acceptodds
Under review as a conference paper at ICLR 2027

Expanding the Quality–Coverage Frontier with Competitive Flow Matching

Abstract

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.

Primary area:
generative models
Keywords:
flow matching, image generation, winner-take-all learning, quality–coverage trade-off, post-training

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VYnvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Competitive Flow Matching, which introduces multiple competing velocity predictions during Flow Matching training, allowing the model to learn different generation modes and control the trade-off between generation quality and diversity through auxiliary conditioning.

However, applying WTA to Flow Matching offers limited methodological novelty, and its advantages over other post-training methods remain to be further investigated. According to the experimental results, the proposed method performs well in high-coverage generation settings, but does not show substantial improvements in overall performance comparisons.

Critical strengths
  1. The paper addresses the issue of conflicting velocity targets in Flow Matching by introducing Winner-Take-All (WTA) training, allowing different auxiliary conditions to predict different velocities and avoiding simple averaging over multiple targets. The overall method is intuitive, easy to understand, and relatively easy to apply to existing Flow Matching models.

  2. The paper focuses not only on generation quality but also on the coverage of the generated distribution. On ImageNet-256, Competitive FM significantly outperforms existing methods, particularly in high-recall settings.

  3. The paper validates the proposed method on different architectures, including SiT and RAEv2, and conducts ablation studies on the number of competing candidates KK, auxiliary conditioning scale, and code sampling strategy.

Critical weaknesses
  1. The core algorithm of this paper applies the existing Winner-Take-All (WTA) learning strategy to velocity regression in Flow Matching. However, similar multi-candidate competition mechanisms have already been extensively studied in previous WTA-based methods. This paper mainly adapts WTA to Flow Matching by introducing auxiliary conditioning. Therefore, the methodological novelty of these improvements is somewhat limited.

  2. The authors argue that predicting the conditional mean velocity in standard Flow Matching leads to averaging over different target velocities, thereby affecting generation quality and coverage. However, these assumptions lack sufficient theoretical justification. The optimal velocity field learned by standard Flow Matching can correctly generate the target distribution under ideal conditions. Therefore, velocity averaging itself may not be the fundamental cause of the quality–coverage trade-off.

  3. The paper lacks comparisons under equal computational budgets. The design involving multiple candidate velocities introduces additional computational overhead, making training FLOPs a potential concern. The authors also do not investigate whether the benefits of WTA outweigh those of allocating the same computational budget to other post-training methods.

Questions for authors
  1. Please provide the training FLOPs required by the multiple candidate velocity design. If possible, it would be helpful to include comparisons with other post-training methods.

  2. Please provide stronger theoretical justification for whether velocity averaging is indeed the cause of the quality–coverage trade-off. Clarifying this point is particularly important for my decision on whether to increase score.

4 comments
AuthorsOpenReview ↗

Computational cost analysis: our gains do not rely on a larger training budget

We thank the reviewer for raising this important concern. We provide detailed analytical FLOP accounting, measured runtime, cross-method cost comparisons in Revised Supplementary Sec. J (Tables 8–9).

  • Computational cost vs. standard FM. Our method is a post-training procedure that does not require training from scratch. We evaluate K candidate velocities without gradients and backpropagate only through the winner. Under standard analytical FLOP accounting, K=8 requires 3.674× the model-core training FLOPs per update of standard FM. Over ten epochs, this totals 33.49 EFLOPs. In practice, our optimized implementation has a lower measured runtime overhead of 2.441× (0.500 vs. 0.205 s/update). The full post-training stage takes approximately ~6.95 hours on a single 8×H200 node (55.57 H200 GPU-hours) of projected active training, corresponding to just 2.63% of SiT's original pretraining FLOPs.

  • Allocating the same or greater compute to standard FM does not reproduce the gain. The runtime budget of our ten-epoch post-training would allow approximately 24.4 additional epochs of standard FM. However, even 50 additional epochs of FM continuation leave FID essentially unchanged (14.41 → 14.31; Main Paper Appendix B.2). Increasing the training batch size also yields no consistent improvement. In contrast, our post-training achieves substantial gains within ten epochs.

  • Alternative post-training methods use more compute yet do not match our quality–coverage frontier. While many recent methods require training from scratch, we focus on two representative post-training methods, CAFM and MixFlow. We measure the per-update runtimes of their released implementations under the same eight-H200 protocol and combine them with the official training schedules used to obtain their reported final checkpoints. CAFM's schedule includes discriminator updates and Jacobian-vector products and corresponds to 121.60 H200 GPU-hours (2.19× ours). MixFlow's final checkpoint is trained for 200 epochs, corresponding to 454.05 H200 GPU-hours (8.17× ours) (Revised Supplementary Sec. J, Table 8). Thus, the official schedules used to produce these methods' reported final checkpoints already require much more post-training compute than ours, yet their quality–coverage frontiers remain behind ours. Our advantage therefore does not come from a larger post-training budget.

