Flow matching (FM) generators often face a quality-coverage tradeoff. Inference-time techniques such as classifier-free guidance raise quality, but at the cost of coverage. We hypothesize this limitation stems from a basic property of the objective: an intermediate noisy image could correspond to many clean images, yet the model can only predict a single velocity, at best averaging over the multiple possibilities. To mitigate this, we introduce competition among predictions during training. We achieve this by augmenting the predictor with a randomly drawn low-dimensional conditioning code z. For each training example, we sample K values of z to produce K candidate velocities that compete to explain the target, and apply the loss only to the best-matching candidate. This winner-take-all objective allows different modes to shape different predictions, rather than being averaged into a single compromise. Furthermore, this term naturally enables a new form of guidance, auxiliary conditioning scaling, which complements traditional guidance methods. Together, these two guidance mechanisms show that our method substantially improves the quality-coverage frontier. We validate on ImageNet-256, using two different testbeds, SiT and RAEv2. Given less than 3% of the original SiT pretraining compute for post-training, at the 67% Recall point, our method reduces FDr,6 from 19.22 to 5.91. On RAEv2, at the 70% Recall point, it reduces FDr,6 from 16.83 to 3.18.
Figures & tables
Figure 2: A 1D toy illustrating conflicting velocity targets in flow matching. (a) Transport from a unimodal source to an imbalanced bimodal target. (b) Standard FM produces off-support mass and incomplete mode coverage. (c) Winner-take-all (WTA)-trained FM sampled with auxiliary conditioning recovers both target modes while largely removing off-support mass; two fixed codes are shown to visualize distinct code-conditioned trajectories. (d) Sampling the same WTA-trained model without the auxiliary conditioning instead concentrates generation on the dominant mode.
Figure 3: Auxiliary conditioning and WTA training. (a) A Gaussian code zk is mapped through cϕ to condition the predictor vθ . (b) WTA compares K predictions from one shared model against the same target. Only the best match backpropagates.
Figure 5: Guidance expands our model’s Pareto frontier toward broad coverage. (a) AutoGuidance trajectories for pretrained SiT and our post-trained model in FDr,6 –Recall space. For each fixed λ , we sweep γ=0,0.25,…,1.5 , with successive points ordered from top to bottom. (b) The resulting Pareto frontier obtained by combining trajectories across λ , compared with SiT. (c) Comparison with recent advances in flow matching over their respective guidance sweeps.
Table 1: Guided comparison under the standard 50K-sample evaluation. (a) Performance at each method’s lowest FDr,6 . (b) Lowest FDr,6 at Recall ≈67% ; light-gray Recall values mark methods whose highest measured Recall is below 67% . The six raw Fréchet distances (including ADM FID) appear in smaller gray type; FDr,6 averages their normalized values.
Figure 7: Our method further improves RAEv2. Left: FDr,6 –Recall frontiers (10K samples): our method extends beyond pretrained RAEv2, most at broad coverage. Right: On the 50K-sample evaluation, our method improves the best overall FDr,6 with a larger gain near 70% Recall.
SiT
JiT
RAE
RAEv2
Metric
Pre-trained
Ours
Pre-trained
Ours
Pre-trained
Ours
Pre-trained
Ours
FDr,6↓
15.76
10.59
15.17
14.15
4.41
3.64
3.89
3.52
Recall (%) ↑
67.65
62.83
66.46
62.03
63.80
58.73
62.08
55.49
Precision (%) ↑
68.71
75.70
68.06
72.05
79.13
83.53
79.84
85.76
Inception Score ↑
131.86
178.47
151.49
168.01
240.36
269.29
226.72
278.71
Table 2: Base model evaluation. For each generator, we compare the pretrained model with its base model after our post-training ( K=8 ), evaluated at λ=0 . Each model is evaluated following the inference and evaluation protocol of its original work (Appendix A.1.7 ).
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
SiT
JiT
RAE
RAEv2
Global batch size
256
1024
1024
1024
Post-training epochs
10
10
10
10
Main learning rate
10−4
2×10−4
2×10−4
5×10−5
EMA decay
0.9999
0.9999
0.9995
0.9995
Gradient-norm limit
—
—
1
1
Random seed
0
0
42
42
Appendix
Table 3: WTA post-training configurations for the four backbones.
λ
γ
Recall (%) ↑
FDr,6↓
−0.50
1.75
70.73
6.071
−1.50
1.50
75.79
7.992
−1.75
1.50
77.04
8.788
−2.00
1.50
77.98
9.658
Appendix
Table 4: Representative 10K operating points for our guided frontier. Our K=8 model with 50-step Euler sampling.
