acceptodds
Under review as a conference paper at ICLR 2027

Osmotic Guidance for Generative Flows

Abstract

This paper studies guidance for flow-based generative models from the perspective of intervention efficiency, an important aspect that is often less explicitly considered in existing guidance research. To address the question of how to achieve the same instantaneous guidance effect with minimal modification to the pretrained dynamics, we formulate guidance as a local constrained optimization problem that minimizes intervention energy subject to a prescribed objective-change rate, yielding a unique canonical minimum-energy solution. Using the marginal equivalence between deterministic and stochastic dynamics, we further show that this solution admits a relative osmotic representation. This leads to osmotic guidance (OG), a training-free framework that provides a unified view of guidance and supports diverse guidance objectives through objective-induced osmotic fields with a shared energy-calibration rule. We instantiate OG for entropy, reward, target-density, and conditional guidance, and introduce a two-anchor preview for efficient strength selection. Experiments on synthetic flow models and image datasets demonstrate effective guidance, improved intervention efficiency, and empirical support for the proposed canonical minimum-energy principle.

Primary area:
generative models
Keywords:
generative flows, flow matching, inference-time guidance, optimal control, distributional control, score-based modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7QYVOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Osmotic Guidance (OG), a training-free guidance method for flow and diffusion models that minimizes intervention energy subject to a prescribed instantaneous rate of change of a desired objective. Whereas standard gradient-based guidance methods scale the objective gradient by a predetermined coefficient λt\lambda_t, OG adaptively normalizes the gradient using its expected squared norm. This normalization ensures that the guidance achieves the prescribed instantaneous objective-improvement rate while minimizing the energy required to do so.

Critical strengths

This paper yields a principled guidance method, whose control guarantees an rate of improvement for the desired objective. The proposed osmotic formulation offers a unified perspective on several guidance objectives, including reward maximization, entropy minimization, target-density matching, and conditional generation. The paper's main proofs and methods are clear and the experiment settings are diverse. The experimental results demonstrate promising improvements in intervention efficiency and sample quality over selected baselines, particularly for entropy and conditional guidance on ImageNet-256.

Critical weaknesses

The paper lacks a stronger comparison to previous guidance methods, and its primary contributions are not well motivated. From my understanding of the paper, the primary contribution is the energy normalized guidance coefficient. However, this coefficient does not yield guarantees to the terminal improvements to the objective. Also, the method does not answer why this specific adaptive time dependent coefficient is preferable over existing adaptive coefficient-selection strategies, including decaying guidance strengths, normalization with respect to the norm of the base flow, or some with respect to some RMS over reward gradient magnitudes. Though the paper offers several experimental settings, the reward-guidance experiments are restricted to low-dimensional synthetic datasets with relatively simple objectives. Ideally, there would also be comparisons with methods such as DPS, FlowDPS, and other established baselines on more complex settings including human preference rewards/inverse problems. The method is also constrained in its estimation of the energy estimation, requiring offline estimation, or additional inference budget.

Questions for authors
  1. Why should controlling the instantaneous rate of objective improvement be preferable to other constrained formulations when the ultimate goal is terminal sample quality?

  2. Does OG's normalization actually outperform alternative adaptive-λt\lambda_t approaches when the guidance direction, computational budget, and terminal objective are held fixed?

  3. Can you demonstrate stronger performance compared to other gradient based guidance methods such as DPS/FlowDPS in inverse problems and human preference reward settings holding the inference time budget the same.

8 comments
AuthorsOpenReview ↗

Rebuttal W1

We thank the reviewer for their careful reading of our paper and for their constructive comments. We appreciate the opportunity to clarify these important points.

Weaknesses

W1. Motivation and Contributions

As for the contributions, while the energy-normalized coefficient is an important consequence of our formulation, it is not our sole contribution. Specifically, we provide:

  • A canonical minimum-energy characterization of guidance, establishing the unique optimal intervention field in L2(ρt)L^2(\rho_t) under a prescribed instantaneous objective-change rate (Proposition 2).
  • An objective-induced relative osmotic representation, connecting the optimal guidance direction to the difference between the osmotic fields of the current marginal and its objective-tilted counterpart (Eqs. 14–18).
  • A unified guidance formulation encompassing entropy concentration, reward maximization, target-density matching, and conditional generation (Table 1).
  • A two-anchor preview strategy for reducing the computational cost of guidance-strength selection (Section 3.3).

Thus, the normalization coefficient is the practical realization of a broader distributional-control formulation, rather than an independently proposed heuristic. In other words, it is an implementation-level contribution that follows from the theory we establish.

Furthermore, as the reviewer already summarized, our primary motivation is intervention efficiency while maintaining comparable quality; we do not aim to improve terminal quality.

We will revise the Introduction to further emphasize our primary motivation and distinguish it from directly optimizing terminal generation quality.