  • One epoch of efficient post-training further advances stronger generators. Just one additional epoch of our post-training further expands the quality–coverage frontiers of already stronger CAFM and REPA generators (Revised Supplementary Sec. F, Fig. 2). For reference, one epoch at our profiled SiT training rate corresponds to approximately 5.56 H200 GPU-hours (~42 minutes on one 8×H200 node).

AuthorsOpenReview ↗

Novelty: Simply adapting WTA to Flow Matching cannot reproduce the large gain.

We thank the reviewer for raising this question, but our contribution is not simply applying WTA to velocity regression. We identify an opportunity to exploit the sample-wise target variation that standard FM does not explicitly retain, and develop competitive post-training to turn this variation into a new code-based, coverage-oriented inference-time control for FM generators. This enables different generative behaviors within a single model, complements guidance, and expands the range of accessible quality–coverage operating points.

We recognize that WTA and multiple-choice learning have been extensively studied. However, directly adapting established WTA formulations to Flow Matching does not automatically yield these improvements. We investigate two representative alternatives, with detailed methods and results provided in Revised Supplementary Sec. I, Figs. 5–6.

  • Established multi-head WTA (MHP) + FM We adapt the well-established Multiple Hypothesis Prediction formulation (Rupprecht et al., ICCV 2017), which implements WTA through multiple independent prediction heads, to FM using eight velocity heads on a shared SiT trunk. Unlike our method, it learns no code-conditioned response whose strength can be scaled by λ, and therefore provides no continuous coverage control that complements guidance. The baseline trains stably, keeps all heads active, and uses approximately 14× as many additional parameters as our auxiliary branch. Nevertheless, at nearly matched Recall, MHP with AutoGuidance obtains FDᵣ,₆=15.58 at 76.12% Recall, compared with 8.31 at 76.46% Recall for our method.

  • Noisy-State WTA + FM Another natural way to introduce competition into FM is to compete among different noisy realizations of the same training example, rather than competing among code-conditioned predictions at the same noisy state as in our formulation. Specifically, For each clean training example, it samples K independent noise realizations to construct K different noisy states, then updates the model only through the realization with the smallest FM regression loss. The method learns no auxiliary code-conditioned fields and provides no inference-time control. Empirically, it only modestly improves the underlying FM predictor before saturating early, while its guided results remains close to the pretrained SiT baseline and does not reproduce our quality–coverage frontier.

Beyond the performance gains, we contribution also includes a systematic study of this inference time control and the generative behaviors induced by competitive training. Our analyses cover the joint effects of auxiliary-conditioning strength and guidance on the quality--coverage frontier; different code-sampling policies; how competition strengthens the shared base field even without auxiliary conditioning; and the effects of scaling competition strength K. We further demonstrate generalizability across flow-generator families (SiT, JiT, and RAE) and stronger pretrained generators (CAFM and REPA), whose frontiers improve further after just one additional epoch of our post-training (Revised Supplementary Sec. F, Fig. 2).

AuthorsOpenReview ↗

Follow-up to Reviewer VYnv: Summary of Our All Responses

Thank you again for your thoughtful review! We hope our responses have been helpful. With the discussion period ending soon, we wanted to briefly recap how we addressed your main concerns.

Our Reply to All Reviewers clarifies the overall motivation and the role of the 1D toy, and provides a mechanistic analysis of our method. The new Additional Supplementary Material provides the full experimental details supporting our rebuttal. We also address your questions as follows:

1. Do we mean that averaging is wrong in Flow Matching?
No. We want to clarify our motivation. A noisy state can correspond to multiple target velocities. Under L2, these velocities are optimally averaged into their conditional mean. With ideal modeling and sufficiently fine integration, this marginal field can correctly recover the full multimodal distribution.

What we claim is: in practice, however, generation relies on approximate learned fields and finite sampling budgets. As illustrated by our toy, under these practical constraints, relying on a single mean field to cover multiple modes can leave samples off-manifold and valid modes underrepresented. In image generation, guidance can bring samples back toward high-density regions, but often at the cost of coverage.

We therefore ask: rather than relying on a single marginal field to cover all modes, can different modes be accommodated by different predictions? Competitive FM realizes this idea through code-conditioned predictions, and our experiments show that it improves both fidelity and coverage in practical generation settings.

We thank the reviewer for raising this important question and will revise the paper to make this distinction explicit and avoid any misunderstanding.

2. Is our method simply a direct adaptation of WTA to Flow Matching?
No. We implemented two representative direct adaptations—classical multi-head WTA and Noisy-State WTA. Both train successfully and provide improvements in selected settings, but neither reproduces our inference-time control or quality–coverage frontier.

