Research map / Image generation / Fast diffusion sampling

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

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ManiFusion, which fuses K samples into one representation before a diffusion backbone and unfuses afterward, with a Trajectory-Fusion Scheduling scheme that varies K across denoising steps, to raise generation throughput.

Problem

Diffusion sampling is expensive because it needs many sequential denoising evaluations; per-sample latency reduction (step reduction, caching) is saturating, so throughput (images per forward pass) needs a different axis.

Approach

Process multiple samples within one denoising evaluation by index-conditioned summation of their representations before a shared backbone and index-conditioned unfusion afterward, with fusion factor scheduled along the trajectory (larger K early, smaller K late), motivated by a manifold-capacity and fiber-ambiguity analysis.

Key results

• On CIFAR-10 DDIM (T=50), ManiFusion reaches 66.80 img/s vs 35.47 for vanilla (x1.88) with FID 10.35 vs 7.99 and MAC reduced to 50.4%.
• On ImageNet with DiT-S/2, ManiFusion achieves 5.98 img/s (x1.97) with FID 74.19 vs 71.85 for baseline at 3.04 img/s.
• Combining ManiFusion with DeepCache on CIFAR-10 DDIM gives 51.98 img/s (x2.94) with FID 9.42, versus 28.05 img/s (x1.59) for DeepCache alone at FID 6.90.

Surprising findings

• Naive fusion with large K (e.g., K=8 uniformly) degrades quality substantially, but a schedule that uses K=8 early can outperform schedules using only {4,2,1}.
• MF-AE shows smaller quality degradation than MF-UNet at the same fusion level, and mixing different class conditions within fused groups is more stable than grouping identical classes.

Limitations

• Requires full pretraining to integrate fused computation, so lightweight finetuning adaptation is not possible.
• Text-to-image foundation models are not evaluated; scalability is only discussed.
• Performance degrades under extreme fusion factors.

Experimental setup

Datasets: CIFAR-10 CelebA CelebA-HQ ImageNet

Benchmarks: CIFAR-10 CelebA CelebA-HQ ImageNet

Metrics: FID sFID IS precision recall throughput (images/s) MACs parameter count

Baselines: DDPM vanilla DDIM (T=50) vanilla Flow Matching vanilla DiT vanilla LDM baseline DeepCache step reduction

Builds on: DDPM DDIM Flow Matching DiT Latent Diffusion Models DeepCache

Closest papers by meaning

• AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.88)
• Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.88)
• Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.88)
• Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.87)
• Boosting Diffusion Models with Moving Average Sampling in Frequency Domain (CVPR 2024 · similarity 0.87)
• FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute (CVPR 2025 · similarity 0.87)