AuthorsOpenReview ↗

Rebuttal W2.A

W2. Comparisons with Existing Guidance Methods and Broader Experimental Validation

We have expanded our experiments across five pretrained generative model families—SiT-XL/2, LightningDiT-XL, Rectified Flow, Lumina-Image-2.0, and Stable Diffusion 3.5 Medium—using ImageNet-256, CelebA-HQ-256, and COCO text-to-image prompts.

The additional experiments cover six settings: entropy guidance, conditional guidance, reward guidance, target-density guidance, inverse problems, and text-to-image generation, including human-preference evaluation. Our primary evaluation criterion is **intervention energy at comparable achieved guidance effects while maintaining comparable quality.

1. Entropy Guidance

1.1 Entropy guidance on SiT-XL/2 (ImageNet-256). We compare OG against Constant Guidance, Temporal Score Rescaling (TSR), and Density Guidance at matched terminal latent-variance contraction. OG uses offline reference-Fisher energy calibration. The TSR results use the original data-scale setting σ=1\sigma=1.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
8.8%OG8.790%7.8765.02
Constant8.790%8.24115.23
TSR8.777%8.20911.89
Density Guidance8.681%8.348190.0
11.7%OG11.746%8.1869.49
Constant11.775%9.52229.68
TSR11.724%9.58423.42
Density Guidance11.666%8.310353.2

1.2 Entropy guidance on LightningDiT-XL (ImageNet-256). We evaluate OG, TSR, and Density Guidance at three target levels of terminal latent-variance contraction.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG5.02%8.985100
TSR4.98%12.977209
Density Guidance5.09%8.0825,958
9%OG8.90%12.706353
TSR8.98%24.255952
Density Guidance9.25%9.51118,884
12%OG11.98%17.728713
TSR11.94%29.9722,393
Density Guidance12.24%10.98332,273

1.3 Entropy guidance on Rectified Flow (CelebA-HQ-256). We compare OG, TSR, and Density Guidance at approximately matched DINOv2 feature contraction. Feature contraction is used because terminal pixel-variance contraction does not consistently reflect the achieved concentration in this model.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG4.95%34.73182
TSR4.99%34.61182
Density Guidance5.16%34.5111,900
10%OG9.95%39.94275
TSR9.96%40.00284
Density Guidance10.03%39.5416,600

2. Conditional Guidance

2.1 Class-conditional guidance on SiT-XL/2 (ImageNet-256). We compare Limited-Interval CFG (LI-CFG) against temporally gated OG as Reviewer xbkq suggests, using the same guidance interval t∈[0.3,0.7]t\in[0.3,0.7] and reference-velocity convention to match the benchmark settings. Both methods are evaluated at statistically equivalent continuous terminal conditional alignment, Acond=E(x,y)[log⁡pclf(y∣x)]A_{\mathrm{cond}}=\mathbb{E}_{(x,y)}[\log p_{\mathrm{clf}}(y\mid x)]. Gated OG retains canonical normalization within the specified interval.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (gated)−0.9439.150.720597.4
LI-CFG−0.9449.240.711589.8

LI-CFG restricts when guidance is applied, whereas OG adaptively normalizes guidance strength within the selected interval. At equivalent conditional alignment, gated OG achieves numerically better FID and recall, with comparable but 1.3% higher intervention energy. These results demonstrate that temporal restriction and OG's adaptive normalization are compatible.

2.2 Class-conditional guidance on LightningDiT-XL (ImageNet-256). We compare a tempered OG variant with SMC-CFG at approximately matched continuous conditional alignment. The tempered coefficient is βt∝(mref/mt)1/4\beta_t\propto(m_{\mathrm{ref}}/m_t)^{1/4}.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (tempered)−0.87829.1610.744150
SMC-CFG−0.87918.6380.7595,146

2.3 Attribute-conditional guidance on Rectified Flow (CelebA-HQ-256). We compare OG against Classifier Guidance and Universal Guidance for facial smiling-attribute control.

MethodSmiling rate (matched)FID ↓EpathE_{\mathrm{path}} ↓
OG67.06%17.9860.1
Classifier Guidance66.86%17.69917
Universal Guidance67.30%25.761,245
AuthorsOpenReview ↗

Rebuttal W2.B

3. Reward Guidance

3.1 Reward guidance on SiT-XL/2 (ImageNet-256). We compare OG-FM against Flow-Map Reward Guidance (FMRG) at approximately matched aesthetic rewards. OG-FM applies OG's energy normalization to a flow-map reward gradient, using the same gradient construction and model-evaluation budget as FMRG.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
~7.0OG-FM6.97222.0925.60
FMRG6.98328.5031.78
~8.5OG-FM8.47643.09178.8
FMRG8.53254.44264.8