Our contribution is not the WTA rule itself, but how we use it to uncover and exploit valuable learning signals in Flow Matching. Auxiliary codes allow different targets to shape specialized code-conditioned velocity fields. Fixing a code makes these responses compose coherently along a trajectory, while using and scaling different codes broadens population coverage and enables additional control over quality–coverage behavior within a single model. This control changes how the model responds to guidance, turning the usual fidelity–coverage trade-off into win-win regimes where fidelity improves without the usual loss of coverage. Together, these effects substantially expand the quality–coverage frontier beyond all the settings we evaluate.

The method also generalizes to stronger pretrained generators, further improving CAFM and REPA after only one epoch of post-training. Beyond the empirical gains, we provide a systematic study of the method, characterizing fixed-code coherence, auxiliary scaling, population-level code mixing, guidance interaction, and scaling with K.

3. Are the gains simply due to additional training compute?
No. Our entire ten-epoch post-training stage uses only 2.63% of SiT’s original pretraining FLOPs. Allocating the same or greater compute to standard FM does not reproduce the gain, and the official CAFM and MixFlow post-training schedules use substantially more compute while remaining behind our quality–coverage frontier. Our advantage therefore does not come from a larger post-training budget.

AuthorsOpenReview ↗

Follow-up Clarification on Conditional Averaging and the Quality–Coverage Trade-off

We sincerely thank the reviewer for emphasizing this important question and for explaining that it is particularly relevant to your assessment. In our Reply to All Reviewers, we provided three complementary discussions: how conditional averaging should be understood in our motivation, what the 1D toy is intended to illustrate, and a detailed toy-based analysis of the mechanism induced by our method. We would like to follow up here with a concise, integrated clarification, while referring to those responses and the Additional Supplementary Material for the complete analyses.

  • Our intended claim is not that velocity averaging is the fundamental cause of the quality–coverage trade-off. Conditional averaging is the correct marginal solution of standard FM: with ideal modeling and sufficiently fine integration, the resulting mean field can recover all modes of the target distribution. Our motivation instead concerns practical generation, where models rely on approximate learned fields and finite sampling budgets. Under these constraints, trajectories following a single marginal field may pass through low-density regions, leaving some samples off-manifold and valid alternatives underrepresented. Although post-hoc guidance can bring samples back toward the data manifold, it often does so by concentrating generation on high-density regions, leading to mode collapse and reduced coverage. This motivates us to explore a different route: can we model modes at each intermediate timestep to cover the full modes of the distribution? Our 1D toy is used to illustrate this distinction. For standard FM, off-support mass decreases from 4.09% with 4 Euler steps to 1.95%, 1.27%, and 0.76% with 8, 16, and 32 steps, respectively, and further decreases to 0.51% with 2,000 steps. This confirms that standard FM can recover the target distribution increasingly well as integration becomes finer. In contrast, Competitive FM achieves 0.41% off-support mass with only 4 steps while covering both modes, showing that, even under such a limited sampling budget, its more mode-directed predictions can reduce off-support generation while retaining broad target-mode coverage.

  • The toy also mirrors the behaviors observed in image generation. Different auxiliary codes support distinct generative behaviors, while scaling the auxiliary conditioning provides a smooth additional control over population coverage. Our image-generation results further show that this different route can improve quality at broad coverage and expand the accessible quality–coverage frontier. In particular, after our post-training, traditional guidance no longer improves quality in the direction of sacrificing coverage, instead, it can improve quality while largely preserving coverage. We consistently observe these benefits across different pretrained generators, model families, and guidance mechanisms.

  • We further provide a detailed mechanism analysis in our general response and Additional Supplementary Material. At ambiguous noisy states, competition causes different code-conditioned predictions to specialize toward different valid target directions. Holding a code fixed allows these local responses to compose consistently over time, whereas disrupting code identity weakens trajectory coherence and increases off-support generation. Across a population, using different codes covers different target modes, while the auxiliary-conditioning scale smoothly controls the amount of expressed variation. Together, these results provide a mechanistic picture of how sample-wise target variation—although not required for the correctness of ideal marginal FM—can serve as a valuable additional learning signal under practical generation conditions.

We have incorporated this clarification into the revised abstract and introduction and added the finite-step comparison as a footnote to Figure 2, with all changes highlighted in blue. We hope these revisions clarify our intended motivation. As the discussion period draws to a close, we would greatly appreciate any further feedback and would be happy to clarify any remaining questions. We hope this clarification will be helpful as you finalize your assessment.

