positions to generate novel views. In contrast, GS provides accelerated training and inference without a decrease in rendering quality by encoding the object's characteristics in a collection of Gaussian distributions. These two techniques have found many use cases in spatial computing and other domains. On the other hand, the emergence of deepfake methods has sparked considerable controversy. Such techniques can have a form of artificial intelligence-generated videos that closely mimic authentic footage. Using generative models, they can modify facial features, enabling the creation of altered identities or facial expressions that exhibit a remarkably realistic appearance to a real person. Despite these controversies, deepfake can offer a next-generation solution for avatar creation and gaming when of desirable quality. To that end, we show how to combine all these emerging technologies to obtain a more plausible outcome. Our ImplicitDeepfake1 uses the classical deepfake algorithm to modify all training images separately and then train NeRF and GS on modified faces. Such relatively simple strategies can produce plausible 3D deepfake-based avatars.


  • ImplicitDeepfake introduces a novel approach combining Neural Radiance Fields (NeRF), Gaussian Splatting (GS), and classical deepfake algorithms to produce plausible 3D avatars with realistic facial modifications.

  • NeRF and GS are integrated to encode objects and enhance training and inference speeds through a representation based on Gaussian distributions, facilitating the creation of convincing 3D deepfake avatars.

  • The method demonstrates the capability for consistent face swapping and robust training datasets for 3D modeling, evaluated using metrics like PSNR, SSIM, and LPIPS.

  • ImplicitDeepfake advances deepfake technology by offering new levels for gaming, virtual reality, and avatar creation, while noting the need for ethical guidelines and detection methods.


The realm of image manipulation and video editing, particularly in the controversial milieu of deepfakes, draws significant attention both for its technological innovation and ethical implications. A recent contribution to this area is illustrated by the work of Stanishevskii et al., which introduces a novel approach termed ImplicitDeepfake. This method uniquely leverages the capabilities of Neural Radiance Fields (NeRF) and Gaussian Splatting (GS) alongside traditional deepfake algorithms, aiming to generate plausible 3D deepfake-based avatars through implicit modeling of facial modifications.

Technological Foundation

The foundational pillars of ImplicitDeepfake rest on the integration of NeRF, Gaussian Splatting, and classical deepfake algorithms. NeRFs have emerged as a powerful mechanism to encode objects into neural networks to generate novel views from a sparse set of images. On the other hand, Gaussian Splatting enhances this capability by offering accelerated training and inference times through a representation based on Gaussian distributions.

The classical deepfake algorithm operates by initially modifying training images through face swapping or altering facial expressions and then applying these modifications in training NeRF and GS. This generates a relatively straightforward pathway to produce 3D avatars that exhibit a high degree of realism and plausible facial alterations.

Contribution and Results

The primary contributions of the paper are threefold:

  • Introduction of a hybrid methodology that merges traditional deepfake algorithms with advanced neural rendering techniques (NeRF and GS), to facilitate the generation of convincing 3D deepfake avatars.

  • Demonstration of the ability of ImplicitDeepfake to effectuate consistent face swapping, enabling direct application of neural rendering on deepfake outputs.

  • Presentation of consistent image output which, when combined with NeRF and GS, creates a robust training dataset for 3D modeling.

In terms of technical achievement, ImplicitDeepfake was quantitatively evaluated using metrics like PSNR, SSIM, and LPIPS across different faces and scenarios. GS exhibited marginally superior results in generating sharper deepfakes compared to NeRF, which sometimes produced blurred outputs due to inconsistencies in the 2D deepfake generation.

Implications and Future Directions

ImplicitDeepfake not only advances the technical capabilities of deepfake generation but also poses significant implications both theoretically and practically. Theoretically, it expands upon the understanding of integrating classical AI methods with advanced neural rendering to enrich 3D modeling and avatar creation. Practically, the technology could revolutionize areas like gaming, virtual reality, and digital avatar creation, offering new levels of immersion and realism.

However, the ethical concerns surrounding deepfake technology remain paramount, emphasizing the need for further research into detection methods and ethical guidelines for usage. Future developments could also explore the application of ImplicitDeepfake in real-time systems, enhancing efficiency and reducing computational requirements, thus broadening its applicability and potential societal impact.


The paper presents a sophisticated advancement in the field of deepfake generation, proposing a novel approach that efficiently combines traditional and contemporary techniques. ImplicitDeepfake represents a significant step forward in creating highly realistic 3D avatars, paving the way for numerous practical applications while underscoring the critical need for ethical considerations and controls in the rapidly evolving domain of deepfake technology.

Get summaries of trending AI/ML papers delivered straight to your inbox

Unsubscribe anytime.