Model
Guidance
Best FDr,6
Recall ≈67%
CAFM
CFG
0.90
0.05
REPA
CFG
1.05
0.05
MixFlow
CFG
1.60
0.00†
Self-Flow
CFG
0.95
0.00†
StableVM
CFG
1.00
0.00
DeepFlow
CFG
1.10
0.00
Appendix
Table 5: Guidance settings for Table 1 . Entries report γ , or (λ,γ) for our K=8 model. For CFG, the conventional scale is s=1+γ . Best settings minimize FDr,6 among evaluated 50K configurations. A dagger denotes a below-target reference.
Figure 10: Precision–Recall comparison in the high-coverage regime. Results on ImageNet-256 using 10K generated samples per operating point and a 50-step Euler ODE solver. Curves summarize guidance sweeps; WTA also varies the auxiliary-conditioning scale. CFG and AG denote classifier-free guidance and AutoGuidance, respectively. Higher Precision and Recall are better.
Figure 11: Additional parameters without competition do not reproduce the high-Recall WTA frontier. Capacity-matched K=1 controls compared with pretrained SiT+AG and our K=8 model. The K=1 generator is paired with either a K=1 post-trained or an original pretrained SiT-B/2 guide. (a) Overall comparison. (b) High-Recall region. All results use 10K samples and 50-step Euler sampling.
Model
FID ↓
IS ↑
Pretrained SiT-XL/2
14.41
118.60
FM continuation, 50 epochs
14.31
120.90
Appendix
Table 6: Continued standard FM changes SiT-XL/2 only modestly. 10K samples, 50-step Euler sampling, γ=0 .
Model / training batch size
FID ↓
IS ↑
Pretrained SiT-B/2
33.40
51.42
FM continuation, batch 256
32.84
53.45
FM continuation, batch 1024
31.76
56.18
FM continuation, batch 8192
32.81
53.20
Appendix
Table 7: Training batch size in standard-FM continuation. SiT-B/2 after 20 additional data epochs, evaluated using 10K samples, 50-step Euler sampling, and γ=0 . The pretrained row is the common initialization before continuation.
Figure 12: Both signs of auxiliary conditioning improve the high-coverage frontier. K=8 guided frontiers for positive and negative λ , compared with pretrained SiT AutoGuidance baseline. Both signs improve the frontier in the high-coverage region. All results use 10K samples and 50-step Euler sampling.
Figure 13: Competition changes the response to auxiliary conditioning. (a) FDr,6 , (b) FID, and (c) Inception Score for K∈{1,4,8,16} with γ=0 . The dashed horizontal lines denote pretrained SiT. All curves use the same λ values and 10K samples with 50-step Euler sampling.
Figure 14: The high-coverage frontier improvement persists across competition levels. Measured guided frontiers for K=4,8,16 , with pretrained SiT AutoGuidance and its γ=0 point as references. The plot shows the high-coverage region. All results use 10K samples and 50-step Euler sampling.
Figure 15: The RAEv2 frontier improvement persists across competition levels. Measured FDr,6 –Recall frontiers for K=8,16 and pretrained RAEv2 with its native guidance. Each setting uses 10K samples and native Euler-50 sampling with timestep shift 8. Two pretrained-reference points have FDr,6 values above the plotted range.
Figure 16: Varying auxiliary conditioning under two sampling policies ( γ=0 ). SiT-XL/2 WTA ( K=8 ), 10K images per setting, 50 Euler updates. Panels (a–c) show FDr,6 , Recall, and Precision versus auxiliary-conditioning scale; dashed black horizontal lines show pretrained SiT at γ=0 . Panel (d) shows the same observations in Recall– FDr,6 space, with a black star for SiT. Gray diamonds denote the shared post-trained base model ( λ=0 ). In (d), filled and open policy markers denote negative and positive scales, respectively; endpoint labels give λ .
Figure 17: Both code policies improve the guided frontier, with a larger broad-coverage gain from fixed codes. All results use 10K samples and 50-step Euler sampling. In (a,b), pale diamonds connect the γ=0 operating points; pale paths vary γ at fixed λ , with arrows pointing toward increasing guidance. Solid and dashed pale paths denote nonpositive and positive λ , respectively; gray denotes the post-trained base model ( λ=0 ). (c) compares these frontiers directly with original SiT+AG, using the same axis ranges.
Figure 18: Samples generated with the same auxiliary code share recurring appearance patterns across source noises and classes. Each row fixes one auxiliary code z and each column one ImageNet class. The four samples at each row–column intersection vary only the initial noise ϵ . Both panels use the same SiT-XL/2 WTA checkpoint, the same four auxiliary codes, and 50-step ODE sampling at λ=1 , γ=0 , holding the code fixed throughout each trajectory. Repeated tendencies in color, background, and composition appear both within and across classes.
Figure 24: Positive and negative auxiliary-conditioning scales. Columns show λ∈{−1.5,−1,−0.5,0,0.5,1,1.5} . Within each class, the initial noise is fixed; within each row, the auxiliary code is fixed. All samples use γ=0 , SiT-XL/2 K=8 , and 50-step Euler sampling.
