FuseBatch: Unlocking the Potential of Diffusion Models in Throughput Perspective

ICLR 2026 Conference Submission16722 Authors

Keywords: Throughput Optimization, Multi-Image Generation, Diffusion Models

TL;DR: FuseBatch boosts diffusion throughput by fusing multiple inputs in one forward pass to generate multiple images (instead of reducing steps). The approach is theoretically grounded and validated across DDPM, DDIM, flow matching, and latent diffusion.

Abstract: Diffusion models achieve state-of-the-art image quality but suffer from slow, iterative denoising. Existing acceleration methods focus on reducing the number of iterations, but these approaches are nearing practical limits. To address this, we take a different perspective by improving efficiency through generating multiple images within a single forward pass. We propose **FuseBatch** which fuses multiple inputs into a shared latent, applies the denoiser once, and unfuses the results to recover all outputs. To extend across domains, we introduce **FB-UNet** for pixel-space models and **FB-AE** for latent diffusion models.We further propose **Timestep-Fusion Scheduling (TFS)**, an inference-only strategy that balances throughput and quality, enabling FuseBatch to surpass the baseline at comparable throughput settings. Across DDPM and step-reduction methods (*e.g* DDIM, Flow Matching), we achieve near-multiplicative throughput gains with modest quality trade-offs, demonstrating its compatibility with existing acceleration techniques. Moreover, it scales effectively to high-resolution LDMs where larger fusion factors become attainable, providing a practical and orthogonal path to faster diffusion sampling.

Primary Area: generative models

Submission Number: 16722