Reviewer_5A8WOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates the inherent fidelity–coverage trade-off in Flow Matching, attributing the degradation of distribution coverage under inference-time guidance to the standard regression objective, which collapses multimodal velocity targets at a given intermediate state into a single conditional expectation. To address this, the authors propose Competitive Flow Matching, a lightweight post-training paradigm that conditions a pretrained backbone on a low-dimensional stochastic latent code to instantiate multiple candidate velocity predictions per noisy state, optimizing parameters exclusively via the loss-minimizing realization under a Winner-Take-All objective. The experiments show that the proposed method substantially improves the quality–coverage frontier.

Critical strengths
  1. By adding a small random code and training with a "winner-take-all" rule, the authors found a very cheap way to upgrade existing pretrained models. This is important because it takes less than 3% of the original training cost, adds no extra work during image generation, and consistently improves four very different types of image generators.
  2. Normally, using guidance to make images look better hurts diversity. With this method, increasing guidance makes images look much better while keeping or even increasing diversity.
Critical weaknesses
  1. The proposed approach appears to be a direct combination of two existing techniques: Winner-Take-All (WTA) training with random latent codes (e.g., Multiple Choice Learning, cIMLE) and standard inference-time guidance (e.g., AutoGuidance, CFG). While this combination works well in practice, the authors need to better clarify their core technical contribution beyond simply putting these two known methods together to improve diversity.

  2. Instead of retraining the model, a standard training-free way to fix the averaged trajectory problem and boost diversity in Flow Matching is simply switching from deterministic ODE sampling to stochastic SDE sampling at test time. By restricting all baselines in the main comparisons to ODE sampling without testing whether tuning SDE noise alongside guidance can achieve similar results, the paper fails to fully justify the need for WTA post-training.

Questions for authors

During training, the auxiliary code is always injected with a positive scale (λ=+1\lambda = +1). Why do negative scales (λ<0\lambda < 0, which flip the learned modulation direction backward) consistently achieve a better quality–coverage frontier than positive scales at test time (Figure 4a and Figure 11)?

4 comments
AuthorsOpenReview ↗

Sweeping SDE noise and guidance does not recover our frontier

We thank the reviewer for pointing out stochastic SDE sampling as an important training-free baseline.

We jointly sweep SDE diffusion strength and Autoguidance scale on pretrained SiT-XL/2, using a coarse-to-fine search (66 configurations in total) under the same evaluation protocol as in paper (Revised Supplementary Sec. H, Fig. 4a). The resulting SDE frontier remains broadly comparable to the standard ODE baseline and substantially behind ours at high coverage.

Under 50k evaluation, none of the SDE results reaches 67% Recall. The closest stochastic SDE point achieves 66.44% Recall, FDᵣ,₆=11.06, and ADM FID =4.87. In comparison, WTA under ODE sampling achieves higher Recall (67.45%) together with substantially better FDᵣ,₆ (5.91) and ADM FID (2.20).

We observe the same conclusion on RAEv2 using its native guidance. All evaluated positive-diffusion configurations are Pareto-dominated by measured WTA-RAEv2 ODE operating points (Revised Supplementary Sec. H, Fig. 4b).

AuthorsOpenReview ↗

Clarifying Our Technical Novelty and Contributions

We thank the reviewer for raising this important question. We acknowledge the connections to existing WTA learning, cIMLE, and inference-time guidance, but we do want to clarify our method cannot be simply summarized as combining WTA for diversity and guidance for quality.

  • Our contribution is not the WTA rule itself, but how we use it to uncover and exploit a valuable learning signal in Flow Matching. We identify that the sample-wise velocity target variation, which is averaged out in standard FM's single marginal velocity, can be particularly useful for generation: (a) improving fidelity, by providing clues about the different valid directions an otherwise ambiguous noisy state can take, and (b) improving coverage, by preserving these alternatives as distinct generative behaviors. We use auxiliary conditioning to capture this variation, with WTA serving as the implicit assignment mechanism that allows different codes to specialize and capture the variation.

  • Different from existing WTA/IMLE methods, we investigate how these code-conditioned velocity fields can compose into coherent Flow Matching trajectories and be continuously scaled to enable controllable generative behaviors. This allows the generator to access quality--coverage operating points previously inaccessible to the pretrained model. Importantly, even without any guidance, our post-trained model already achieves a stronger and smoothly controllable quality--coverage curve (Main Paper Fig. 4), including operating points that improve both quality and coverage over pretrained FM. With only ten epochs of post-training, even the shared base generator, with auxiliary conditioning disabled, improves unguided FID from 8.61 to 4.02 (K=8), and gets even stronger (FID from 8.61 to 2.90 when K=32) when the competition scales (Main paper, figure 8).

