REVIEW 4 major objections 4 minor 18 references
REG: Rectified Gradient Guidance for Conditional Diffusion Models
T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read The paper proves that the standard marginal-scaling objective behind diffusion guidance is theoretically invalid, replaces it with a joint-scaling objective whose unique optimal update takes a ratio form, and introduces REG, a gradient…
desk verdict Joint-scaling theory is the real contribution; REG is a heuristic with an unproven diagonal-Jacobian approximation, and one Table 3 entry mislabels a worse IS as better. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the induced expected reward $E_t(x_t,y) = \int p_\theta(x_0|x_t,y)\,R_0(x_0,y)\,dx_0$, the average terminal reward over all clean images the model could still reach from the noisy state $x_t$. The argument works by showing that under joint scaling the unique reverse kernels are the ratio $E_t(x_t,y)/E_{t+1}(x_{t+1},y)$ times the original kernels, so the marginal at step $t$ is $\frac{E_t(x_t,y)}{E(y)}\,p_\theta(x_t|y)$ and the optimal noise prediction is $\epsilon_{\theta,t} - \sqrt{1-\bar\alpha_t}\,\nabla_{x_t}\log E_t(x_t,y)$. Because $E_t$ requires running the chain to the end, practical guidance substitutes the pointwise reward $R_t$; REG improves that substitution by noting that with a deterministic sampler $E_t(x_t,y) = R_0(\hat{x}_0,y)$, and applying the chain rule yields a Jacobian $\partial\hat{x}_0/\partial x_t$, which REG approximates as diagonal to produce the elementwise correction in Eq. (21). Theorem 4.1 also establishes that strict guidance changes the denoising start distribution to $\frac{E_T(x_T,y)}{E(y)}\,\mathcal{N}(x_T|0,I)$.
What would settle it
Measure the off-diagonal entries of $\partial\hat{x}_0/\partial x_t = \frac{1}{\sqrt{\bar\alpha_t}}\left(I - \sqrt{1-\bar\alpha_t}\,\partial\epsilon_{\theta,t}/\partial x_t\right)$ on a trained DiT or EDM2 model: if the off-diagonal block of the Jacobian is comparable to or larger than the diagonal, and a sampler using the exact Eq. (22) matrix-vector product does not clearly beat REG on FID and IS, then REG's justification for dropping the off-diagonals collapses.
Extended reading notes
Core claim
The central claim is that guidance should be constructed from a scaled joint distribution, $\bar{p}_\theta(x_{0:T}|y) \propto p_\theta(x_{0:T}|y)\,R_0(x_0,y)$, not from scaled marginal distributions as the literature assumes. Scaling a marginal, either the endpoint $p_\theta(x_0|y)$ or every $p_\theta(x_t|y)$, is proven invalid because the denoising transitions implicitly determine $R_{t-1}$ from $R_t$ up to a constant, so the prescribed rewards over-constrain the chain and no DDPM can realize the target. For the joint objective, Theorem 4.1 proves that the unique transition kernels are $\bar{p}_\theta(x_t|x_{t+1},y) = \frac{E_t(x_t,y)}{E_{t+1}(x_{t+1},y)}\,p_\theta(x_t|x_{t+1},y)$ with $E_t(x_t,y) = \mathbb{E}_{p_\theta(x_0|x_t,y)}[R_0(x_0,y)]$, which forces the optimal noise prediction $\bar{\epsilon}^*_{\theta,t} = \epsilon_{\theta,t} - \sqrt{1-\bar\alpha_t}\,\nabla_{x_t}\log E_t(x_t,y)$ and, notably, a modified denoising start distribution. Because computing $E_t$ requires foresight of the whole chain, the paper interprets CG, CFG, and AutoG as approximations that replace $\nabla_{x_t}\log E_t$ with $\nabla_{x_t}\log R_t$, and quantifies that approximation error in Theorems 4.2 and 4.3. REG is proposed as a better approximation: with a deterministic sampler, $E_t(x_t,y)$ becomes $R_0(\hat{x}_0,y)$, and the chain rule introduces the Jacobian $\partial\hat{x}_0/\partial x_t$, which REG keeps in diagonal form to obtain a cheap correction term; 1D and 2D synthetic experiments confirm REG tracks $\nabla_{x_t}\log E_t$ more closely than vanilla CFG, and large-scale tests show consistent FID, IS, and CLIP improvements.
