SubFlow: Sub-mode Conditioned Flow Matching for Diverse One-Step Generation
Authors: Yexiong Lin, Jia Shi, Shanshan Ye, Wanyu Wang, Yu Yao, Tongliang Liu
Organizations: Sydney AI Centre, The University of Sydney, Sydney, NSW, Australia · School of Artificial Intelligence, Xidian University, Xi’an, Shaanxi, China · Australian Artificial Intelligence Institute, University of Technology Sydney, Sydney, NSW, Australia · Department of Information Systems, City University of Hong Kong, Hong Kong, China
Flow matching has emerged as a powerful generative framework, with recent few-step methods achieving remarkable inference acceleration. However, we identify a critical yet overlooked limitation: these models suffer from severe diversity degradation, concentrating samples on dominant modes while neglecting rare but valid variations of the target distribution. We trace this degradation to averaging distortion: when trained with MSE objectives, class-conditional flows learn a frequency-weighted mean over intra-class sub-modes, causing the model to over-represent high-density modes while systematically neglecting low-density ones. To address this, we propose SubFlow, Sub-mode Conditioned Flow Matching, which eliminates averaging distortion by decomposing each class into fine-grained sub-modes via semantic clustering and conditioning the flow on sub-mode indices. Each conditioned sub-distribution is approximately unimodal, so the learned flow accurately targets individual modes with no averaging distortion, restoring full mode coverage in a single inference step. Crucially, SubFlow is entirely plug-and-play: it integrates seamlessly into existing one-step models such as MeanFlow and Shortcut Models without any architectural modifications. Extensive experiments on ImageNet-256 demonstrate that SubFlow yields substantial gains in generation diversity (Recall) while maintaining competitive image quality (FID), confirming its broad applicability across different one-step generation frameworks. Project page: https://yexionglin.github.io/subflow.
Figures & tables
Figure 1. Toy experiment on a 4-peak Gaussian mixture target with imbalanced subclusters (majority weight 0.35, minority weight 0.15 per class). (a) Traditional flow matching (100 inference steps) covers all modes. (b) MeanFlow suffers from severe mode collapse, concentrating only on dominant subclusters and losing diversity. (c) In contrast, our SubFlow successfully restores full mode coverage in a single step by decomposing each class into fine-grained sub-modes and conditioning the flow on sub-mode indices.
Figure 2. Overview of SubFlow. (a) Offline pre-processing : semantic features are extracted from the training images and clustered within each class, yielding sub-mode assignments {ki} and empirical priors p(k∣c) . (b) Training : the vector field vθ(xt,t,c,k) is optimized with the SubFlow objective. For classifier-free guidance, only the class label c is randomly dropped, while the sub-mode index k is always retained. (c) Inference : given a target class c , a sub-mode index k is sampled from p(k∣c) , and the conditioned vector field generates a sample associated with the selected sub-mode.
Method
Param
Epoch
NFE
FID ↓
Precision (%)
Recall (%)
GANs
BigGAN ( Brock et al., 2019 ) [ICLR’19]
112M
-
1
7.70
82.17
27.67
StyleGAN-XL ( Sauer et al., 2022 ) [SIGGRAPH’22]
166M
500
1
2.30
71.93
48.71
GigaGAN ( Kang et al., 2023 ) [CVPR’23]
569M
128
1
3.45
75.57
48.36
Multi-step Diffusion/Flows
ADM ( Dhariwal and Nichol, 2021 ) [NeurIPS’21]
554M
400
250
4.59
82
52
Table 1. Class-conditional ImageNet-256 generation results. NFE denotes the number of function evaluations (sampling steps). Results in the Multi-step Diffusion/Flows section are cited from ( Zheng et al., 2026b ) , and all other results are evaluated using officially released checkpoints. SubFlow consistently improves Recall (diversity) when integrated into different one-step baselines, while maintaining competitive FID and Precision.
Figure 3. Qualitative comparison between MeanFlow (left) and MeanFlow with SubFlow (right) on ImageNet-256. Each row shows five samples from the same class. Green boxes highlight samples where SubFlow produces visibly higher image quality with sharper details and fewer artifacts.
Method
NFE
FID ↓
Precision (%)
Recall (%)
MeanFlow
1
6.17
70.35
43.45
MeanFlow
2
5.46
80.82
38.91
MeanFlow
4
5.54
82.86
37.59
MeanFlow
8
5.58
83.35
37.62
MeanFlow
16
5.60
83.27
37.83
MeanFlow
32
5.64
83.12
38.16
Table 2. Effect of the number of sampling steps (NFE) on MeanFlow-B/2 with and without SubFlow ( K=5 ). Multi-step integration could improve FID but Recall slightly decreases, confirming that reduced discretization error amplifies rather than corrects the dominant-mode bias. SubFlow provides consistent Recall gains at all step counts.