3.2 Reward guidance on LightningDiT-XL (ImageNet-256). We compare OG against Universal Guidance using a pretrained aesthetic reward model at approximately matched moderate and strong rewards.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG5.10729.7342.5
Universal Guidance5.07529.6019,000
StrongOG5.52829.81569
Universal Guidance5.50229.7478,800

3.3 Reward guidance on Rectified Flow (CelebA-HQ-256). We evaluate OG and Universal Guidance for aesthetic reward maximization at approximately matched achieved rewards.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG6.58318.44339
Universal Guidance6.57622.956,780
StrongOG7.92726.625,030
Universal Guidance7.92631.7716,500

4. Target-Density Guidance

4.1 Target-density guidance on SiT-XL/2 (ImageNet-256). We compare OG against Monte Carlo guidance gMCg^{\mathrm{MC}}, measuring the Wasserstein discrepancy W2W_2 between generated samples and the target distribution. Each method is evaluated at moderate and strong target shifts.

TargetMethodW2W_2 (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG0.31318.531.37
gMCg^{\mathrm{MC}}0.30619.8115.7
StrongOG0.22518.295.60
gMCg^{\mathrm{MC}}0.22920.0535.4

4.2 Target-density guidance on LightningDiT-XL (ImageNet-256). We compare OG against Gradient Guidance and gMCg^{\mathrm{MC}} at approximately matched moderate target-density discrepancy W2W_2.

MethodW2W_2 (matched)FID ↓EpathE_{\mathrm{path}} ↓
OG0.2287.140.26
Gradient Guidance0.2277.150.78
gMCg^{\mathrm{MC}}0.2307.2522.3

5. Inverse Problems

5.1. Image inpainting and deblurring on Rectified Flow (CelebA-HQ-256). We compare OG against DPS and deterministic FlowDPS ( η=0\eta=0 , the “w/o ϵ\epsilon” ablation), calibrating methods to statistically equivalent measurement-consistency errors within a prespecified 1% equivalence margin. All methods use the same inference configuration: Euler-25, guidance stride 2, batch size 8, and the same device. Reported runtimes are synchronized, generation-only GPU wall-clock times.

TaskMethodMeasurement error (matched)EpathE_{\mathrm{path}} ↓Runtime (s/img) ↓
Inpainting vs DPSOG0.5172580.483
DPS0.5162590.482
Inpainting vs FlowDPSOG0.5063740.483
FlowDPS (η=0\eta=0)0.5048060.485
Deblurring vs DPSOG0.4027820.497
DPS0.4006730.496
Deblurring vs FlowDPSOG0.4027820.497
FlowDPS (η=0\eta=0)0.4001,3640.524

At approximately matched measurement consistency and comparable inference time, OG reduces intervention energy by 53.6% and 42.7% relative to deterministic FlowDPS for inpainting and deblurring, respectively. Against DPS, OG achieves nearly identical energy for inpainting but requires 16.2% more energy for deblurring. Importantly, these differences are observed under the same sampling budget, with OG and DPS exhibiting nearly identical runtime (within 0.2%) and OG running approximately 5.2% faster than deterministic FlowDPS for deblurring.

Reviewer_Z8UXOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to build a theory framework at inference stage for pre-trained flow model based on minimizing the intervention required to achieve a prescribed guidance effect and accordingly proposes Osmotic Guidance (OG), which can achieve multiple guidance tasks under one architecture.

Critical strengths

Firstly, the paper tries to achieve a prescribed instantaneous objective improvement while minimizing perturbations. And they starts with a quite simple and mathematically straightforward theory, which provides a useful perspective for analyzing and comparing inference-time interventions. Secondly, the proposed framework accommodates entropy, reward, target-density, and conditional objectives under a common formulation. This makes the approach conceptually accessible and potentially applicable to different inference-time control problems.

Critical weaknesses

Firstly, I think this paper's theoretical optimality is local and instantaneous, which means, it doesn't provide an improved end-to-end sampling quality prove. Consequently, minimizing intervention energy locally does not necessarily yield better final samples or outperform existing guidance methods across diverse tasks. Secondly, the empirical validation is limited. The experiments cover a relatively narrow set of guidance objectives and benchmarks. Besides, the results are mixed: for example, OG improves FID and recall over CFG at matched Top-1 accuracy but sacrifices precision. This makes it difficult to establish that the proposed framework offers a broadly reliable advantage over existing methods.

Questions for authors
  1. How sensitive are the empirical gains to the accuracy of the estimated score fields and energy calibration? Do the authors observe similar improvements when these quantities can not be estimated perfectly?

  2. Can you justify the broader effectiveness of the proposed framework given its local, instantaneous optimality guarantee and mixed empirical results?

7 comments
AuthorsOpenReview ↗

Rebuttal W1

We thank the reviewer for their careful reading of our paper and for their constructive comments. We appreciate the opportunity to clarify these important points.

