Diffusion models produce high-quality images but pay for it with costly iterative denoising. Most acceleration efforts have focused on making each individual sample faster, yet the abstract of this new preprint suggests that gains along that axis are running into limits.
The paper introduces ManiFusion, a method that instead targets throughput by generating multiple samples in superposition within manifold space. Rather than refining a single latent trajectory, the approach appears to batch or overlap samples in a shared manifold representation, potentially amortizing the denoising cost across many outputs.
The provided abstract is truncated, so full implementation details and experimental results are not yet available from this source. The core idea, however, is a clear shift in emphasis: from cutting per-sample latency to increasing the number of samples generated per unit of compute.