Load-bearing premise
The load-bearing premise, admitted in the last paragraph of Section 4, is that the Jacobian $\partial\hat{x}_0/\partial x_t$ is approximately diagonal, which lets REG replace a full matrix-vector product with an elementwise product; the paper gives no error bound for that step, so if the Jacobian is strongly non-diagonal, REG's gains are heuristic rather than principled.
Editorial extensions
If this is right
- Strictly optimal guidance modifies the denoising start prior to $\bar{p}_\theta(x_T|y) = \frac{E_T(x_T,y)}{E(y)}\,\mathcal{N}(x_T|0,I)$, so any implementation that keeps a standard Gaussian start is an approximation; the paper leaves sampling from the corrected start to future work.
- CG, CFG, and AutoG are reclassified as approximations to the optimal update under a no-foresight constraint, with an approximation error that shrinks as denoising progresses and a bias that persists even in expectation (Theorems 4.2 and 4.3).
- Adding REG improves FID and Inception score on class-conditional ImageNet at 64×64, 256×256, and 512×512 across DiT-XL/2 and EDM2 models, for vanilla, cosine, linear, and interval CFG and for AutoG.
- Adding REG improves FID and CLIP score on COCO-2017 text-to-image generation with SD-v1-4 and SD-XL under vanilla, cosine, and linear CFG, at under 2× runtime and memory overhead in almost all measured settings.
- REG is method-agnostic: it is a correction applied on top of any reward-based guidance rather than a new guidance signal, so it inherits the reward design of the base method.
Reading between the lines
- Editorial inference: the decisive unproven step in REG is the diagonal approximation of $\partial\hat{x}_0/\partial x_t$ in Eq. (21); a direct test would compare REG against the exact matrix-vector product in Eq. (22), and if the exact product beats REG by a wide margin, the diagonal assumption, not the joint-scaling theory, is doing the work.
- Editorial inference: because the diagonal of $\partial\epsilon_{\theta,t}/\partial x_t$ is a divergence-like quantity, REG has the flavor of a trace correction to the score, which suggests an unexplored connection to the probability-flow ODE and to likelihood or flux estimates.
- Editorial inference: the no-foresight framing implies a ladder of approximations, one-step lookahead (REG), two-step lookahead, and so on up to the exact $E_t$; a testable extension is whether differentiating through two denoising steps instead of one further reduces the approximation error and improves FID.
- Editorial inference: the paper treats the corrected start distribution as an open question, but it is also a tunable lever; sampling the initial noise from a distribution biased by $E_T(x_T,y)$ could close the remaining gap between practice and the optimal joint-scaling sampler.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper identifies a mismatch between the standard marginal-scaling interpretation of guidance in diffusion models and the way guidance is actually implemented at sampling time. It argues that marginal scaling is generally infeasible, and proposes instead a joint-scaling objective over the whole denoising chain. Theorem 4.1 derives the unique transition kernels for this joint-scaled distribution and the corresponding optimal noise predictor, epsilon*_theta,t = epsilon_theta,t - sqrt(1 - alpha_bar_t) grad_{x_t} log E_t(x_t,y). The paper then interprets existing guidance methods as no-foresight approximations, states approximation-error bounds, and introduces rectified gradient guidance (REG), a correction term intended to make existing guidance methods approximate the optimal update more closely. Experiments on 1D and 2D synthetic problems compare REG against the exact optimal gradient, and large-scale experiments on class-conditional ImageNet and text-to-image tasks report consistent FID/IS/CLIP improvements. The joint-scaling construction is clean and self-contained, but the derivation of REG contains a chain-rule step that is not correctly stated and relies on an unproven diagonal-Jacobian approximation.