  • We want to clarify that our improvements do not arise from simply combining WTA with guidance, rather, it is our method that enables guidance to operate in a fundamentally different way. Without guidance, our post-trained generator is already stronger and provides a continuously controllable range of quality--coverage operating points. Whereas a standard generator starts from a single unguided point, with guidance typically improving fidelity at the cost of coverage, our method provides an entire range of coverage-controllable operating points on which guidance can act. This new control is complementary to guidance, allowing it to substantially improve fidelity while preserving coverage, and even increase coverage and diversity at high-coverage operating points (Main Figs. 5–6).

Beyond the performance gains, we systematically study this new inference-time control and the generative behaviors induced by competitive training. Our analyses investigate mode-seeking and mode-covering behaviors within a single model, the complementary effects of auxiliary conditioning and guidance, code-sampling policies, how competition strengthens the auxiliary-off generator, and scaling with competition strength (K). We also compare against classical multi-head WTA (MHP) and Noisy-State WTA, neither of which reproduces our high-coverage gains (Revised Supplementary Sec. I, Figs. 5–6). Finally, we demonstrate generalizability across FM generator families (SiT, JiT, and RAEv2), and further improve stronger pretrained CAFM and REPA generators with just one additional epoch of competitive post-training (Revised Supplementary Sec. F, Fig. 2)

AuthorsOpenReview ↗

Why can negative auxiliary-conditioning scales work?

We thank the reviewer for this insightful question. We empirically find that negative scaling is effective, and show that it is a smooth and structured behavior of the learned model rather than an isolated effect. As shown in Main Fig. 4, varying λ\lambda in either direction produces smooth and broadly similar changes in generative behavior, with increasing ∣λ∣|\lambda| progressively moving toward broader coverage. In Main Appendix B.2.1 (Fig. 11), we further compare the guided frontiers obtained using positive and negative scales separately. Both improve the high-coverage frontier over pretrained SiT, with negative scales achieving somewhat stronger results.

To understand this behavior, we examine the AdaLN modulation in our SiT model. AdaLN produces scale, shift, and gating coefficients that control the transformer blocks based on time and class conditioning. Our method adds a parallel, code-conditioned pathway whose modulation coefficients are combined with the original ones. The two modulation projections have separate output biases: the pretrained native bias β\beta and the zero-initialized auxiliary bias α\alpha.

During training (λ=1\lambda=1), these biases enter additively as β+α\beta+\alpha and receive identical gradients and nearly identical optimizer updates, yielding βt−αt≈βpre\beta_t-\alpha_t\approx\beta_{\mathrm{pre}}.

Writing dtd_t for the learned common bias shift and Rt(z)R_t(z) for the centered code-dependent modulation, the effective modulation can be approximately expressed as

mλ(z)≈At(c)+βpre+(1+λ)dt+λRt(z).m_\lambda(z)\approx A_t(c)+\beta_{\mathrm{pre}}+(1+\lambda)d_t+\lambda R_t(z).

This explains why the two signs behave differently: at λ=+1\lambda=+1, the learned common shift is reinforced (2dt+Rt(z)2d_t+R_t(z)); at λ=−1\lambda=-1, it is largely canceled while the sign-reversed code-dependent variation remains (−Rt(z)-R_t(z)). Intuitively, negative scaling can therefore retain coverage-oriented variation around a more favorable fidelity-oriented base. It is not equivalent to simply sampling an opposite code, since the auxiliary response has a substantial nonzero code mean.

We further verify this explanation in the trained model. The native and auxiliary biases indeed undergo nearly identical changes during training, with only 0.043% relative mismatch in the raw checkpoint (1.82% for EMA). Moreover, the auxiliary modulation contains a substantial component shared across codes, accounting for 69.23% of its squared-norm energy. Under matched 50-step evaluation at γ=1.25\gamma=1.25, λ=−1\lambda=-1 achieves FDᵣ,₆=7.10 and Recall=73.98%, compared with 8.31 and 72.76% for λ=+1\lambda=+1. Finally, explicitly canceling the common shift while retaining code-dependent variation closely reproduces the negative branch's performance.

We will include the full analysis in the revised supplementary material.

AuthorsOpenReview ↗

Follow-up to Reviewer 5A8W: Summary of Our All Responses

Thank you again for your thoughtful review! We hope our responses have been helpful. With the discussion period ending soon, we wanted to briefly recap how we addressed your main concerns.

Our Reply to All Reviewers clarifies the overall motivation and the role of the 1D toy, and provides a mechanistic analysis of our method. The new Additional Supplementary Material provides the full experimental details supporting our rebuttal. We also address your questions as follows:

1. Novelty & Contribution: Is our method a simple combination of WTA and guidance?
No. Our contribution is not the WTA rule itself, but how we use it to uncover and exploit valuable learning signals in Flow Matching. Auxiliary codes allow different targets to shape specialized code-conditioned velocity fields. Fixing a code makes these responses compose coherently along a trajectory, while using and scaling different codes broadens population coverage and enables additional control over quality–coverage behavior within a single model. Importantly, these improvements already emerge without guidance. This new control then changes how the model responds to guidance, turning the usual fidelity–coverage trade-off into win-win regimes where fidelity improves without the usual loss of coverage. Together, these effects substantially expand the quality–coverage frontier beyond all the settings we evaluate.