Figure 4. Diverse generation from the same noise x0 . First column: MeanFlow output. Columns 2–6: MeanFlow+SubFlow with different sub-mode indices k .
Figure 5. Sensitivity analysis for the number of sub-modes K . Dashed lines indicate the baseline without SubFlow. Recall improves steadily with moderate K , while Precision remains stable and FID stays competitive across a broad range.
Strategy
FID ↓
Prec.(%)
Recall(%)
Baseline (no SubFlow)
18.70
55.23
43.51
Random sub-mode assignment
18.80
55.45
43.92
Uniform sampling
20.00
52.86
48.22
SubFlow
18.49
54.21
46.54
Table 3. Ablation on sub-mode clustering and sampling strategy (MeanFlow, DiT-B/4, 80 epochs, K=5 ). SubFlow ( K -Means clustering and prior-based sampling p(k∣c) ) achieves a better quality–diversity trade-off than random sub-mode assignment and uniform sampling.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Clustering Algorithm
FID ↓
Prec.(%)
Recall(%)
Baseline (no SubFlow)
18.70
55.23
43.51
DBSCAN
19.49
54.65
44.67
K -Means (SubFlow)
18.49
54.21
46.54
Appendix
Table 4. Comparison of clustering algorithms for sub-mode discovery (MeanFlow, DiT-B/4, 80 epochs). K -Means yields better FID and comparable Recall, suggesting that evenly partitioned sub-modes are more effective for flow matching conditioning.
Strategy
FID ↓
Prec.(%)
Recall(%)
Baseline (no SubFlow)
18.70
55.23
43.51
SubFlow (drop both c and k )
18.59
54.79
46.25
SubFlow (drop c only)
18.49
54.21
46.54
Appendix
Table 5. Effect of dropping the sub-mode index k during CFG training (MeanFlow, DiT-B/4, 80 epochs, K=5 ). Both c and k are dropped independently with the same probability. Dropping c only (our default) achieves the best overall trade-off.
Figure 6. t-SNE visualization of DINOv3 features for ImageNet classes at opposite ends of the HDBSCAN clustering spectrum. Classes in (a) are well served by standard class conditioning, while classes in (b) suffer from averaging distortion that SubFlow’s sub-mode conditioning resolves.
Flow matching models have emerged as a powerful framework for realistic image generation by learning to reverse a corruption process that progressively adds Gaussian noise. However, because noise is injected in the latent domain, its impact on different frequency components is non-uniform. As a result, during inference, flow matching models tend to generate low-frequency components (global structure) in the early stages, while high-frequency components (fine details) emerge only later in the reverse process. Building on this insight, we propose Frequency-Aware Flow Matching (FreqFlow), a novel approach that explicitly incorporates frequency-aware conditioning into the flow matching framework via time-dependent adaptive weighting. We introduce a two-branch architecture: (1) a frequency branch that separately processes low- and high-frequency components to capture global structure and refine textures and edges, and (2) a spatial branch that synthesizes images in the latent domain, guided by the frequency branch's output. By explicitly integrating frequency information into the generation process, FreqFlow ensures that both large-scale coherence and fine-grained details are effectively modeled low-frequency conditioning reinforces global structure, while high-frequency conditioning enhances texture fidelity and detail sharpness. On the class-conditional ImageNet-256 generation benchmark, our method achieves state-of-the-art performance with an FID of 1.38, surpassing the prior diffusion model DiT and flow matching model SiT by 0.79 and 0.58 FID, respectively. Code is available at https://github.com/OliverRensu/FreqFlow.
Flow Matching enables high-quality visual generation via continuous-time dynamics, but inference remains costly due to multiple sequential function evaluations. Existing acceleration methods reduce the number of function evaluations but often introduce additional training overhead, degrade quality, or fail to account for input-dependent variability. We propose COFLOW, an inference-time method that adaptively selects the step counts each generation based on the prompt features. Our context-aware COFLOW is trained online with an unsupervised reward that balances inference efficiency and generation fidelity. Our method is plug-and-play, requiring no retraining of the underlying generative model. It generalizes to image and video generation, achieving over 2.5x speedup while preserving perceptual and semantic quality. We further provide a theoretical analysis establishing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35-50 to 4-8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flow-matching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.
Chuyang Zhao, Yifei Song, Hongfa Wang +7
1Joy Future Academy · 2Fudan University · 3Tsinghua University +1