Weaknesses

W1. Local and Instantaneous Optimality

We agree that instantaneous minimum-energy optimality does not directly guarantee improved terminal sample quality or universal superiority over existing guidance methods. We would like to clarify that our primary objective is to minimize intervention energy while maintaining comparable end-to-end sampling quality, rather than to improve it. We will revise the Introduction to further emphasize our primary motivation and distinguish it from directly optimizing terminal generation quality.

In addition, we adopt local, instantaneous optimization because it is more tractable than a globally coupled formulation. When the marginal path and instantaneous rates are fixed, it also minimizes accumulated intervention energy over the full time horizon.

AuthorsOpenReview ↗

Rebuttal W2.A

W2. Broader Empirical Validation and Downstream Generation Quality

We have expanded our experiments across five pretrained generative model families—SiT-XL/2, LightningDiT-XL, Rectified Flow, Lumina-Image-2.0, and Stable Diffusion 3.5 Medium—using ImageNet-256, CelebA-HQ-256, and COCO text-to-image prompts.

The additional experiments cover six settings: entropy guidance, conditional guidance, reward guidance, target-density guidance, inverse problems, and text-to-image generation, including human-preference evaluation. Our primary evaluation criterion is intervention energy at comparable achieved guidance effects while maintaining comparable quality.

1. Entropy Guidance

1.1 Entropy guidance on SiT-XL/2 (ImageNet-256). We compare OG against Constant Guidance, Temporal Score Rescaling (TSR), and Density Guidance at matched terminal latent-variance contraction. OG uses offline reference-Fisher energy calibration. The TSR results use the original data-scale setting σ=1\sigma=1.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
8.8%OG8.790%7.8765.02
Constant8.790%8.24115.23
TSR8.777%8.20911.89
Density Guidance8.681%8.348190.0
11.7%OG11.746%8.1869.49
Constant11.775%9.52229.68
TSR11.724%9.58423.42
Density Guidance11.666%8.310353.2

1.2 Entropy guidance on LightningDiT-XL (ImageNet-256). We evaluate OG, TSR, and Density Guidance at three target levels of terminal latent-variance contraction.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG5.02%8.985100
TSR4.98%12.977209
Density Guidance5.09%8.0825,958
9%OG8.90%12.706353
TSR8.98%24.255952
Density Guidance9.25%9.51118,884
12%OG11.98%17.728713
TSR11.94%29.9722,393
Density Guidance12.24%10.98332,273

1.3 Entropy guidance on Rectified Flow (CelebA-HQ-256). We compare OG, TSR, and Density Guidance at approximately matched DINOv2 feature contraction. Feature contraction is used because terminal pixel-variance contraction does not consistently reflect the achieved concentration in this model.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG4.95%34.73182
TSR4.99%34.61182
Density Guidance5.16%34.5111,900
10%OG9.95%39.94275
TSR9.96%40.00284
Density Guidance10.03%39.5416,600

2. Conditional Guidance

2.1 Class-conditional guidance on SiT-XL/2 (ImageNet-256). We compare Limited-Interval CFG (LI-CFG) against temporally gated OG as Reviewer xbkq suggests, using the same guidance interval t∈[0.3,0.7]t\in[0.3,0.7] and reference-velocity convention to match the benchmark settings. Both methods are evaluated at statistically equivalent continuous terminal conditional alignment AcondA_{\mathrm{cond}}, defined as the expected classifier log-likelihood E(x,y)log⁡pclf(y∣x)\mathbb{E}_{(x,y)}\log p_{\mathrm{clf}}(y|x). Gated OG retains canonical normalization within the specified interval.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (gated)−0.9439.150.720597.4
LI-CFG−0.9449.240.711589.8

LI-CFG restricts when guidance is applied, whereas OG adaptively normalizes guidance strength within the selected interval. At equivalent conditional alignment, gated OG achieves numerically better FID and recall, with comparable but 1.3% higher intervention energy. These results demonstrate that temporal restriction and OG's adaptive normalization are compatible.

2.2 Class-conditional guidance on LightningDiT-XL (ImageNet-256). We compare a tempered OG variant with SMC-CFG at approximately matched continuous conditional alignment. The tempered coefficient is βt∝(mref/mt)1/4\beta_t\propto(m_{\mathrm{ref}}/m_t)^{1/4}.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (tempered)−0.87829.1610.744150
SMC-CFG−0.87918.6380.7595,146

2.3 Attribute-conditional guidance on Rectified Flow (CelebA-HQ-256). We compare OG against Classifier Guidance and Universal Guidance for facial smiling-attribute control.