Significance. If the joint-scaling theory is accepted, the paper gives a principled resolution of known inconsistencies in guidance theory, and REG is a simple, method-agnostic plug-in that appears to improve FID and CLIP scores across several models and samplers. The 1D and 2D synthetic experiments are a genuine strength: they compare directly against the exact optimal gradient grad log E_t rather than against indirect metrics. The public code release and the parametric nature of Theorem 4.1 are additional assets. However, the theoretical grounding of REG itself is incomplete: Eq. (22) is not a correct chain-rule identity under the paper's gradient convention, the identification E_t = R_0(hat x_0) for deterministic samplers is not justified, and the diagonal-Jacobian approximation has no error analysis. These issues do not invalidate Theorem 4.1, but they substantially weaken the claim that REG is the theoretically grounded 'better approximation' advertised in the abstract; at present the method is best described as a heuristic correction with empirical support.
major comments (4)
- [Section 4, Eq. (22)] Under the column-vector gradient convention used elsewhere in the paper, e.g. Eq. (7), the chain rule for grad_{x_t} log R_0(hat x_0, y) is J^T grad_{hat x_0} log R_0(hat x_0, y), where J_{ij} = partial hat x_{0,i} / partial x_{t,j}. Eq. (22) instead writes J times the gradient. This is not a correct identity. Because a diagonal Jacobian satisfies J = J^T, the later diagonal approximation accidentally coincides with the correct formula, but the stated derivation is mathematically wrong. The authors should either correct the transpose, or explicitly reframe Eq. (22) as a heuristic construction that is only meant to motivate Eq. (21).
- [Section 4, derivation of Eq. (21)] The paper states that 'when a deterministic sampler is adopted' E_t(x_t,y) becomes R_0(hat x_0,y), with hat x_0 the one-step estimate. This equates the terminal point of the deterministic reverse trajectory with the one-step Tweedie estimate. These are not the same object: in a deterministic multi-step sampler the final x_0 is a function of the entire future trajectory, not merely of the current one-step prediction. This identification is load-bearing because it justifies replacing grad log E_t by grad log R_0(hat x_0), and no argument is given that the one-step estimate captures the full deterministic denoising map. The derivation should be repaired or the approximation stated explicitly.
- [Theorem 4.2 and Appendix D] The proof of Eq. (18) uses Eq. (33) at t = 0 to bound Terms 1 and 2. However, the paper itself notes in Appendix A that the score formula Eq. (7) is not valid at t = 0 in discrete time and that epsilon_theta(.,0,.) is meaningless; moreover sqrt(1 - alpha_bar_0) = 0 makes Eq. (33) degenerate at t = 0. Thus the bound is not proved as stated. In addition, the term 2wLt in Eq. (18) requires a Lipschitz condition in t that is not stated in Theorem 4.2. The result may be recoverable in the continuous-time limit, but the current proof has a gap.
- [Section 4, Eq. (21) and Section 5.1] The REG correction replaces a D-by-D Jacobian-vector product with an elementwise product, justified only as an 'approximately diagonal' assumption. No error bound, structural condition, or empirical characterization of this approximation is provided. The 1D experiments cannot test the assumption, since a scalar Jacobian is trivially diagonal, and the 2D win ratios in Table 1 are only 57-74%, which is not a strong validation. If the diagonal approximation can fail, REG is not a principled approximation of Eq. (22) but a heuristic gradient modification, and the paper should present it as such rather than as a theoretically grounded consequence of Eq. (17).
minor comments (4)
- [Section 4, Eq. (21)] The notation 'nabla_{x_t} log R_t(hat x_t, y)' is used, but hat x_t is not defined anywhere; presumably this should be x_t or the definition should be supplied.
- [Footnote 1] The footnote contains the typo 'another subtly' and is long enough that it should be integrated into the main text or shortened.
- [Table 3] For the 64x64 EDM2-S + AutoG row, the IS value is reported as 69.01 both without and with REG, yet the green upward arrow is shown. An unchanged metric should not be marked as an improvement.
- [Theorem 4.3 proof] The integration-by-parts argument in the proof of Theorem 4.3 assumes the boundary terms vanish; this assumption should be stated explicitly.
Circularity Check
No significant circularity: Theorem 4.1 is a direct derivation from a stated joint-scaling objective, and REG is an explicitly declared approximation rather than a renamed input.
