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

Minkyu Kim; Baekseung Kim; Junhoo Lee; Jangho Kim; Nojun Kwak

cs.CVOct 15, 2026

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

Authors: Minkyu Kim, Baekseung Kim, Junhoo Lee, Jangho Kim, Nojun Kwak

Organizations: Seoul National University · Kookmin University

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.

Figures & tables

Figure 1: Illustration of the ManiFusion framework.

K\times B

inputs is fused into

B

representations, jointly processed by a shared denoising backbone, then unfused via the unfusion module into

K

representations per group, and finally passed through a projection layer to produce

K

outputs.

Figure 2: Comparison of throughput vs. quality and diversity on CIFAR-10 (left) and CelebA (right). The table (top) reports FID, sFID, IS, Precision, and Recall across throughput levels for Vanilla and Ours). The plots (bottom) visualize FID (right) and sFID (left) for DDIM, highlighting how Ours (red) compares against Vanilla DDIM (blue).

Figure 3: Analysis of the representation geometry induced by ManiFusion. (a) Geometric diagnostics measure prediction separation using response-distance ratios, and vector-field consistency and dynamics alignment using cosine similarity. (b) Training dynamics of Fusion Weight Orthogonality for MF-AE with $K=4$ , showing the development of slot-specific separation during training.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Clean-capacity and representation-alignment diagnostics under multi-sample fusion. Left. Swiss-roll samples as ambient dimension $D$ and fusion factor $K$ vary. Right. An affine transformation aligns the visualized fused representations with the single-image reference in this toy setting. The RMSEs are $4.83\times 10^{-4}$ for $(D{=}16,K{=}2)$ , $7.03\times 10^{-4}$ for $(D{=}16,K{=}4)$ , and $7.8\times 10^{-5}$ for $(D{=}32,K{=}8)$ .

Figure 5: Empirical local response ratios under multi-sample fusion. We measure local Lipschitz ratios of the denoising dynamics under both near-neighbor and fusion-induced perturbations across fusion factors. The estimated ratios for

K{=}2

and

K{=}4

remain clearly positive while staying well controlled, showing nonzero output responses in the sampled directions. For

K{=}8

, the ratio remains positive but becomes noticeably smaller, indicating attenuated responses under more aggressive fusion in this diagnostic.

Figure 6: Interpolation behavior under ManiFusion on CelebA. The top panel shows semantic interpolation between two identities by interpolating noise and running the reverse process. The bottom panel shows geometric interpolation between pose/appearance factors. Compared with the baseline (K = 1), ManiFusion (K = 2) yields smoother transitions and fewer abrupt mode switches. Red boxes highlight representative transition regions.

Figure 7: t-SNE visualization of feature distributions on CIFAR-10. The plots visualize the feature embeddings of real data (orange) and generated samples (blue) across different fusion factors. The panels show standard diffusion ( $K=1$ ) and ManiFusion with $K=2$ and $K=4$ from left to right.

Kernel	FID $\downarrow$	sFID $\downarrow$	IS $\uparrow$	Prec $\uparrow$	Rec $\uparrow$
3	5.93	4.58	8.67	0.63	0.58
5	6.94	4.74	8.58	0.61	0.59
7	7.47	4.74	8.33	0.62	0.58

Appendix

Table 10: Ablation on fusion kernel size. Larger kernels increase spatial entanglement, degrading sample quality. Smaller kernels (e.g., size 3) yield the best performance across FID, sFID, and IS.

Panel of Table 11 — Table 11: Qualitative comparison of ManiFusion ( $K=4$ ) versus a naive $2\times 2$ patch-based baseline on CIFAR-10 and CelebA. ManiFusion maintains sharper details and avoids mosaic artifacts.

Panel of Table 15 — Table 15: Randomly generated samples from DDPM, DDIM (T=50), and Flow Matching on CIFAR-10 for $K=1,2,4$ and TFS. ManiFusion maintains coherence and diversity across different samplers.

Panel of Table 16 — Table 16: Randomly generated samples from DDPM, DDIM (T=50), and Flow Matching on CelebA for $K=1,2,4$ and TFS. Results confirm that ManiFusion preserves sample fidelity and diversity across datasets.

Panel of Table 17 — Table 17: Qualitative comparison of vanilla LDM, MF-AE, and MF-UNet in latent diffusion models. Each fusion design is shown at $K=2$ and $K=4$ .

Panel of Table 18 — Table 18: Qualitative scaling with throughput on CIFAR-10 , showing random samples from ManiFusion (K=2) and TFS using DDIM with timesteps 10, 20, 50, and 100. Throughput ranges from 66.80–565.03 img/s.

Panel of Table 19 — Table 19: Qualitative scaling with throughput on CelebA , showing random samples from ManiFusion (K=2) and TFS using DDIM with timesteps 10, 20, 50, and 100. Throughput ranges from 26.83–228.77 img/s.

Diffusion Model Inference Acceleration

Diffusion model inference acceleration reduces the time and compute needed to generate or reconstruct outputs while preserving quality.

See all 506 papers

DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease
Mar 26, 2026Runsheng Bai, Chengyu Zhang, Yangdong DengDiffusion Model Sampling Diffusion Model Inference Acceleration
CSAIL, MIT · IST, Nanjing University · School of Software, Tsinghua University; FuturististAI Lab, Shanghai Tsinghua International Innovation Center
Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network
May 18, 2026Grigory Bartosh, David Ruhe, Emiel Hoogeboom +3Diffusion Sampling Diffusion-Based Image Generation
*Work done while interning at Google DeepMind Amsterdam, 1University of Amsterdam, 2Google DeepMind Amsterdam · 2Google DeepMind Amsterdam
Diffusion Image Generation with Explicit Modeling of Data Manifold Geometry
May 25, 2026Duoduo Xue, Zhiyu Zhu, Junhui HouDiffusion Transformer Diffusion Model Sampling
Department of Computer Science, City University of Hong Kong, China · City University of Hong Kong (Dongguan), China

Mar 26, 2026J/K move · Enter open · S save

DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease

Runsheng Bai, Chengyu Zhang, Yangdong Deng

CSAIL, MIT · IST, Nanjing University · School of Software, Tsinghua University; FuturististAI Lab, Shanghai Tsinghua International Innovation Center

Diffusion Model Sampling Diffusion Model Inference Acceleration Diffusion Sampling Diffusion-Based Image Generation Diffusion-Based Denoising

Open paper AskPDF arXiv

Diffusion models have achieved remarkable success in generating high-fidelity content but suffer from slow, iterative sampling, resulting in high latency that limits their use in interactive applications. We introduce DRiffusion, a parallel sampling framework that parallelizes diffusion inference through a draft-and-refine process. DRiffusion employs skip transitions to generate multiple draft states for future timesteps and computes their corresponding noises in parallel, which are then used in the standard denoising process to produce refined results. Theoretically, our method achieves an acceleration rate of

\tfrac{1}{n}

\tfrac{2}{n+1}

, depending on whether the conservative or aggressive mode is used, where

n

denotes the number of devices. Empirically, DRiffusion attains 1.4

\times

-3.7

\times

speedup across multiple diffusion models while incur minimal degradation in generation quality: on MS-COCO dataset, both FID and CLIP remain largely on par with those of the original model, while PickScore and HPSv2.1 show only minor average drops of 0.17 and 0.43, respectively. These results verify that DRiffusion delivers substantial acceleration and preserves perceptual quality.

Abstract

Diffusion Model Inference Acceleration

Explore similar work

DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease

Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network

Diffusion Image Generation with Explicit Modeling of Data Manifold Geometry