We also implemented two representative direct adaptations—classical multi-head WTA and Noisy-State WTA. Both train successfully and provide improvements in selected settings, but neither reproduces our inference-time control or quality–coverage frontier. This demonstrates that simply adapting WTA to Flow Matching is insufficient to recover our advantage.

The method further generalizes to stronger pretrained generators, improving CAFM and REPA after only one epoch of post-training. Beyond the empirical gains, we provide a systematic study of the method, characterizing fixed-code coherence, auxiliary scaling, population-level code mixing, guidance interaction, and scaling with K.

2. Can stochastic SDE sampling recover the same advantage without post-training?
No. We jointly sweep SDE diffusion strength and guidance over 66 configurations. Under 50K evaluation, none reaches 67% Recall. The closest SDE point achieves 66.44% Recall with FDᵣ,₆=11.06 and FID =4.87, whereas our ODE-sampled model reaches 67.45% Recall with substantially better FDᵣ,₆=5.91 and FID =2.20. We observe the same conclusion on RAEv2, where all evaluated positive-diffusion settings are Pareto-dominated by our measured ODE operating points.

3. Why can negative auxiliary-conditioning scales outperform positive scales?
Both positive and negative scales produce smooth, structured changes in generative behavior. The negative branch can be stronger because the native and auxiliary AdaLN biases receive nearly identical updates during training. At λ=-1, their learned common shift is largely canceled while the code-dependent variation remains, preserving coverage around a more favorable fidelity-oriented base. Checkpoint diagnostics and an explicit cancellation control verify this explanation.

Reviewer_LQNeOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Competitive Flow Matching, a winner-take-all (WTA) post-training objective intended to improve the quality–coverage trade-off of flow-matching generators. The method augments the velocity predictor with a low-dimensional random code. For each training state, it samples (K) codes, produces (K) candidate velocity predictions, and backpropagates only through the candidate closest to the target velocity. At inference time, the strength of this auxiliary conditioning can be varied with a scale, providing an additional control over quality and coverage and complementing conventional guidance.

Critical strengths
  • The main empirical result is compelling: WTA substantially improves quality at matched high Recall on both SiT and RAEv2. The magnitude of the improvement is large enough that it is unlikely to be explained by minor evaluation differences alone.

  • The paper carefully controls for additional capacity, continued standard FM training, competition level (K), code-sampling policy, CFG/AutoGuidance, and training cost. These experiments make the empirical contribution substantially more convincing than a simple WTA application alone.

Critical weaknesses
  • The paper motivates WTA by arguing that averaging multiple valid velocity targets into their conditional mean creates an undesirable compromise. However, in ideal flow matching, this conditional expectation is precisely the marginal velocity field that the standard objective is intended to learn. Therefore, multimodal conditional targets do not by themselves imply an incorrect generative distribution. This issue is especially relevant to the 1D example, where standard FM is evaluated using only four Euler steps. The observed off-support mass may therefore reflect finite model capacity and/or numerical integration error rather than an inherent flaw in velocity averaging. The empirical effectiveness of WTA is convincing, but the claimed mechanism is less established.

  • The strongest results combine WTA with AutoGuidance and jointly sweep both auxiliary-conditioning scale (\lambda) and guidance strength, while many competing methods are evaluated using CFG with only a guidance-strength sweep. The controlled SiT comparison strongly supports the method itself, but the broader claim of dominating recent methods is not completely apples-to-apples. This concern is amplified by the relatively large hyperparameter search used to construct the proposed frontier. Reporting uncertainty or a more tightly matched guidance/search protocol would strengthen the comparison.

Questions for authors
  • Since the conditional-mean velocity is the optimal marginal FM solution, can the authors clarify theoretically what failure WTA is correcting? In the 1D toy experiment, does standard FM still exhibit off-support mass when using a sufficiently accurate solver with many integration steps?

  • Can the authors provide a cross-method comparison using the same guidance mechanism—for example, applying AutoGuidance to strong competing methods or comparing all methods under CFG? This would help separate the benefit of WTA from the choice of guidance.

  • Since the proposed frontier is selected over a large joint search, can the authors report uncertainty or results across training/evaluation seeds for the key 50K high-coverage comparisons? This is particularly relevant for the smaller differences near the best-quality operating point.

  • The method is evaluated almost entirely on class-conditional ImageNet-256. Do the authors have evidence that the same quality–coverage behavior persists in a more multimodal setting such as text-to-image generation? This would materially strengthen the generality of the contribution.

