acceptodds
Under review as a conference paper at ICLR 2027

P-Guide: Parameter-Efficient Prior Steering for Single-Pass CFG Inference

Abstract

Classifier-Free Guidance (CFG) is widely used for conditional generation in flow matching, yet its standard dual-pass form requires two backbone evaluations at each sampling step. We introduce P-Guide, a framework that applies guidance to the mean of a learned initial latent distribution and then uses one backbone evaluation per step. The resulting prior shift induces a local first-order trajectory response without explicit velocity-field extrapolation during sampling. We consider both homoscedastic and heteroscedastic priors; their comparable ImageNet performance indicates that learning the variance is secondary to mean steering in this setting. Experiments on class-conditional generation show that P-Guide reduces guided sampling GFLOPs by approximately 50% at the same step count while adding only a 1.31M-parameter prior module; on ImageNet, prior guidance reduces FID from 27.78 to 22.33 (U-Net) and from 32.84 to 24.61 (DiT-B/2) while remaining single-pass, offering a cost-effective balance between sampling efficiency and generation quality at moderate guidance scales.

Primary area:
generative models
Keywords:
Classifier-Free Guidance, Flow Matching, Prior Steering, Conditional Generation, Generative Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_StiVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper investigates whether CFG-like transformations of the source distribution can replace per-step classifier-free guidance during Flow Matching inference. Specifically, it proposes P-Guide, which extrapolates the conditional and unconditional means of a learned Gaussian prior to enable single-pass inference. Experiments on ImageNet and CIFAR-10 demonstrate improvements in generation quality with reduced inference costs.

Critical strengths

The paper presents a meaningful modeling perspective by shifting conventional per-step classifier-free guidance from velocity-space extrapolation to source-distribution-level guidance. This approach aims to reduce the computational overhead of CFG by requiring only a single backbone evaluation per sampling step. The research problem is well-motivated and practically relevant, particularly for improving the inference efficiency of Flow Matching models.

Critical weaknesses
  1. Limited effectiveness and unclear benefits of CFG-inspired extrapolation.

Although P-Guide achieves certain improvements in generation quality, these gains are primarily observed when the guidance scale is close to 1, while stronger extrapolation leads to significant performance degradation. This raises concerns about the effectiveness of the proposed CFG-inspired linear extrapolation. In particular, existing methods have already explored learning transformations of the source distribution, and P-Guide appears to reformulate such transformations by introducing a residual term controlled by the guidance scale. Given that the method is only effective under relatively small extrapolations, it remains unclear whether the observed improvements arise from a meaningful CFG-like guidance mechanism or simply from a residual adjustment to the learned source distribution.

  1. Insufficient theoretical justification for the connection to conventional CFG.

Although the paper establishes a local first-order relationship between source-space perturbations and trajectory-level guidance, it does not sufficiently justify why extrapolating the learned conditional and unconditional source means should approximate the effect of conventional CFG. In particular, exponential tilting of the source distribution does not necessarily correspond to the desired tilting of the generated data distribution. The theoretical analysis identifies an alignment condition, but the proposed prior-learning objective does not explicitly enforce it. Further theoretical or empirical evidence would help establish the connection between the proposed source-space extrapolation and conventional CFG.

  1. Insufficient comparisons with existing source-distribution modification methods.

Several previous works, such as CAR-Flow, CSFM, and NoiseRefine, have explored improving generation by modifying or learning the source distribution. However, the experiments primarily compare P-Guide against standard CFM and CFG, without providing direct comparisons with representative source-modification methods under unified settings. Since the proposed approach is closely related to these methods, such comparisons are important for establishing the effectiveness and distinct contribution of the proposed extrapolation strategy.

Questions for authors
  1. Could the authors provide further theoretical justification or empirical evidence demonstrating that the observed improvements specifically arise from the proposed CFG-inspired linear extrapolation, rather than simply from introducing a residual parameterization of the source-distribution transformation? In particular, it remains unclear whether the conditional–unconditional mean difference provides a unique advantage over other source-space perturbations.

  2. Could the authors include comparisons with existing source-distribution transformation methods, such as CAR-Flow or CSFM, under comparable experimental settings? Such comparisons would help clarify the necessity and effectiveness of the proposed extrapolation mechanism beyond existing learned-source approaches.

7 comments

Thank you for your suggestions on source-space guidance. The following ablations and fixed-checkpoint comparisons examine the conditional–unconditional direction and its value beyond source learning.

W1 / Q1

The norm-matched direction ablation identifies the conditional–unconditional mean contrast as an effective guidance direction. Sampling starts from