MethodSmiling rate (matched)FID ↓EpathE_{\mathrm{path}} ↓
OG67.06%17.9860.1
Classifier Guidance66.86%17.69917
Universal Guidance67.30%25.761,245
AuthorsOpenReview ↗

Rebuttal W2.B

3. Reward Guidance

3.1 Reward guidance on SiT-XL/2 (ImageNet-256). We compare OG-FM against Flow-Map Reward Guidance (FMRG) at approximately matched aesthetic rewards. OG-FM applies OG's energy normalization to a flow-map reward gradient, using the same gradient construction and model-evaluation budget as FMRG.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
~7.0OG-FM6.97222.0925.60
FMRG6.98328.5031.78
~8.5OG-FM8.47643.09178.8
FMRG8.53254.44264.8

3.2 Reward guidance on LightningDiT-XL (ImageNet-256). We compare OG against Universal Guidance using a pretrained aesthetic reward model at approximately matched moderate and strong rewards.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG5.10729.7342.5
Universal Guidance5.07529.6019,000
StrongOG5.52829.81569
Universal Guidance5.50229.7478,800

3.3 Reward guidance on Rectified Flow (CelebA-HQ-256). We evaluate OG and Universal Guidance for aesthetic reward maximization at approximately matched achieved rewards.

TargetMethodReward (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG6.58318.44339
Universal Guidance6.57622.956,780
StrongOG7.92726.625,030
Universal Guidance7.92631.7716,500

4. Target-Density Guidance

4.1 Target-density guidance on SiT-XL/2 (ImageNet-256). We compare OG against Monte Carlo guidance gMCg^{\mathrm{MC}}, measuring the Wasserstein discrepancy W2W_2 between generated samples and the target distribution. Each method is evaluated at moderate and strong target shifts.

TargetMethodW2W_2 (matched)FID ↓EpathE_{\mathrm{path}} ↓
ModerateOG0.31318.531.37
gMCg^{\mathrm{MC}}0.30619.8115.7
StrongOG0.22518.295.60
gMCg^{\mathrm{MC}}0.22920.0535.4

4.2 Target-density guidance on LightningDiT-XL (ImageNet-256). We compare OG against Gradient Guidance and gMCg^{\mathrm{MC}} at approximately matched moderate target-density discrepancy W2W_2.

MethodW2W_2 (matched)FID ↓EpathE_{\mathrm{path}} ↓
OG0.2287.140.26
Gradient Guidance0.2277.150.78
gMCg^{\mathrm{MC}}0.2307.2522.3

5. Inverse Problems

5.1 Image inpainting and deblurring on Rectified Flow (CelebA-HQ-256). We compare OG against DPS and deterministic FlowDPS (η=0\eta=0, corresponding to its “w/o ϵ\epsilon” ablation). Methods are calibrated to statistically equivalent measurement-consistency errors within a prespecified 1% equivalence margin.

TaskMethodMeasurement error (matched)EpathE_{\mathrm{path}} ↓
Inpainting vs DPSOG0.517258
DPS0.516259
Inpainting vs FlowDPSOG0.506374
FlowDPS (η=0\eta=0)0.504806
Deblurring vs DPSOG0.402782
DPS0.400673
Deblurring vs FlowDPSOG0.402782
FlowDPS (η=0\eta=0)0.4001,364

OG achieves essentially identical energy to DPS for inpainting but requires approximately 16.2% more energy for deblurring. In contrast, OG reduces energy by 53.6% and 42.7% relative to deterministic FlowDPS for inpainting and deblurring, respectively. The FlowDPS comparison concerns its deterministic variant, not the default stochastic method.

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

This paper studies guidance for flow-based generative models from the perspective of intervention efficiency, rather than just sample quality or condition alignment.

The core question is: for a desired guidance effect, can we achieve it with minimal perturbation to the pretrained dynamics? The authors argue that strong guidance often pushes trajectories far from the original generative flow, hurting unrelated image properties.

To address this, they formulate guidance as a local constrained optimization problem at each time step. The goal is to minimize instantaneous intervention energy, measured as the expected squared norm of the added guidance field, while enforcing a prescribed rate of change for a distributional objective. The solution to this problem is shown to be unique in the square-integrable space and proportional to the spatial gradient of the objective's first variation. This is presented as the canonical minimum-energy guidance field.

The paper then connects this solution to the marginal equivalence between deterministic ODE flows and stochastic SDE flows. Using this equivalence, they define an osmotic field as a scaled score function, and a relative osmotic field as the difference between two such scores. The canonical guidance direction can then be reinterpreted as an objective-induced relative osmotic field, which is the difference between the osmotic field of a tilted version of the current marginal that favors the objective and the osmotic field of the current marginal itself. This leads to the Osmotic Guidance (OG) framework, which is training-free. The framework has two parts: a direction determined by the objective and a scalar strength determined by energy calibration. The energy calibration is estimated empirically from a batch of samples.

The authors instantiate OG for four common objectives: entropy concentration, reward maximization, target-density guidance, and conditional guidance. In each case the direction reduces to a known score difference, for example conditional guidance becomes the difference between conditional and unconditional scores, similar to classifier-free guidance.

