SplatFlux, described in a new arXiv preprint, is a feed-forward framework that converts video latents directly into 4D Gaussian splatting representations. The work targets two related tasks: reconstructing dynamic 3D scenes from observed videos and generating new 4D content from conditional inputs.
The authors' key observation is that observed videos and conditionally generated content can be handled in a unified way, though the abstract cuts off before fully elaborating on this point. By operating in latent space rather than pixel space, the framework aims to avoid per-scene optimization and enable faster, feed-forward processing.
Because the source is a single preprint abstract, details on architecture, training data, and quantitative results are not yet available. The full paper is posted on arXiv under reference 2610.28257.