+(w-1)\alpha(\mu_c-\mu_u).$$ At $w=1$, this is the learned conditional source. We fix the source predictor, flow backbone, noise, and conditional standard deviation, replacing only the direction in the last term. Random and label-permuted directions are norm-matched in the conditional covariance metric. Evaluation uses the $\alpha=0.1$, 300K DiT-B/2 checkpoint, 50 Euler steps, and 50K images. Entries are FID; the $w=1$ reference is 30.32. | Source direction | w=1.2 | w=1.5 | | --- | ---: | ---: | | Conditional − null | 28.79 | 26.78 | | Reversed | 32.04 | 34.70 | | Random, norm-matched | 30.34 | 30.37 | | Permuted label, norm-matched | 29.86 | 29.30 | The conditional–unconditional direction gives the lowest FID at both strengths. Random directions remain near the unsteered control, while reversal worsens FID. These differences at fixed model weights and perturbation magnitude link the gain to the conditional information in the direction. **The useful extrapolation range depends on actual source displacement.** The same checkpoint gives: | PG | FID ↓ | KID ↓ | IS ↑ | Precision ↑ | Recall ↑ | Top-1 (%) ↑ | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | 0.9 | 31.17 | 0.02576 | 44.51 | 0.4705 | 0.6514 | 42.64 | | 1 | 30.32 | 0.02465 | 45.69 | 0.4757 | 0.6477 | 43.46 | | 1.5 | 26.78 | 0.02007 | 50.45 | 0.4948 | 0.6297 | 46.87 | | 2 | 24.36 | 0.01654 | 54.47 | 0.5062 | 0.6113 | 49.03 | | 3 | 23.24 | 0.01314 | 57.21 | 0.4904 | 0.5841 | 49.63 | At $\alpha=0.1$, FID improves through $w=3$, reaching 23.24, while recall decreases. The actual perturbation is $(w-1)\alpha(\mu_c-\mu_u)$, so guidance strength should be read together with the source scale and mean contrast. This configuration supports FID gains over a wider coefficient range, with a fidelity–coverage trade-off. **Planned manuscript revision.** Emphasize source-space guidance in Section 4.4 (p. 5) and add the direction/scale ablations to Section 5.3 (pp. 7–9) and Appendix C (pp. 24–27). **W3 / Q2** **Thank you for suggesting a direct comparison. CAR-Flow and CSFM use conditional sources to improve flow-matching learning; P-Guide adds sampling-time guidance to a learned source.** On a fixed CSFM model, PG=1 retains the original conditional source and other scales adjust only its mean, directly testing guidance after source learning. The comparison fixes the 280K CSFM checkpoint, source encoder, flow backbone, and conditional variance, varying only PG and CFG. All settings use the same native-BF16 sampler, 20 Euler steps, 50K images, and the original CSFM FID protocol. | PG | CFG | NFE | FID ↓ | IS ↑ | Top-1 (%) ↑ | CLIP ↑ | | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | 0.9 | 1 | 20 | 7.24 | 160.11 | 65.77 | 0.33641 | | 1 | 1 | 20 | 7.07 | 164.14 | 66.19 | 0.33731 | | 1.5 | 1 | 20 | 6.91 | 180.17 | 67.66 | 0.33926 | | 2 | 1 | 20 | 7.16 | 183.93 | 67.80 | 0.33996 | | 0.9 | 2 | 40 | 6.53 | 156.10 | 67.02 | 0.34409 | | 1 | 2 | 40 | 6.40 | 159.93 | 67.67 | 0.34480 | | 1.5 | 2 | 40 | 5.79 | 177.30 | 69.32 | 0.34671 | | 2 | 2 | 40 | 5.73 | 186.78 | 69.96 | 0.34759 | At CFG=1, increasing PG from 1 to 1.5 reduces FID from 7.07 to 6.91 and raises CLIP from 0.33731 to 0.33926. PG=2 further improves IS, Top-1, and CLIP, with FID of 7.16. At CFG=2, PG=2 reduces FID from 6.40 to 5.73 and raises CLIP from 0.34480 to 0.34759. Source guidance thus offers an adjustable quality–alignment trade-off on the same model and improves both FID and CLIP when combined with velocity CFG. The comparison demonstrates P-Guide's extension of an existing conditional-source model and its complementary role alongside velocity CFG. The related-work revision will distinguish source construction or refinement in CAR-Flow, CSFM, and NoiseRefine from sampling-time control of a learned source. **Planned manuscript revision.** Clarify the related methods in Section 2 (p. 2) and add the paired fixed-CSFM-checkpoint comparison to Section 5.3 (pp. 7–9).

W2

Thank you for this question. Mean extrapolation at fixed conditional covariance exactly reweights the source density, and the same conditional flow transfers that reweighting to the output. The derivation below specifies its connection to conditional–marginal data tilting and to velocity CFG.

Write my=αμym_y=\alpha\mu_y, mu=αμum_u=\alpha\mu_u, and d=my−mud=m_y-m_u. For fixed positive-definite conditional covariance Σy\Sigma_y, define