6 comments
AuthorsOpenReview ↗

Our advantage persists under matched AutoGuidance, and our post-training generalizes to and further improves stronger pretrained generators

We thank the reviewer for this thoughtful and important suggestion. We agree that comparing methods under the same guidance mechanism helps isolate the contribution of WTA from the choice of guidance.

  • Our advantage persists under matched AutoGuidance. Our original cross-method comparison follows the guidance protocols adopted in competing methods' papers, predominantly CFG. To provide a more direct comparison, we apply AutoGuidance to two strong baselines, CAFM and REPA (Revised Supplementary Sec. F, Fig. 1). AG improves both baselines over their CFG counterparts. Even so, our SiT-based WTA frontier lies below both AG baselines at comparable Recall and extends the Pareto envelope farther toward the lower right. Under 10k evaluation, at around 75% Recall, our method achieves FDᵣ,₆=7.99, compared with 10.52 for CAFM+AG and 15.86 for REPA+AG. This confirms that our high-coverage advantage is not simply attributable to using AG instead of CFG.

  • Our post-training generalizes to stronger pretrained generators and further advances the frontier. Our method is a general post-training procedure that can be applied to different pretrained generators. We apply it to the stronger CAFM and REPA checkpoints for only one additional epoch. As shown in Revised Supplementary Sec. F, Fig. 2, this brief post-training further expands both models' quality–coverage frontiers under AutoGuidance. At 50K evaluation, REPA+WTA achieves FDᵣ,₆=4.80 at 70.27% Recall, while CAFM+WTA achieves FDᵣ,₆=5.40 at 70.29% Recall. (Both points simultaneously improve over the highest-coverage point on the SiT-based WTA frontier reported in our paper, which achieves FDᵣ,₆=5.91 at 67.45% Recall.) These results show that our post-training can build on advances in pretrained generators to push the quality–coverage frontier even further.

AuthorsOpenReview ↗

An efficient search is sufficient to recover a strong quality–coverage frontier

Although our method introduces two inference-time controls, their effects follow a consistent and complementary pattern. This allows us to search the two controls in a structured and efficient way, without an exhaustive two-dimensional sweep. As shown in Revised Supplementary Sec. G, Fig. 3(a,b), the auxiliary-conditioning scale primarily changes the accessible Recall range, while guidance improves quality within that range. Favorable guidance strengths also remain similar across auxiliary scales.

We therefore first identify a favorable guidance range at λ=0, and then evaluate only three guidance strengths across five auxiliary scales. As shown in Supplementary Fig. 3(c), the resulting 5×3 grid, consisting of only 15 configurations, already recovers a strong quality–coverage frontier.

AuthorsOpenReview ↗

Our 50K results are stable across evaluation seeds

We thank the reviewer for requesting this uncertainty analysis. We repeat the evaluation of six representative operating points along our 50K quality–coverage frontier using three independent evaluation seeds. The trained checkpoint and inference settings are fixed, while the initial noise, class labels, and auxiliary codes vary across seeds. Across all six points, the sample standard deviations are at most 0.0137 in FDᵣ,₆ and 0.0051 in Recall (Supplementary Sec. K, Table 10). At a representative high-coverage point, we obtain FDᵣ,₆=5.9033±0.0079 and Recall =0.6784±0.0034.

We also evaluate the best-quality operating points of strong baselines under the same three-seed, 50K protocol:

MethodFDᵣ,₆↓Recall ↑
Ours4.4544±0.01320.6064±0.0021
SiT+AG4.5306±0.00680.6208±0.0012
CAFM4.8562±0.01640.5722±0.0011

All three methods show low evaluation-seed variability. Among these selected best-quality operating points, our method retains the lowest mean FDᵣ,₆. We will add these 50K uncertainty results and clarify the evaluation-seed protocol in the revised paper.

Comments on the paper

7 comments
AuthorsOpenReview ↗

General Response and Additional Supplementary Material

We sincerely thank all reviewers for their thoughtful and constructive feedback. Your comments have helped us clarify the motivation, mechanism, and contributions of our work. In response, we have conducted additional experiments, provided further analyses and explanations, and will incorporate the corresponding clarifications and results into the revised paper.

To keep our responses concise, we state the key findings and conclusions directly in the discussion. Complete experimental details, analyses, additional controls, figures, and tables are provided in a separate PDF Additional Supplementary Material, uploaded under the Supplementary Material field on our submission page. References to “Revised Supplementary Sec. X” in our replies point to this PDF, not to the appendix included after the main paper in our original submission.

AuthorsOpenReview ↗

Response to All Reviewers — A: Conditional mean is correct, but sample-wise targets offer more to learn from.

