- 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 , 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
-
Why should controlling the instantaneous rate of objective improvement be preferable to other constrained formulations when the ultimate goal is terminal sample quality?
-
Does OG's normalization actually outperform alternative adaptive- approaches when the guidance direction, computational budget, and terminal objective are held fixed?
-
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
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 .
| Target | Method | Contraction (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| 8.8% | OG | 8.790% | 7.876 | 5.02 |
| Constant | 8.790% | 8.241 | 15.23 | |
| TSR | 8.777% | 8.209 | 11.89 | |
| Density Guidance | 8.681% | 8.348 | 190.0 | |
| 11.7% | OG | 11.746% | 8.186 | 9.49 |
| Constant | 11.775% | 9.522 | 29.68 | |
| TSR | 11.724% | 9.584 | 23.42 | |
| Density Guidance | 11.666% | 8.310 | 353.2 |
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.
| Target | Method | Contraction (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| 5% | OG | 5.02% | 8.985 | 100 |
| TSR | 4.98% | 12.977 | 209 | |
| Density Guidance | 5.09% | 8.082 | 5,958 | |
| 9% | OG | 8.90% | 12.706 | 353 |
| TSR | 8.98% | 24.255 | 952 | |
| Density Guidance | 9.25% | 9.511 | 18,884 | |
| 12% | OG | 11.98% | 17.728 | 713 |
| TSR | 11.94% | 29.972 | 2,393 | |
| Density Guidance | 12.24% | 10.983 | 32,273 |
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.
| Target | Method | Contraction (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| 5% | OG | 4.95% | 34.73 | 182 |
| TSR | 4.99% | 34.61 | 182 | |
| Density Guidance | 5.16% | 34.51 | 11,900 | |
| 10% | OG | 9.95% | 39.94 | 275 |
| TSR | 9.96% | 40.00 | 284 | |
| Density Guidance | 10.03% | 39.54 | 16,600 |
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 and reference-velocity convention to match the benchmark settings. Both methods are evaluated at statistically equivalent continuous terminal conditional alignment, . Gated OG retains canonical normalization within the specified interval.
| Method | (matched) | FID ↓ | Recall ↑ | ↓ |
|---|---|---|---|---|
| OG (gated) | −0.943 | 9.15 | 0.720 | 597.4 |
| LI-CFG | −0.944 | 9.24 | 0.711 | 589.8 |
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 .
| Method | (matched) | FID ↓ | Recall ↑ | ↓ |
|---|---|---|---|---|
| OG (tempered) | −0.8782 | 9.161 | 0.744 | 150 |
| SMC-CFG | −0.8791 | 8.638 | 0.759 | 5,146 |
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.
| Method | Smiling rate (matched) | FID ↓ | ↓ |
|---|---|---|---|
| OG | 67.06% | 17.98 | 60.1 |
| Classifier Guidance | 66.86% | 17.69 | 917 |
| Universal Guidance | 67.30% | 25.76 | 1,245 |
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.
| Target | Method | Reward (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| ~7.0 | OG-FM | 6.972 | 22.09 | 25.60 |
| FMRG | 6.983 | 28.50 | 31.78 | |
| ~8.5 | OG-FM | 8.476 | 43.09 | 178.8 |
| FMRG | 8.532 | 54.44 | 264.8 |
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.
| Target | Method | Reward (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| Moderate | OG | 5.107 | 29.73 | 42.5 |
| Universal Guidance | 5.075 | 29.60 | 19,000 | |
| Strong | OG | 5.528 | 29.81 | 569 |
| Universal Guidance | 5.502 | 29.74 | 78,800 |
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.
| Target | Method | Reward (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| Moderate | OG | 6.583 | 18.44 | 339 |
| Universal Guidance | 6.576 | 22.95 | 6,780 | |
| Strong | OG | 7.927 | 26.62 | 5,030 |
| Universal Guidance | 7.926 | 31.77 | 16,500 |
4. Target-Density Guidance
4.1 Target-density guidance on SiT-XL/2 (ImageNet-256). We compare OG against Monte Carlo guidance , measuring the Wasserstein discrepancy between generated samples and the target distribution. Each method is evaluated at moderate and strong target shifts.
| Target | Method | (matched) | FID ↓ | ↓ |
|---|---|---|---|---|
| Moderate | OG | 0.313 | 18.53 | 1.37 |
| 0.306 | 19.81 | 15.7 | ||
| Strong | OG | 0.225 | 18.29 | 5.60 |
| 0.229 | 20.05 | 35.4 |
4.2 Target-density guidance on LightningDiT-XL (ImageNet-256). We compare OG against Gradient Guidance and at approximately matched moderate target-density discrepancy .
| Method | (matched) | FID ↓ | ↓ |
|---|---|---|---|
| OG | 0.228 | 7.14 | 0.26 |
| Gradient Guidance | 0.227 | 7.15 | 0.78 |
| 0.230 | 7.25 | 22.3 |
5. Inverse Problems
5.1. Image inpainting and deblurring on Rectified Flow (CelebA-HQ-256). We compare OG against DPS and deterministic FlowDPS ( , the “w/o ” 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.
| Task | Method | Measurement error (matched) | ↓ | Runtime (s/img) ↓ |
|---|---|---|---|---|
| Inpainting vs DPS | OG | 0.517 | 258 | 0.483 |
| DPS | 0.516 | 259 | 0.482 | |
| Inpainting vs FlowDPS | OG | 0.506 | 374 | 0.483 |
| FlowDPS () | 0.504 | 806 | 0.485 | |
| Deblurring vs DPS | OG | 0.402 | 782 | 0.497 |
| DPS | 0.400 | 673 | 0.496 | |
| Deblurring vs FlowDPS | OG | 0.402 | 782 | 0.497 |
| FlowDPS () | 0.400 | 1,364 | 0.524 |
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.
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:
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.