\bar{q}_u^{\\,y}=\mathcal{N}(m_u,\Sigma_y),\quad q_w=\mathcal{N}(m_u+wd,\Sigma_y).$$ The reference $\bar{q}_u^{\\,y}$ uses the null-condition mean and conditional covariance, matching mean-only guidance. Expanding the Gaussian log densities gives $$q_w=Z_w^{-1}q_y^w(\bar{q}_u^{\\,y})^{1-w},\qquad \nabla\log q_w=w\nabla\log q_y+(1-w)\nabla\log\bar{q}_u^{\\,y}.$$ For the same invertible conditional flow $\Phi_y$, let $$p_y=(\Phi_y)_{\\#}q_y,\quad r_u^y=(\Phi_y)_{\\#}\bar{q}_u^{\\,y},\quad p_w=(\Phi_y)_{\\#}q_w.$$ The change-of-variables Jacobian is shared, and the exponents sum to one, giving $$p_w(x)=Z_w^{-1}p_y(x)^w r_u^y(x)^{1-w}.$$ Thus, source tilting transfers exactly to output tilting between $p_y$ and $r_u^y$. It recovers the desired tilt relative to the marginal data density $p_u$ when $p_y$ matches the conditional data density and $r_u^y=p_u$. Conditional-transport training constrains the former; the latter is an additional reference-distribution condition. **Endpoint responses characterize the first-order relation to velocity CFG.** With finite second moments, a realizable mean model at the population optimum, and independent condition dropout, NLL gives $\mu_y=\mathbb{E}[X\mid y]$ and $\mu_u=\mathbb{E}[X]$. Along $\dot{x}_t=v_y(x_t,t)$, let $g_t=v_y(x_t,t)-v_u(x_t,t)$, $A_t=\partial_xv_y(x_t,t)$, and $\dot{J}_t=A_tJ_t$ with $J_0=I$. Differentiating at $\eta=0$ for $w=1+\eta$ gives $$\dot{P}_t=A_tP_t,\quad P_0=d;\qquad \dot{G}_t=A_tG_t+g_t,\quad G_0=0,$$ $$P_1=J_1d,\qquad G_1=J_1d_{\star},\qquad d_{\star} = \int_0^1J_t^{-1}g_t\\,dt.$$ Under local twice differentiability and bounded relevant derivatives, $$\lVert x_1^{\mathrm{PG}}(\eta)-x_1^{\mathrm{CFG}}(\eta)\rVert \le |\eta|\\,\lVert J_1(d-d_{\star})\rVert +C\eta^2.$$ First-order endpoint matching requires $d=d_{\star}$. NLL learns a class-shared mean direction $d$, whereas $d_{\star}$ can vary across trajectories. We measure the resulting response difference directly. Using 1,000 paired class/noise inputs, centered finite differences in FP32, and perturbation size 0.03, the mean cosine similarities to the velocity-CFG endpoint response are: | Source direction | 20 steps | 50 steps | | --- | ---: | ---: | | Conditional − null | 0.2630 | 0.1882 | | Random | 0.0014 | 0.0026 | | Permuted label | 0.0721 | 0.0667 | The conditional–unconditional direction has weak positive correlation with the velocity-CFG response, above the random and label-permuted controls; mean relative errors are 1.08 and 1.17 at 20 and 50 steps. Together with the W1 / Q1 ablation, the results establish the generation benefit of this direction and quantify its limited first-order alignment with velocity CFG. **Planned manuscript revision.** Clarify first-order responses and density reweighting in Section 4.3 (p. 5) and Appendix B (pp. 16–23), with endpoint diagnostics in Appendix C (pp. 24–27).
Reviewer_StiVOpenReview ↗

Response to Authors

Thank you for the detailed response and the additional experiments. The norm-matched direction ablation and the extended scale sweep are informative. However, my main concerns remain.

  1. Clarification on "residual." By "residual," I did not mean an arbitrary perturbation direction in source space. I meant the residual parameterization of the source transformation: instead of learning a direct transformation f(x), P-Guide parameterizes the source as x + f(x), i.e., z = ε + αμ(y), and the guided source z_cfg = z_c + α(w−1)(μ(y) − μ(∅)) keeps this residual form. The direction ablation shows that the conditional–null direction carries conditional information. It does not tell whether the improvement comes from the CFG-inspired linear extrapolation or from the residual parameterization itself.

  2. Comparison with direct source-transformation baselines. To separate these two effects, P-Guide should be compared with methods that learn a direct (non-residual) source transformation z = f(ε, y), such as CAR-Flow, CSFM, and NoiseRefine, under unified settings (same backbone, training budget, NFE, and evaluation protocol). Together with P-Guide at w = 1 (residual source without extrapolation) and at w > 1, this comparison would show how much of the gain comes from the residual parameterization and how much comes from the extrapolation. The fixed-CSFM experiment shows that P-Guide can be applied on top of an existing conditional-source model, but it is not such a comparison, and NoiseRefine is not addressed in the response. Moreover, at CFG = 1, P-Guide improves FID only from 7.07 to 6.91 on the CSFM checkpoint, and FID worsens at PG = 2, which suggests that much of the benefit may come from source learning rather than from the extrapolation itself.

Summary. I appreciate the authors' efforts and the honest analysis in the response. Because the gains from linear extrapolation have not been separated from those of the residual parameterization, and direct comparisons with existing source-transformation methods are still missing, I will maintain my original rating.

Reviewer_8dfCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces P-Guide, a parameter-efficient framework that replaces iterative velocity-space classifier-free guidance (CFG) with prior-space steering to enable single-pass inference in flow matching models. The method learns condition-dependent Gaussian priors and performs CFG-style extrapolation over their mean and variance parameters, supported by a first-order trajectory approximation. Experiments on MNIST, CIFAR-10, and ImageNet-1k demonstrate reduced inference computation and competitive generation quality in selected settings, although the theoretical equivalence and the effectiveness of prior-space guidance remain insufficiently validated.