full rationale
The paper's central theoretical result (Theorem 4.1, Eqs. 13-17) is self-contained: the scaled joint objective is stated as a goal, the auxiliary quantity E_t is defined independently as an expectation of R_0 under p_theta(x_0|x_t,y), and the unique transition kernels and noise prediction update are derived, not assumed. Eq. (17) follows by differentiating the derived marginal and applying the score formula; no fitted parameter is renamed as a prediction. The REG formula Eq. (21) is explicitly presented as a sequence of approximations to Eq. (22): the paper states it replaces the terminal gradient with an intermediate one, simplifies the Jacobian using the DDPM relation, and approximates the matrix-vector product by elementwise multiplication assuming approximate diagonality. These are declared heuristics, not concealed inputs. The 1D/2D 'better approximation' experiments compare REG against a numerically integrated nabla log E_t defined from the same reward R_0; this is an internal consistency check without circularity. ImageNet and COCO evaluations use external FID, IS, and CLIP metrics. No load-bearing step rests on self-citation. The section 'Invalid Marginal Scaling' independently proves that the older marginal-scaling objective is infeasible, so the joint-scaling premise is not smuggled in. The reviewer-flagged transpose issue in Eq. (22) is a mathematical correctness concern, not a circular step. Overall, the derivation chain is self-contained and does not reduce to its inputs by construction.
Assumptions & free parameters
assumptions (4)
- domain assumption Score function formula \nabla_{x_t} \log p_\theta(x_t|y) = -\epsilon_\theta(x_t,t,y)/\sqrt{1-\bar\alpha_t} holds for all t including t=0 (via continuous-time SDE extension).
- domain assumption The noise prediction network \epsilon_\theta is L-Lipschitz and bounded by B.
- ad hoc to paper The Jacobian matrix \partial \hat x_0/\partial x_t is approximately diagonal, allowing elementwise product in Eq. (21).
- domain assumption A deterministic sampler is used in the reverse process.
Cite this review
Pith. "Pith review of REG: Rectified Gradient Guidance for Conditional Diffusion Models." pith.science (2026). https://pith.science/paper/P7QHNUDI
@misc{pith2026250118865,
author = {Pith},
title = {Pith review of: REG: Rectified Gradient Guidance for Conditional Diffusion Models},
year = {2026},
howpublished = {\url{https://pith.science/paper/P7QHNUDI}},
note = {Machine review of arXiv:2501.18865}
}
read the original abstract
Guidance techniques are simple yet effective for improving conditional generation in diffusion models. Albeit their empirical success, the practical implementation of guidance diverges significantly from its theoretical motivation. In this paper, we reconcile this discrepancy by replacing the scaled marginal distribution target, which we prove theoretically invalid, with a valid scaled joint distribution objective. Additionally, we show that the established guidance implementations are approximations to the intractable optimal solution under no future foresight constraint. Building on these theoretical insights, we propose rectified gradient guidance (REG), a versatile enhancement designed to boost the performance of existing guidance methods. Experiments on 1D and 2D demonstrate that REG provides a better approximation to the optimal solution than prior guidance techniques, validating the proposed theoretical framework. Extensive experiments on class-conditional ImageNet and text-to-image generation tasks show that incorporating REG consistently improves FID and Inception/CLIP scores across various settings compared to its absence.
Figures
Figures from the paper (7 more)
Reference graph
Works this paper leans on
-
[1]
A noise is worth diffusion guidance
Ahn, D., Kang, J., Lee, S., Min, J., Kim, M., Jang, W., Cho, H., Paul, S., Kim, S., Cha, E., et al. A noise is worth diffusion guidance. arXiv preprint arXiv:2412.03895 ,
-
[2]
In the main text, REG is derived using the ϵ- prediction parametrization, where we simplify the Jacobian matrix by using ˆx0 = 1√ ¯αt (xt − √1 − ¯αtϵθ,t), leading to Eq. (21). For EDM2 (Karras et al., 2024b), which uses the x0 parametrization, the REG correction term trivially becomes ∂(1T · Dθ(xt, t, y))/∂xt, where Dθ represents the x0-prediction network...
work page 2022
-
[3]
Chung, H., Kim, J., Park, G. Y ., Nam, H., and Ye, J. C. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070,
-
[6]
The target conditional data distribution is 0.5 × N(0.5, 0.252) + 0.5 × N(1.5, 0.252)
Histograms of 8000 samples drawn from a trained simple diffusion model are shown for three cases: (a) without any guidance, (b) using the vanilla CFG without REG, and (c) using the vanilla CFG with our proposed REG. The target conditional data distribution is 0.5 × N(0.5, 0.252) + 0.5 × N(1.5, 0.252). While all approaches correctly recover the mean locati...
work page 2023
-
[8]
Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M
Accessed: 2025- 01-11. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125,
arXiv 2025
-
[9]
Rec- tified diffusion: Straightness is not your need in rectified flow
Wang, F.-Y ., Yang, L., Huang, Z., Wang, M., and Li, H. Rec- tified diffusion: Straightness is not your need in rectified flow. arXiv preprint arXiv:2410.07303, 2024a. Wang, X., Dufour, N., Andreou, N., Cani, M.-P., Abre- vaya, V . F., Picard, D., and Kalogeiton, V . Analysis of classifier-free guidance weight schedulers. Transactions on Machine Learning ...
