Manifusion speeds up image generation by processing multiple samples together
ManiFusion: Unlocking the High-Throughput Generation via Superposition in Manifold Space
Computer Vision and Pattern RecognitionMachine Learning
Summary
Generating high-quality images with diffusion models usually takes a lot of time because the process involves many slow steps. This work introduces ManiFusion, which speeds things up by combining multiple images into one representation during the processing, then separating them afterward. The authors explain how this combined representation still keeps the unique details for each image so they can be generated correctly. ManiFusion works with different types of diffusion models and can be used alongside other speed-up techniques, increasing the number of images produced without much loss in quality.
What this means in practice
- •For graphic software developers: Integrate ManiFusion to generate multiple images faster, improving performance in applications requiring batch image synthesis.
- •For cloud service providers: Run ManiFusion to deliver higher throughput in image generation APIs, allowing more users to receive results simultaneously.$Commercial implications: Enables cloud providers to offer faster image generation services at scale, improving product offerings for AI-based visual content creation.
Authors
Minkyu Kim, Baekseung Kim, Junhoo Lee, Jangho Kim, Nojun Kwak
Abstract
Diffusion models achieve high-quality image generation but remain computationally expensive due to iterative denoising across many timesteps. Existing acceleration methods mainly reduce per-sample latency, but gains along this axis are becoming saturated. We instead explore an orthogonal throughput-oriented direction by processing multiple samples within a single denoising evaluation. To this end, we propose ManiFusion, a backbone-agnostic framework for multi-sample generation through fused-space computation. ManiFusion jointly processes multiple samples by fusing them before the denoising backbone and separating them afterward through unfusion. We provide a geometric perspective on how multiple samples can share a fused representation while retaining the information needed to guide their individual generation trajectories. To better align fusion allocation with denoising stages, ManiFusion uses Trajectory-Fusion Scheduling (TFS), which dynamically adjusts fusion factors throughout sampling. ManiFusion generalizes across diffusion families, including both U-Net and transformer backbones in pixel and latent spaces. We further show that ManiFusion composes with existing acceleration methods and reaches throughput regimes unattainable by prior approaches with minimal quality degradation.