Critical strengths
  1. well-motivated problem. The paper addresses the computational overhead of standard CFG, which requires conditional and unconditional network evaluations at every sampling step. Reducing this overhead is practically meaningful, particularly for resource-constrained generative systems.
  2. The theoretical analysis is comprehensive and reasonably rigorous.
  3. P-Guide introduces a lightweight prior module of approximately 1.31M parameters and substantially reduces inference FLOPs compared with standard dual-pass CFG.
  4. The experiments cover MNIST, CIFAR-10, and ImageNet-1k, including both U-Net and DiT-based backbones. The authors also investigate prior parameterization, module capacity, pretrained initialization, and guidance strength. The additional experiments provide insights into the effects of mean and variance steering, guidance-scale sensitivity, and compatibility with standard CFG.
Critical weaknesses

Weakness:

  1. Several important experiments, such as SANA text-to-image generation, are relegated to the appendix. I suggest moving these experiments to the main paper and shifting the MNIST results to the appendix. Key ablations should be conducted on more challenging datasets, preferably ImageNet or CIFAR-10 if computational resources are limited.
  2. Table 2 is more representative than Table 1 but reports only FID. I encourage the authors to include KID, IS, classification accuracy, and precision/recall for a more comprehensive evaluation of generation quality and conditional fidelity.
  3. The SANA experiments are important and deserve a more thorough evaluation in the main paper. In particular:
    • Report additional metrics, such as FID and CLIP Score, to assess generation quality and text-image alignment.
    • Evaluate P-Guide without standard CFG to validate its single-pass guidance capability.
    • Investigate different CFG guidance scales to assess robustness and effectiveness.
  4. Although Corollary B.1 establishes the first-order alignment condition Δzα=∫01Jt−1gt,dt\Delta z_\alpha=\int_0^1J_t^{-1}g_t,dt, the Gaussian NLL objective used to train P-Guide does not guarantee that this condition is satisfied. The connection between the theoretical criterion and the proposed learning objective remains unclear.
Questions for authors

Question:

  1. Why do the authors not consider a learnable isotropic Gaussian prior, i.e., N(μϕ(y),σϕ2(y)I)\mathcal{N}(\mu_\phi(y),\sigma_\phi^2(y)I), as an intermediate setting between fixed and diagonal covariance?
  2. Why is the backbone trained using the original conditional prior in Stage 2, while extrapolated priors are used during inference?
  3. Could condition dropout be extended to Stage 2 Flow Matching training to learn both conditional and unconditional transports? Would this improve robustness at larger guidance scales?
  4. Could the behavior of standard CFG be distilled into P-Guide through fine-tuning or knowledge distillation, enabling P-Guide-only inference? I believe this would significantly enhance the practical value of the method.
  5. How sensitive is P-Guide to the number of ODE solver steps? I encourage the authors to evaluate ImageNet at NFE = 10, 20, 50, and 100 to better characterize the quality–efficiency trade-off.
  6. In Appendix E, Figure 7(a), the PG-LoRA + P-Guide result appears to exhibit a noticeable anatomical artifact in the left hand. Is this an isolated case or a recurring issue?
  7. When w>1w>1, the extrapolated source deviates from the distribution used during backbone training, ptrain(z∣y)=N(αμc,Σc)p_{\mathrm{train}}(z\mid y)=\mathcal{N}(\alpha\mu_c,\Sigma_c). Could the authors quantify this distribution shift using DKL(pguidew∣ptrain)D_{\mathrm{KL}}(p_{\mathrm{guide}}^w|p_{\mathrm{train}}) and plot both KL divergence and FID against the guidance scale ww (using separate y-axes)? This could help investigate whether source distribution mismatch is associated with performance degradation at larger guidance scales.
6 comments

Thank you for your detailed suggestions on theory, evaluation, and presentation. We address each point with derivations, matched ablations, and text-to-image results.

W1

We will move text-conditional CSFM and full-COCO SANA experiments into the main paper, and detailed MNIST results into the appendix. The original SANA study used a COCO subset as an initial text-to-image test. The new evaluations cover full-COCO adaptation and guidance on an existing conditional-source model.

The comparison fixes the 280K CSFM checkpoint, source encoder, flow backbone, and conditional variance, varying only PG and CFG. All settings use the same native-BF16 sampler, 20 Euler steps, 50K images, and the original CSFM FID protocol.

PGCFGNFEFID ↓IS ↑Top-1 (%) ↑CLIP ↑
0.91207.24160.1165.770.33641
11207.07164.1466.190.33731
1.51206.91180.1767.660.33926
21207.16183.9367.800.33996
0.92406.53156.1067.020.34409
12406.40159.9367.670.34480
1.52405.79177.3069.320.34671
22405.73186.7869.960.34759

At CFG=1, increasing PG from 1 to 1.5 reduces FID from 7.07 to 6.91 and raises CLIP from 0.33731 to 0.33926. PG=2 further improves IS, Top-1, and CLIP, with FID of 7.16. At CFG=2, PG=2 reduces FID from 6.40 to 5.73 and raises CLIP from 0.34480 to 0.34759. Source guidance thus offers an adjustable quality–alignment trade-off on the same model and improves both FID and CLIP when combined with velocity CFG.

The key ablations use ImageNet and cover covariance parameterization, condition dropout, guidance direction, and sampling budget. We report them under the corresponding questions below.

Planned manuscript revision. Include text-conditional CSFM, full-COCO SANA, and key ImageNet ablations in Section 5 (current pp. 6–9). Move detailed MNIST results to the appendix and retain detailed SANA protocols in Appendix E.