Finally, they propose a two-anchor preview method to efficiently select guidance strength. Instead of running full generation for many strengths, they run only two anchor strengths and linearly interpolate the terminal latents to approximate intermediate strengths, using this preview to screen candidates before a final full solve.

Experiments are done on synthetic 2D distributions (two moons, checkerboard, Gaussian mixture) with a pretrained flow matching model, and on ImageNet-256 with a pretrained SiT-XL/2 model.

Critical strengths
  1. Linking intervention energy to terminal deviation from the unguided distribution is a reasonable motivation.

  2. Writing all guidance as relative score difference is pedagogically clean. The summary table correctly unifies entropy, reward, target, conditional under same conceptual formula.

  3. No backbone retraining, only inference-time field plus scalar normalization.

Critical weaknesses
  1. The canonical minimum-energy result is a quadratic program solved by Cauchy-Schwarz - identical to prior control barrier function QP work. Defining osmotic field as scaled score is just renaming the score. Relative osmotic field is score difference, which is exactly how classifier-free guidance and many flow guidance methods already operate. No new algorithmic primitive.

  2. Theory-practice mismatch (admitted in appendix)

  3. Proposition 1 exponential in integral of Lipschitz constant - vacuous for neural velocity fields. Defining method to minimize energy then reporting energy as efficiency metric is circular. Need external Pareto frontiers with confidence intervals.

  4. Weak, noisy, inconsistent empirical gains

  • 2D: Reward per energy gains are within one standard deviation across methods (e.g., moons 17.18 vs 17.16 vs 17.01). Target energy to achieve same W2 reduction has large overlapping variance.

  • ImageNet: FID improvement at matched contraction is small (7.876 vs 8.209). At matched Top-1 accuracy around 0.95, OG uses about twice the energy of CFG (749 vs 339 in left-Riemann logs) for slightly better FID, contradicting efficiency claim. Precision drops (0.857 vs 0.889) at matched Top-1, trading precision for recall.

  • Sample sizes: 2048 for 2D, 128 for entropy screen, 1000 screen vs 5000 confirmation for conditional - far from 50k standard. Historical 50k benchmark shows TSR more contracted than OG. Unfinished 50k extension has empty metrics directory.

  • Baselines: No limited-interval guidance, no autoguidance, no self-guidance. gMC baseline uses small 512-bank plus Gaussian VP path - weak.

  1. Bound is standard linear interpolation error for twice-differentiable functions. Speedup 1.64x excludes one-time calibration cost of over 2120 seconds (55s schedule + 2064s pilots). Break-even around 9 sessions. Cold-start is slower. Width sweep does not show quadratic scaling.
Questions for authors
  1. Given appendix discussion, what functional does your 2D and ImageNet entropy implementation actually optimize? Can you estimate true entropy via kNN or report expected log of reference density vs true entropy separately?

  2. Quantify transfer error between calibration and evaluation norms on ImageNet when using frozen 256-sample calibration. Does online per-batch normalization cause mode collapse? Why floor vs additive stabilizer?

  3. Why project onto 3 of 4 latent channels? What is impact of remaining conditional velocity on logged energy? Since CFG with scale 1 has non-zero logged energy under your definition, how is Top-1 matching fair?

  4. Energy logs show OG uses about twice the energy of CFG at similar FID and Top-1. In what sense is OG more efficient? Provide paired statistical test for energy saving with fixed initial noise and same quadrature.

  5. You report mean log target density but claim negative KL improvement. Provide KL estimate or at least expected log density of guided marginal to separate terms.

  6. Include one-time calibration in speedup calculation. What is wall-clock for cold-start single session? Why linear latent interpolation vs spherical or flow-aware interpolation?

  7. Why not compare to limited-interval CFG which also aims to reduce unnecessary intervention? How does reference-Fisher strength compare to TSR with tuned data-scale parameter?

11 comments
AuthorsOpenReview ↗

Rebuttal W1-2

We thank the reviewer for their careful reading of our paper and for their constructive comments. We appreciate the opportunity to clarify these important points.

Weaknesses

W1. Novelty and Contributions

Our contribution is not the invention of these individual ingredients, but their formulation and integration into a minimum-intervention framework for generative guidance.

Specifically, we formulate guidance as an optimization problem in L2(ρt)L^2(\rho_t) under a prescribed instantaneous distributional-objective rate (Proposition 2), establish its objective-induced relative osmotic representation (Eqs. 14–18), and derive a unified guidance formulation for four objectives (Table 1). This yields a principled, objective-dependent normalization rule rather than a manually specified guidance coefficient.

We acknowledge that the instantaneous quadratic program admits a Cauchy–Schwarz / minimum-norm solution, and that the resulting guidance direction can coincide with existing methods such as CFG. Our contribution is not a new QP solver or a renamed score field. Rather, OG provides a single rate-calibrated mechanism that turns four guidance objectives into comparable interventions—something that “CFG-style direction + manually tuned scale” does not specify. The practical payoff of this energy-calibrated strength selection is further evaluated against alternative adaptive normalization rules (see W4, Adaptive normalization ablation). We will clarify these contributions and their relationship to prior work in the revision.