Figure 25: Positive and negative auxiliary-conditioning scales on additional classes. The layout and sampling settings follow Figure 24 .
Figure 26: Positive and negative auxiliary-conditioning scales on additional classes. The layout and sampling settings follow Figure 24 .
Quantity
Original pretraining
K=8 post-training
Training epochs
—
10
Optimizer updates
7,000,000
50,040
Global batch size
256
256
Example presentations
1,792,000,000
12,810,240
Relative update budget
100%
0.715%
Relative model-core computation
100%
2.63%
Appendix
Table 8: Training budget for SiT-XL/2. The pretraining schedule follows Ma et al. (2024) . Model-core computation uses the analytical FLOP accounting in Equation 19 .
Method
Candidate chunk
Images/s ↑
GiB/GPU
Ordinary FM
—
834.51
36.55
Ours, K=1
32
656.86
39.66
Ours, K=2
64
581.01
39.67
Ours, K=4
128
474.36
39.66
Ours, K=8
256
343.75
39.66
Ours, K=16
512
224.12
40.96
Appendix
Table 9: Training throughput and peak reserved GPU memory on eight H100 80GB GPUs, using the highest-throughput tested configuration for each method. Throughput is reported across all eight GPUs, and memory per GPU.
Flow matching (FM) trains a time-dependent vector field that transports samples from a simple prior to a complex data distribution. However, for high-dimensional images, each training sample supervises only a single trajectory and intermediate point, yielding an extremely sparse and high-variance training signal. This under-constrained supervision can cause flow collapse, where the learned dynamics memorize specific source-target pairings, mapping diverse inputs to overly similar outputs, failing to generalize. We introduce Posterior-Augmented Flow Matching (PAFM), a theoretically grounded generalization of FM that replaces single-target supervision with an expectation over an approximate posterior of valid target completions for a given intermediate state and condition. PAFM factorizes this intractable posterior into (i) the likelihood of the intermediate under a hypothesized endpoint and (ii) the prior probability of that endpoint under the condition, and uses an importance sampling scheme to construct a mixture over multiple candidate targets. We prove that PAFM yields an unbiased estimator of the original FM objective while substantially reducing gradient variance during training by aggregating information from many plausible continuation trajectories per intermediate. Finally, we show that PAFM improves over FM by up to 3.4 FID50K across different model scales (SiT-B/2 and SiT-XL/2), different architectures (SiT and MMDiT), and in both class and text conditioned benchmarks (ImageNet and CC12M), with a negligible increase in the compute overhead. Code: https://github.com/gstoica27/PAFM.git.
George Stoica, Sayak Paul, Matthew Wallingford +6
Georgia Tech · University of Washington · Hugging Face +2
While Flow Matching theoretically guarantees constant-velocity trajectories, we identify a critical breakdown in high-dimensional practice: the Velocity Deficit. We show that the MSE objective systematically underestimates velocity magnitude, causing generated samples to fail to reach the data manifold-a phenomenon we term Integration Lag. To rectify this, we propose Initial Energy Injection, instantiated via two complementary methods: the training-based Magnitude-Aware Flow Matching (MAFM) and the training-free Scale Schedule Corrector (SSC). Both are grounded in our discovery of a crucial asymmetry: velocity contraction causes harmful kinetic stagnation at the trajectory's start, yet acts as a beneficial denoising mechanism at its end. Empirically, SSC yields significant efficiency gains with zero retraining and just one line of code. On ImageNet-1k (256x256), it improves FID by 44.6% (from 13.68 to 7.58) and achieves a 5x speedup, enabling a 50-step generator (FID 7.58) to beat a 250-step baseline (FID 8.65). Furthermore, our methods generalize to Text-to-Image tasks and high-resolution generation, improving FID on MS-COCO by ~22%.
Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to be learned implicitly. In this paper, we introduce Energy-Guided Flow Matching(EG-FM) that explicitly models a coarse-to-fine generative trajectory by moving endpoint. Specifically, EG-FM replaces the fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from low-frequency image to clean image. The fraction of high-frequency signal in moving endpoint is released by an image-specific energy-guided scheduling, leading to the re-targeting of velocity in flow matching. Our framework requires no adaptation of the backbone and training data, bringing negligible cost on the training and inference stages. In our experiment, EG-FM consistently achieves lower FID on the ImageNet class-conditional image generation task at 256×256 with fewer epochs, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. We continue training the generation task on the setting of 512×512 resolution, yielding a FID of 1.58 after only 40 high-resolution adaptation epochs. Furthermore, we transfer EG-FM on text-to-image generation and achieve 0.85 on GenEval score and 83.9 on DPG-Bench. Code is available at https://github.com/ysng123/EG-FM.
Haoyang Tong, Yu He, Fang Li +6
MAIS & NLPR, CASIA · JD.com · Xi'an Jiaotong University +1