W2

The additional metrics characterize the fidelity, conditional accuracy, and coverage trade-off of source guidance. We fix the α=0.1\alpha=0.1, 300K checkpoint and use 50 Euler steps, NFE=50, and 50K images per setting. FID follows the original LFM protocol; KID is unscaled.

PGFID ↓KID ↓IS ↑Precision ↑Recall ↑Top-1 (%) ↑
0.931.170.0257644.510.47050.651442.64
130.320.0246545.690.47570.647743.46
1.526.780.0200750.450.49480.629746.87
224.360.0165454.470.50620.611349.03
323.240.0131457.210.49040.584149.63

Increasing ww from 1 to 2 reduces FID from 30.32 to 24.36 and lowers KID, while improving IS and classification accuracy. Precision rises and recall falls, characterizing the fidelity–coverage trade-off induced by guidance.

Planned manuscript revision. Expand the ImageNet metric table in Section 5.3 (pp. 7–9) and specify the evaluation protocols in Appendix A (pp. 14–15).

W3

SANA results show improved FID with one backbone prediction per step and compatibility with velocity CFG. The LoRA models use full-COCO adaptation weights. Each configuration generates 50K images from the same 5K-prompt pool using 20 Euler steps; FID uses 5K COCO validation images as reference. CFG=1 gives NFE=20; CFG=4.5 gives NFE=40.

ModelPGCFGNFEFID ↓CLIP ↑
Pretrained SANA—12027.720.23680
Pretrained SANA—4.54031.360.26563
Baseline LoRA—12027.800.21300
Source LoRA112027.650.21273
Source LoRA212026.260.21410
Source LoRA312025.370.21509
Baseline LoRA—4.54025.050.26824
Source LoRA14.54024.950.26801
Source LoRA24.54025.030.26800
Source LoRA34.54025.180.26792

With source guidance alone, increasing PG from 1 to 3 reduces FID from 27.65 to 25.37 and raises CLIP from 0.21273 to 0.21509. PG=3 approaches baseline LoRA's FID at CFG=4.5 (25.05) with half the backbone evaluations; the latter has higher CLIP (0.26824), illustrating the compute–alignment trade-off.

At CFG=4.5, PG=1 to 3 gives FID of 24.95–25.18 and CLIP of 0.26792–0.26801, with little variation. Source guidance gives a clearer FID benefit without velocity CFG, while stronger velocity CFG provides higher text alignment.

Planned manuscript revision. Present the full-COCO SANA text-to-image results alongside CSFM in Section 5 (current pp. 6–9), with detailed protocols and extended evaluation in Appendix E.

W4

Thank you for asking us to clarify the link between the objective and theoretical criterion. NLL learns conditional and marginal means; Corollary B.1 gives the condition for their contrast to match velocity CFG's first-order endpoint response. Decomposing the error into mean estimation and dynamical alignment makes their roles explicit.

Let XX denote the data latent and Y~\widetilde{Y} the label after independent condition dropout. With finite second moments and fixed positive-definite covariance ScS_c, Gaussian NLL satisfies

=\frac12\lVert \mu-\mu_c^{\star}\rVert _{S_c^{-1}}^2, \qquad \mu_c^{\star}=\mathbb{E}[X\mid\widetilde{Y}=c].$$ With a realizable mean model at the population optimum, $\mu_y^{\star}=\mathbb{E}[X\mid Y=y]$ and $\mu_u^{\star}=\mathbb{E}[X]$. The same mean optimum holds when learned variance has a positive population optimum. Thus, $d=\alpha(\mu_y-\mu_u)$ is a conditional–marginal mean contrast. **For fixed conditional covariance, mean extrapolation exactly reweights the source density.** Write $m_y=\alpha\mu_y$, $m_u=\alpha\mu_u$, and fix positive-definite $\Sigma_y$: $$q_y=\mathcal{N}(m_y,\Sigma_y),\quad \bar{q}_u^{\\,y}=\mathcal{N}(m_u,\Sigma_y),\quad q_w=\mathcal{N}(m_u+w(m_y-m_u),\Sigma_y).$$ The reference $\bar{q}_u^{\\,y}$ uses the null-condition mean and conditional covariance, as in mean-only guidance. Expanding the Gaussian log densities gives $$q_w=Z_w^{-1}q_y^w(\bar{q}_u^{\\,y})^{1-w},\qquad \nabla\log q_w=w\nabla\log q_y+(1-w)\nabla\log\bar{q}_u^{\\,y},$$ $$d=\Sigma_y\nabla_z\log\frac{q_y}{\bar{q}_u^{\\,y}}.$$ **Endpoint responses relate this source direction to velocity CFG.** Along $\dot{x}_t=v_y(x_t,t)$, let $A_t=\partial_xv_y(x_t,t)$, $g_t=v_y(x_t,t)-v_u(x_t,t)$, and $\dot{J}_t=A_tJ_t$ with $J_0=I$. Differentiating at $\eta=0$ for $w=1+\eta$ gives $$\dot{P}_t=A_tP_t,\quad P_0=d;\qquad \dot{G}_t=A_tG_t+g_t,\quad G_0=0.$$ The integrating-factor solution is $$P_1=J_1d,\qquad G_1=J_1d_{\star},\qquad d_{\star} = \int_0^1J_t^{-1}g_t\\,dt.$$ Under local twice differentiability in state and perturbation, with bounded relevant derivatives, $$\lVert x_1^{\mathrm{PG}}(\eta)-x_1^{\mathrm{CFG}}(\eta)\rVert \le |\eta|\\,\lVert J_1(d-d_{\star})\rVert +C\eta^2.$$ For a fixed reference flow, identity covariance, and $\alpha\ge0$, let $\varepsilon_y,\varepsilon_u$ be the population excess NLL risks of the two means. Then $$\lVert \widehat{d}-d_{\star}\rVert \le\alpha\bigl(\sqrt{2\varepsilon_y}+\sqrt{2\varepsilon_u}\bigr) +\lVert d_{\mathrm{pop}}-d_{\star}\rVert ,\qquad d_{\mathrm{pop}}=\alpha\bigl(\mathbb{E}[X\mid y]-\mathbb{E}[X]\bigr).$$ NLL controls the first term, the mean-estimation error. The second measures the gap between the population mean contrast and the dynamical direction. First-order matching in Corollary B.1 requires $d=d_{\star}$, an additional dynamical condition. We measure mean cosine similarity to the velocity-CFG endpoint response using 1,000 paired class/noise inputs, centered finite differences in FP32, and perturbation size 0.03: | Source direction | 20 steps | 50 steps | | --- | ---: | ---: | | Conditional − null | 0.2630 | 0.1882 | | Random | 0.0014 | 0.0026 | | Permuted label | 0.0721 | 0.0667 | The conditional–unconditional direction has weak positive correlation, above the random and label-permuted controls. Mean relative response errors are 1.08 and 1.17 at 20 and 50 steps. NLL therefore supplies a condition-dependent mean direction, while the endpoint diagnostic quantifies its dynamical difference from velocity CFG. **Planned manuscript revision.** Clarify source-statistic learning and the alignment criterion in Section 4.3 (p. 5) and Appendix B (pp. 16–23), and add endpoint-response diagnostics to Appendix C (pp. 24–27).