W2. Theory–Practice Consistency

Although practical OG implementations employ approximations to the theoretical formulation, our additional experiments demonstrate that these approximations preserve the intended entropy-improvement behavior across the tested settings.

As discussed in Appendix B.7, entropy guidance using the pretrained reference score corresponds to the surrogate ΔJproxy/ΔJtrue≈0.3–0.6\Delta J_{\mathrm{proxy}}/\Delta J_{\mathrm{true}}\approx 0.3\textrm{--}0.6 , rather than the true negative entropy Jtrue=−H(ρt)J_{\mathrm{true}}=-H(\rho_t). Their relationship is

Jproxy=Jtrue−DKL(ρt∥πt).J_{\mathrm{proxy}}=J_{\mathrm{true}}-D_{\mathrm{KL}}(\rho_t\|\pi_t).

To quantify the practical impact of this approximation, we conducted an additional 2D diagnostic experiment on Two Moons, Checkerboard, and Gaussian Mixture, using two guidance strengths and three independent seeds.

At approximately 5% and 9% terminal contraction, the surrogate and true negative entropy improve in the same direction across all 30 evaluated OG cases, with a strong within-dataset correlation (r≈0.99r\approx0.99) between contraction and true entropy improvement. Although the surrogate underestimates the magnitude of improvement (ΔJproxy/ΔJtrue≈0.3\Delta J_{\mathrm{proxy}}/\Delta J_{\mathrm{true}}\approx0.3–0.60.6), no cases of surrogate improvement accompanied by deterioration in true negative entropy were observed.

We also evaluate frozen normalization transfer. The ratio between the guided-marginal squared-gradient norm and its frozen calibration estimate has a median of 0.90 (range 0.64–1.28), indicating moderate calibration discrepancies rather than exact rate preservation.

Overall, these results suggest that the practical approximations retain the intended entropy-improvement behavior in the evaluated regimes, despite quantitative deviations from the ideal formulation. We will include these results in the revised manuscript and clarify the distinction between exact theoretical objectives and their practical approximations.

AuthorsOpenReview ↗

Rebuttal W3

W3. Stability Bound and Intervention-Efficiency Evaluation

Regarding Proposition 1, we agree that the exponential dependence on the integrated Lipschitz constant may yield a loose bound for neural velocity fields. Our intention is not to provide a tight numerical certificate of terminal deviation, but to establish a conditional relationship between intervention energy and terminal distributional deviation. Importantly, our minimum-energy characterization in Proposition 2 is independent of this stability bound. We will clarify their distinct roles and limitations.

Regarding the intervention-energy metric, we agree that lower energy alone is insufficient to demonstrate practical benefits. However, Proposition 2 establishes instantaneous optimality at a fixed marginal, whereas the reported accumulated energy is evaluated along the resulting guided trajectories and is not guaranteed to be globally minimized.

More importantly, we evaluate energy alongside independent downstream metrics at approximately matched guidance effects. For example, on Rectified Flow aesthetic guidance at matched reward (~6.58), OG achieves lower energy (339 vs. 6,780), smaller paired LPIPS (0.0022 vs. 0.0571), and smaller DINO distance (0.0170 vs. 0.0819) compared with Universal Guidance, while improving FID (18.44 vs. 22.95).

These results provide empirical evidence that intervention efficiency can be accompanied by stronger preservation of pretrained outputs, beyond the energy metric itself. We acknowledge that this relationship is not universally guaranteed. As a step toward the reviewer's suggestion of Pareto-style analysis with statistical uncertainty, our expanded matched-effect comparisons report energy alongside independent quality metrics, and Q4 provides bootstrap confidence intervals and paired permutation tests for the reported energy savings. We will further develop explicit Pareto frontiers in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal W4.A

W4. Empirical Reliability and Broader Comparisons

We have substantially expanded our experiments across five pretrained generative model families—SiT-XL/2, LightningDiT-XL, Rectified Flow, Lumina-Image-2.0, and Stable Diffusion 3.5 Medium—using ImageNet-256, CelebA-HQ-256, and COCO text-to-image prompts.

The additional experiments cover six settings: entropy guidance, conditional guidance, reward guidance, target-density guidance, inverse problems, and text-to-image generation, including human-preference evaluation. Our primary evaluation criterion is intervention energy at comparable achieved guidance effects, with generation-quality metrics reported alongside it.

