ManiFusion: Unlocking the High-Throughput Generation via Superposition in Manifold Space

Key points

• Diffusion models achieve high-quality image generation but remain computationally expensive due to iterative denoising across many timesteps.1
• We instead explore an orthogonal throughput-oriented direction by processing multiple samples within a single denoising evaluation.1
• ManiFusion jointly processes multiple samples by fusing them before the denoising backbone and separating them afterward through unfusion.1
• To better align fusion allocation with denoising stages, ManiFusion uses Trajectory-Fusion Scheduling (TFS), which dynamically adjusts fusion factors throughout sampling.1

Sources (1)

• [1]ManiFusion: Unlocking the High-Throughput Generation via Superposition in Manifold Space
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 07:44 AM

  To this end, we propose ManiFusion, a backbone-agnostic framework for multi-sample generation through fused-space computation.

  Diffusion models achieve high-quality image generation but remain computationally expensive due to iterative denoising across many timesteps.

Extractive summary: sentences quoted from the sources.