Thank you for your questions about prior parameterization, training, and source shift. We address them with matched ablations and the KL diagnostic below.

Q1

Thank you for this suggestion. Adding a learned isotropic prior gives similar FID across the three covariance parameterizations in the matched comparison. The groups share backbone initialization, 40K prior-fitting steps, 300K backbone-adaptation steps, α=0.1\alpha=0.1, global batch size 128, and Stage-2 dropout 0.1. Each adapted checkpoint generates 50K images with 50 Euler steps and NFE=50.

Covariancew=0.8w=1w=1.2w=2
Fixed24.2722.6721.2917.34
Learned isotropic24.8822.9821.3117.05
Learned diagonal24.9823.1321.4917.36

Entries are FID. At each guidance scale, the three parameterizations differ by less than one FID point, and all improve as ww increases. Mean guidance is effective with fixed, isotropic, and diagonal covariance; each model's conditional covariance stays fixed during the ww sweep.

Planned manuscript revision. Add the covariance ablation to Appendix C (pp. 24–27) and summarize it in Section 5.3.

Q2

A fixed training source allows one backbone to support multiple inference-time guidance strengths. Stage 2 learns transport from the conditional source to the data. At sampling time, w=1w=1 recovers the training source, w>1w>1 extrapolates along the conditional–unconditional mean direction, and w<1w<1 weakens that direction.

With the flow map fixed, changing the source mean changes the generated distribution. Training a separate backbone for each ww to map its source to the same conditional data distribution would produce the same endpoint distribution under ideal fitting. We therefore train on a fixed source and use mean extrapolation for inference-time control. Q7 quantifies this change through source KL and FID.

Planned manuscript revision. Clarify the relation between two-stage training and inference control in Sections 4.2 (p. 4) and 4.4 (p. 5).

Q3

We compared matched Stage-2 training with condition dropout of 0 and 0.1: dropout=0 gives slightly lower FID, and both improve as PG increases. The groups share the prior, backbone initialization, and training budget; prior-fitting dropout is 0.1 in both. Evaluation uses 50 Euler steps and 50K images per setting.

Stage-2 dropoutw=1w=1.2w=2
022.4621.0417.13
0.122.6721.2917.34

Prior-fitting dropout learns the null-condition source mean used by P-Guide; Stage-2 dropout learns the unconditional velocity used by velocity CFG. Source guidance works with either Stage-2 setting, while learning the unconditional velocity gives no further FID benefit at stronger source guidance in this comparison.

Planned manuscript revision. Add the matched Stage-2 dropout ablation to Appendix C (pp. 24–27) and distinguish the two stages’ dropout roles in Section 4.2 (p. 4).