-
[10]
Rectified diffusion guidance for conditional generation
Xia, M., Xue, N., Shen, Y ., Yi, R., Gong, T., and Liu, Y .-J. Rectified diffusion guidance for conditional generation. arXiv preprint arXiv:2410.18737,
-
[11]
11 REG: Rectified Gradient Guidance for Conditional Diffusion Models A. Remarks on Score Function Formula We will illustrate the score function formula for the unconditional generation case, corresponding to the second line of Eq. (7), noting that the first line of Eq. (7) associated with the conditional generation is similar. The continuous-time SDE form...
work page 2021
Show all 18 references
-
[12]
(33)) ≤ w · ||ϵθ(xt, 0, y) − ϵθ(xt, t,y)|| + w · ||ϵθ(xt, t) − ϵθ(xt, 0)|| (Triangular inequality) ≤ 2wLt
− ϵθ(xt, t,y) + ϵθ(xt, t)|| (Use Eq. (33)) ≤ w · ||ϵθ(xt, 0, y) − ϵθ(xt, t,y)|| + w · ||ϵθ(xt, t) − ϵθ(xt, 0)|| (Triangular inequality) ≤ 2wLt . (ϵθ,t is Lipschitz continuous) (34) Combining them together, we obtain: √ 1 − ¯αt||∇xt log Et(xt, y) − ∇xt log Rt(xt, y)|| ≤2wB|| dˆ...
2024
-
[13]
by using a shared-weight noise prediction network with label dropout during training, and we find no significant difference compared to our implementation in this specific case. We emphasize that this example only considers a single class with label y = 0, and, similar to (Kyn...
2024
-
[16]
Nevertheless, we use Eq
and the Euler Discrete sampler, respectively, which differ from DDPM. Nevertheless, we use Eq. (21) in these cases as well, and it performs effectively in practice. We conduct our experiments using the open-source DiT (Peebles & Xie, 2023), EDM2 (Karras et al., 2024b), and Hug...
2023
-
[17]
bad” models used in AutoG, we rely on the publicly available checkpoints shared by Karras et al. (2024a). However, when the interval range or the “bad
study provides the interval ranges for DiT and EDM2, which we adopt here for consistency. For the “bad” models used in AutoG, we rely on the publicly available checkpoints shared by Karras et al. (2024a). However, when the interval range or the “bad” model checkpoint is unavai...
2024
-
[18]
For text-to-image generation, we randomly select one caption per image from the COCO-2017 validation dataset, creating 5,000 pairs of images and captions
and EDM2 (Karras et al., 2024b). For text-to-image generation, we randomly select one caption per image from the COCO-2017 validation dataset, creating 5,000 pairs of images and captions. FID and CLIP scores are evaluated using TorchMetrics. Additional Pareto front results for...
2017
-
[2020]
and Salimans, T
Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598,
-
[2021]
L., Verbeek, J., Biroli, G., and Mezard, M
Pavasovic, K. L., Verbeek, J., Biroli, G., and Mezard, M. Un- derstanding classifier-free guidance: High-dimensional theory and non-linear generalizations. arXiv preprint arXiv:2502.07849,
-
[2022]
Guiding a diffusion model with a bad version of itself
Karras, T., Aittala, M., Kynk¨a¨anniemi, T., Lehtinen, J., Aila, T., and Laine, S. Guiding a diffusion model with a bad version of itself. In Advances in Neural Information Processing Systems,, 2024a. Karras, T., Aittala, M., Lehtinen, J., Hellsten, J., Aila, T., and Laine, S....
-
[2023]
Sdxl: Im- proving latent diffusion models for high-resolution image synthesis
Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M ¨uller, J., Penna, J., and Rombach, R. Sdxl: Im- proving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952,
-
[2024]
and Nakkiran, P
Bradley, A. and Nakkiran, P. Classifier-free guidance is a predictor-corrector. arXiv preprint arXiv:2408.09000,
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.