MoE-JEPA boosts V-JEPA with a mixture of experts for synthetic image detection

Researchers presented MoE-JEPA — an architecture for detecting generated and manipulated images based on V-JEPA 2 with the addition of Residual Mixture-of-Experts and a separate branch for noise analysis. The work was published on September 15, 2026 on arXiv. The authors claim that the understanding of the visual world built into JEPA serves as strong prior knowledge for a deepfake detector, while the Gated Attention Multiple Instance Learning module refines spatial understanding. The model was evaluated on the SID-Set benchmark of 300 thousand images.

What it means

This work is not directly related to the tools in the 3D and video generation section: MoE-JEPA solves the task of detecting fakes, not creating content. However, it touches on the broader context in which generative models exist. Among our tools, the ones closest in spirit to this topic are Hunyuan3D 2.1 and Stable Fast 3D — both take images as input and produce 3D models. If a synthetic image is fed into such a generator, the result may inherit its artifacts. Detectors on the level of MoE-JEPA are a tool for checking source images, but the fact sheets of our generators contain no such check: neither Hunyuan3D 2.1 nor Stable Fast 3D describes filtering of input images for synthetic origin. OpenLRM also works with images, but its fact sheet contains no information about verifying the authenticity of input data. For all three tools, the question of what happens when a generated image is fed in remains open. The authors of MoE-JEPA released the code as open source, but did not specify the license in the presentation. VRAM requirements and model parameters are not given in the description of the work.