Q7

Source shift has an exact KL expression, which we plot alongside FID on the requested dual axes. Let d=α(μc−μu)d=\alpha(\mu_c-\mu_u). Guidance preserves the conditional covariance Σc\Sigma_c, so

=\frac{(w-1)^2}{2}d^\top\Sigma_c^{-1}d.$$ The table reports class-averaged KL for the $\alpha=0.1$, 300K model, together with FID from 50 Euler steps and 50K images per setting. | PG scale | Mean source KL | FID ↓ | | --- | ---: | ---: | | 1 | 0.0000 | 30.32 | | 1.2 | 0.0358 | 28.79 | | 1.5 | 0.2237 | 26.78 | | 2 | 0.8946 | 24.36 | | 3 | 3.5785 | 23.24 | KL grows quadratically with source displacement, while FID improves through $w=3$ in this sweep. Thus, larger source shifts coexist with lower FID over this range. The figure will identify $\alpha=0.1$ so the relation between guidance scale and actual source displacement is explicit. **Planned manuscript revision.** Add the dual-axis KL–FID figure and source-shift analysis to Appendix C (pp. 24–27), with a reference in Section 4.4 (p. 5).
Reviewer_2VZGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

P-Guide learns a conditional Gaussian source and trains or adapts a flow-matching backbone to it. During sampling, it extrapolates the source mean away from the null-condition mean, then uses only the conditional backbone. Class-conditional experiments show a quality–compute trade-off, while a small text-to-image pilot tests compatibility with conventional CFG.

Critical strengths
  1. The method provides a simple way to reduce sampling cost without guided-output supervision. It gives an encouraging result at a fixed budget: P-Guide obtains FID 26.90 versus 30.01 for CFG using ten backbone evaluations.

  2. The paper reports both useful operating points and failure cases. The distributed-injection study and the distinction between local response and CFG alignment help explain the limits of a single initial mean shift.

Critical weaknesses
  1. Guidance-Free Training (GFT; Chen et al., 2025) [1] and guidance distillation (Meng et al., 2023) [2] also avoid dual-pass CFG at inference. The paper discusses distillation but offers no experimental comparison with either approach, leaving its practical advantage over these alternatives unclear. Table 3 is encouraging; a controlled comparison at matched sampling budgets, with adaptation costs reported, would strengthen the evidence.

  2. The cost of adopting the method is difficult to assess. Both prior training and backbone adaptation are required, but training time or FLOPs are not reported. The small SANA pilot also retains conventional CFG. It demonstrates compatibility, as the authors acknowledge, but does not yet show a text-to-image sampling benefit that could offset the adaptation cost.

[1] Huayu Chen, Kai Jiang, Kaiwen Zheng, Jianfei Chen, Hang Su, and Jun Zhu. Visual Generation Without Guidance. ICML, 2025.

[2] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. CVPR, 2023.

Questions for authors
  1. Could you compare against GFT or the guidance-distillation stage of Meng et al. (2023), using a common backbone and matched sampling-call budget while reporting adaptation costs? Including P-Guide without extrapolation at the Table 3 budget would also help isolate the value of mean steering.

  2. Could you clarify the separate prior-training and backbone-adaptation budgets and their training time, particularly in the SANA setup? For the text extension, a run without conventional CFG would help assess whether cheaper sampling preserves image quality and prompt following.

3 comments

Thank you for your suggestions on guidance comparisons, adaptation cost, and text-to-image generation. We address these points with matched-budget composition experiments, training costs, and SANA results.

W1 / Q1

P-Guide combines with GFT and guidance distillation to further improve FID at the same inference budget. GFT incorporates guidance into a single conditional prediction through training; Meng et al.'s first stage trains a student to match the teacher's combined conditional and unconditional predictions. P-Guide adjusts the initial source mean and can be used with either velocity model.

We implement both guidance objectives in DiT-B/2 flow-matching models, with 20K additional training steps per model. The P-Guide combinations start from the α=0.1\alpha=0.1, 300K conditional-source backbone. The Meng-style teacher uses CFG=1.5; GFT uses β=2/3\beta=2/3 at inference. Gaussian and conditional sources have separately adapted velocity models. All settings use 10 Euler steps, NFE=10, 50K images, and the original LFM FID protocol.

SourcePGMeng-style FID ↓GFT FID ↓
Gaussian—21.8415.16
Conditional0.919.3414.54
Conditional118.9614.16
Conditional1.218.2213.48
Conditional1.517.2812.53
Conditional216.1211.52
Conditional315.1710.71

The w=1w=1 row is the conditional-source control without extrapolation. Each PG sweep shares the prior, velocity model, and conditional standard deviation, changing only the initial mean. Increasing PG from 1 to 3 reduces FID from 18.96 to 15.17 for Meng-style distillation and from 14.16 to 10.71 for GFT, the lowest values in the evaluated range of 0.8 to 3. Source guidance therefore further improves FID after guidance training or distillation, retaining one backbone prediction per step.

Prior-fitting and backbone-adaptation times are reported under W2 / Q2.

Planned manuscript revision. Add the composition comparison to Section 5.3 (current pp. 7–9), with initialization and additional training budgets in Appendix A (pp. 14–15).

W2 / Q2

ImageNet-256 prior fitting takes approximately 0.5 hours and backbone adaptation 32 hours, both on one RTX 4090.

ImageNet training stage (one RTX 4090)Optimization stepsTime
Prior fitting40KApproximately 0.5 hours
Backbone adaptation400KApproximately 32 hours

Prior fitting takes about 1.6% of the backbone-adaptation time; the two stages total approximately 32.5 hours. SANA jointly trains the prior predictor and rank-16 LoRA. Both full-COCO runs use one RTX 4090 (24GB) for 4,620 steps: 360 minutes without PG and 395 minutes with PG, an increase of 35 minutes, or approximately 9.7%.