Overall Premise:

We fully agree that standard FM learns the conditional-mean velocity, which is correct in the ideal continuous-time limit. We thank the reviewer for highlighting this distinction and will revise the paper to make it explicit, avoiding any impression that averaging makes ideal FM incorrect.

Although the conditional mean is exactly what ideal FM requires, it does not explicitly use the variation among sample-wise velocity targets as a learning signal. We ask whether retaining this variation can benefit practical finite models and finite-step sampling.

Our core idea can be summarized as:

  • At a given noisy state, different clean samples can induce different valid velocity targets, creating an ambiguity that standard FM handles with a single conditional-mean response. If an auxiliary label indicated which target alternative is associated with each sample, conditioning on it could disambiguate these targets and allow the resulting label-conditioned fields to specialize toward different alternatives, making each field simpler and more mode-directed while preserving distinct behaviors for population-level coverage.
  • Because such labels are unavailable, we sample auxiliary codes and use the established WTA rule to learn the assignment implicitly: only the code-conditioned prediction closest to each target is updated. This performs unsupervised local routing over the conditional velocity distribution.
  • This local routing induces a family of code-conditioned velocity fields specialized toward different target directions, while also reshaping the shared base field. A fixed code produces a coherent trajectory, and scaling its influence with λ\lambda smoothly controls the expressed variation.
  • Specialization toward valid target directions supports fidelity, while mixing codes improves population coverage. Increasing the number of codes available to a generated population monotonically improves Recall while reducing FID\mathrm{FID}. This code-based coverage control complements guidance by changing how the model responds to it: instead of following the usual fidelity–coverage trade-off, guidance can improve fidelity without the usual loss of coverage, thereby expanding the quality–coverage frontier.

Novelty & Contribution:

The novelty is not the WTA rule itself, but how we use it to organize sample-wise FM target variation into useful generative structure and thereby enable controllable generative behaviors that expand the accessible quality–coverage frontier. Concretely, competitive training learns a family of code-conditioned velocity fields specialized toward different target directions, reshapes and strengthens the shared base field even when auxiliary conditioning is disabled, broadens population coverage by using different fixed codes across generated samples, and improves finite-step generation without KK-way inference.

Overall, we develop an effective post-training method for FM generators and introduce a new coverage-oriented inference-time control that complements guidance. We provide systematic analyses of the behaviors induced by competitive training, including the joint effects of auxiliary conditioning and guidance and their scaling with KK. We discover that auxiliary conditioning changes how models respond to guidance: fidelity can improve without the usual loss of coverage, yielding the strongest quality–coverage frontier among the methods we evaluate.

AuthorsOpenReview ↗

Response to All Reviewers — B: Clarifying the role of the 1D toy

The 1D toy is designed to illustrate the advantage of Competitive FM under practical finite-step sampling. Because this toy is exceptionally simple, we deliberately use only four Euler steps to create a stringent finite-step setting in which differences between the learned fields become visible.

Competitive FM organizes sample-wise target variation into more mode-directed fields, allowing each trajectory to reach a valid mode more directly with few sampling steps. Individual specialized routes support accurate generation, while mixing codes broadens population coverage. Consequently, even with only four steps, Competitive FM reaches both target modes with little off-support mass. It therefore achieves both accurate generation and broad coverage in a setting where standard FM performs poorly.

This toy is not intended to suggest a flaw in standard FM. For such a simple target, standard FM recovers the distribution increasingly accurately as the integration becomes finer, as expected from its correct continuous-time solution. Indeed, increasing the number of Euler steps from 4 to 8, 16, and 32 progressively reduces its off-support mass from 4.089% to 1.953%, 1.270%, and 0.757%. The distinction illustrated by the toy is therefore practical: standard FM progressively catches up with finer integration, whereas Competitive FM already generates accurately and covers both modes under a very limited sampling budget. (We thank the reviewer for highlighting this distinction and will revise the paper to make explicit that the toy does not indicate a flaw in standard FM and that standard FM recovers the target increasingly accurately with finer integration.)

The toy is also used to intuitively illustrate the behaviors induced by Competitive FM, and these behaviors closely mirror our ImageNet results. The auxiliary scale λ controls how strongly the sample-wise target variation captured by the codes is expressed at inference. At λ=0, auxiliary conditioning is disabled, and generation concentrates on the dominant mode, producing mode-seeking behavior. As λ increases, more of the code-conditioned variation is expressed, progressively broadening population coverage across both modes. Specifically, for λ=0→0.25→0.5→0.75→1, the mass on the less-covered mode increases from 4.96%→23.31%→34.58%→39.87%→42.65%, while off-support mass remains below 1%. Thus, a single model can smoothly move between mode-seeking and broader-coverage behaviors at inference.

We observe the same pattern on ImageNet: code specialization supports generation fidelity, mixing codes broadens population coverage, and λ provides continuous inference-time control over the resulting operating points.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.