1. Entropy Guidance

1.1 Entropy guidance on SiT-XL/2 (ImageNet-256). We compare OG against Constant Guidance, Temporal Score Rescaling (TSR), and Density Guidance at matched terminal latent-variance contraction. OG uses offline reference-Fisher energy calibration. The TSR results use the original data-scale setting σ=1\sigma=1.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
8.8%OG8.790%7.8765.02
Constant8.790%8.24115.23
TSR8.777%8.20911.89
Density Guidance8.681%8.348190.0
11.7%OG11.746%8.1869.49
Constant11.775%9.52229.68
TSR11.724%9.58423.42
Density Guidance11.666%8.310353.2

1.2 Entropy guidance on LightningDiT-XL (ImageNet-256). We evaluate OG, TSR, and Density Guidance at three target levels of terminal latent-variance contraction.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG5.02%8.985100
TSR4.98%12.977209
Density Guidance5.09%8.0825,958
9%OG8.90%12.706353
TSR8.98%24.255952
Density Guidance9.25%9.51118,884
12%OG11.98%17.728713
TSR11.94%29.9722,393
Density Guidance12.24%10.98332,273

1.3 Entropy guidance on Rectified Flow (CelebA-HQ-256). We compare OG, TSR, and Density Guidance at approximately matched DINOv2 feature contraction. Feature contraction is used because terminal pixel-variance contraction does not consistently reflect the achieved concentration in this model.

TargetMethodContraction (matched)FID ↓EpathE_{\mathrm{path}} ↓
5%OG4.95%34.73182
TSR4.99%34.61182
Density Guidance5.16%34.5111,900
10%OG9.95%39.94275
TSR9.96%40.00284
Density Guidance10.03%39.5416,600

2. Conditional Guidance

2.1 Class-conditional guidance on SiT-XL/2 (ImageNet-256). We compare Limited-Interval CFG (LI-CFG) against temporally gated OG as the reviewer suggests, using the same guidance interval t∈[0.3,0.7]t\in[0.3,0.7] and reference-velocity convention to match the benchmark settings. Both methods are evaluated at statistically equivalent continuous terminal conditional alignment, Acond=E(x,y)[log⁡pclf(y∣x)]A_{\mathrm{cond}}=\mathbb{E}_{(x,y)}[\log p_{\mathrm{clf}}(y\mid x)]. Gated OG retains canonical normalization within the specified interval.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (gated)−0.9439.150.720597.4
LI-CFG−0.9449.240.711589.8

LI-CFG restricts when guidance is applied, whereas OG adaptively normalizes guidance strength within the selected interval. At equivalent conditional alignment, gated OG achieves numerically better FID and recall, with comparable but 1.3% higher intervention energy. These results demonstrate that temporal restriction and OG's adaptive normalization are compatible.

2.2 Class-conditional guidance on LightningDiT-XL (ImageNet-256). We compare a tempered OG variant with SMC-CFG at approximately matched continuous conditional alignment. The tempered coefficient is βt∝(mref/mt)1/4\beta_t\propto(m_{\mathrm{ref}}/m_t)^{1/4}.

MethodAcondA_{\mathrm{cond}} (matched)FID ↓Recall ↑EpathE_{\mathrm{path}} ↓
OG (tempered)−0.87829.1610.744150
SMC-CFG−0.87918.6380.7595,146

2.3 Attribute-conditional guidance on Rectified Flow (CelebA-HQ-256). We compare OG against Classifier Guidance and Universal Guidance for facial smiling-attribute control.

MethodSmiling rate (matched)FID ↓EpathE_{\mathrm{path}} ↓
OG67.06%17.9860.1
Classifier Guidance66.86%17.69917
Universal Guidance67.30%25.761,245

Comments on the paper

3 comments
AuthorsOpenReview ↗

Manuscript Revision Update

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We have carefully revised the manuscript to address the concerns raised during the review process. The main revisions are summarized below:

  1. We revised the second paragraph of the Introduction to emphasize that our primary goal is to minimize intervention energy for a prescribed guidance effect, rather than directly optimize terminal generation quality.
  2. We incorporated additional experiments, baseline comparisons, and validation studies in the Experiments section to address the reviewers' suggestions and provide stronger empirical evidence.
  3. We revised the Conclusion corresponding to the expanded experimental findings.

We kindly invite the reviewers to examine the updated manuscript and would greatly appreciate any further feedback.

Area_Chair_4vmgOpenReview ↗

Reviewer-author discussion kickoff

Dear Reviewers,

Thank you for your review. The authors have now provided their rebuttal. We would appreciate it if you could take some time to read their responses and engage in the discussion, particularly on whether your concerns have been adequately addressed.

Thank you for your time and contributions to the review process!

AuthorsOpenReview ↗

Manuscript Revision Update version 2

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We have carefully revised the manuscript to address the concerns raised during the review process. The main revisions are summarized below:

  1. All revised content is highlighted in blue.

  2. we added a paragraph to the Related Work section clarifying the distinctions between OG, standard optimal-control frameworks and DPS series.

We kindly invite the reviewers to examine the updated manuscript and would greatly appreciate any further feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.