SANA results show improved FID with one backbone prediction per step and compatibility with velocity CFG. The LoRA models use full-COCO adaptation weights. Each configuration generates 50K images from the same 5K-prompt pool using 20 Euler steps; FID uses 5K COCO validation images as reference. CFG=1 gives NFE=20; CFG=4.5 gives NFE=40.

ModelPGCFGNFEFID ↓CLIP ↑
Pretrained SANA—12027.720.23680
Pretrained SANA—4.54031.360.26563
Baseline LoRA—12027.800.21300
Source LoRA112027.650.21273
Source LoRA212026.260.21410
Source LoRA312025.370.21509
Baseline LoRA—4.54025.050.26824
Source LoRA14.54024.950.26801
Source LoRA24.54025.030.26800
Source LoRA34.54025.180.26792

With source guidance alone, increasing PG from 1 to 3 reduces FID from 27.65 to 25.37 and raises CLIP from 0.21273 to 0.21509. PG=3 approaches baseline LoRA's FID at CFG=4.5 (25.05) with half the backbone evaluations; the latter has higher CLIP (0.26824), illustrating the compute–alignment trade-off.

At CFG=4.5, PG=1 to 3 gives FID of 24.95–25.18 and CLIP of 0.26792–0.26801, with little variation. Source guidance gives a clearer FID benefit without velocity CFG, while stronger velocity CFG provides higher text alignment.

Planned manuscript revision. Add adaptation costs to Appendix A (current pp. 14–15) and full-COCO SANA results, including controls without velocity CFG, to Section 5. Provide detailed evaluation protocols in Appendix E.

Thank you for the detailed responses and additional experiments, which have adequately addressed my main concerns. I find the contribution more convincing and will raise my score.

Dear Reviewer 2VZG,

Thank you for reconsidering our work and for your willingness to raise your score. We sincerely appreciate the time and care you have devoted to reviewing and discussing our paper. Your constructive feedback helped us clarify P-Guide’s contribution and strengthen its experimental evaluation.

Best regards, The Authors

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your careful reviews and continued discussion. Your suggestions helped us separate the effects of source learning and guidance, strengthen the comparisons, and clarify the contribution.

1. Separating extrapolation from source learning and residual parameterization (2VZG, StiV)

The paired experiments keep the source predictor, backbone, residual parameterization, and conditional standard deviation fixed, changing only the extrapolation coefficient. Improvements on SANA, CSFM, Meng-style distillation, GFT, and the drifting model identify the additional benefit of sampling-time extrapolation over an already learned conditional source.

2. Source supervision and the connection to velocity-field CFG (8dfC, StiV)

We clarified how data-mean estimation relates to the trajectory-alignment condition. The revised analysis gives source-response and finite-scale error bounds, while direction ablations and endpoint diagnostics examine how source guidance relates to velocity-field CFG. This connects the statistical meaning of the source contrast to its effect on generation.

3. Efficient guidance and training cost (2VZG, 8dfC)

Matched-budget comparisons and composition experiments show that P-Guide can improve Meng-style guidance distillation and GFT without increasing backbone evaluations. The cost report separates prior fitting from backbone adaptation and includes SANA training times with and without PG, addressing 2VZG's question about the added training cost.

4. Text-to-image generation (2VZG, 8dfC)

CSFM and full-COCO SANA results now appear in the main text with text-alignment metrics and paired generation trajectories. Source-only SANA guidance reaches FID close to standard CFG with fewer backbone evaluations; standard CFG has higher CLIP. CSFM also shows gains from combining source guidance with velocity-field CFG.

5. Prior design, sampling, and distribution shift (8dfC)

New ablations cover covariance parameterization, condition dropout during backbone adaptation, solver steps, and guidance strength. Quality and diversity metrics and the requested KL–FID analysis describe how these choices affect generation. Paired samples across seeds document the hand artifact under different sampling settings.

6. Comparisons with source-transformation methods (StiV)

We compared the source-learning objectives of P-Guide, CAR-Flow, and CSFM. CAR-Flow evaluations and paired trajectories show that source extrapolation fails to provide effective guidance at the evaluated checkpoint, leaving the generated results nearly unchanged. CSFM's normalized alignment retains conditional directions useful for extrapolation in some settings. These comparisons distinguish the statistics learned by each source objective. We also discuss NoiseRefine's guided-image supervision and the unavailable code and weights that limited its reproduction.

Our main contributions

  • Extending conditional-source learning to guidance before transport. Learned-source methods place conditional information in the initial distribution. P-Guide uses the contrast between conditional and null source means as an adjustable sampling-time control. Changing its strength leaves the source predictor, flow backbone, and conditional standard deviation fixed.
  • Giving mean extrapolation an explicit basis in data statistics. Under fixed-variance squared-error source regression, the population-optimal conditional and null predictions are E[X∣y]\mathbb E[X\mid y] and E[X]\mathbb E[X]. Their difference defines a conditional shift relative to the overall data mean. The theory relates this source contrast to velocity-field CFG, and the source-objective comparisons show why a learned conditional source alone does not ensure effective extrapolation.
  • Providing efficient guidance that composes with existing generation methods. Source guidance is applied before integration and uses only forward generation. It can improve generation alongside velocity-field CFG, Meng-style distillation, and GFT, and extends to text-to-image models and few-step generators adapted to the learned source. The fixed-weight comparisons establish the incremental role of extrapolation across these settings.

Thank you again for helping us strengthen the paper and explain these contributions more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.