acceptodds
Under review as a conference paper at ICLR 2027

Fitted CFG: Stabilization Bands and Contraction Matching

Abstract

Classifier-free guidance (CFG) changes the flow that a diffusion sampler integrates, but the DDIM update was designed for the unguided flow. In an ordered, commuting Gaussian model, guidance strengthens contraction along discriminative directions, and a coarse DDIM step can reverse such a coordinate. On this model, a band of guidance coefficients prevents the reversal and controls discriminative residual amplification without prescribing the exact contraction; it contains the sign cap and a parameter range of deterministic CFG++. Terminal matching fixes the factor left free and selects fitted CFG, a one-coefficient change at no extra denoiser cost. Fitted CFG and the caps share first-order absolute accuracy; matching additionally bounds the accumulated log-relative contraction error independently of the noise horizon, for fixed guidance and maximum log step. On CIFAR-10, fitted CFG lowers FID in all cells of two nine-cell grids and in a 50k comparison (27.87 to 20.88); the sign cap matches this stabilization in paired controls. A calibrated finite-scale extension improves FID, KID and saturation on 30k SDXL samples at some cost in text alignment. The framework separates three design criteria: stability, consistency with the prescribed field, and contraction accuracy.

Primary area:
generative models
Keywords:
diffusion models, classifier-free guidance, numerical stability, exponential fitting, diffusion sampling

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_EroUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates numerical instability in classifier-free guidance (CFG) when combined with coarse DDIM sampling. The authors show that strong guidance can cause sign reversals and residual amplification along discriminative directions, potentially degrading generation quality.

The main contributions are:

  1. A theoretical analysis identifying instability thresholds and stable coefficient ranges under an ordered, commuting Gaussian model.
  2. Fitted CFG, a modified guidance coefficient derived by matching the exact contraction of a terminal Gaussian mode, without additional denoiser evaluations.
  3. A horizon-independent bound on accumulated relative contraction error under the stated Gaussian assumptions.
  4. A finite-scale crossover extension with a calibrated parameter.

Experiments on CIFAR-10 demonstrate improved FID and reduced residual amplification. On SDXL, the basic Fitted CFG method exhibits mixed results, while the calibrated crossover variant provides modest improvements in FID and KID with reduced text alignment.

Critical strengths
  1. Principled theoretical analysis of CFG instability (Sections 2-3).

The paper provides a useful mathematical explanation of why coarse DDIM steps under strong guidance can cause coordinate sign reversals and residual amplification. The distinction between stability, consistency, and contraction accuracy is meaningful, and Theorems 1-3 provide a rigorous framework for understanding these effects under the stated assumptions. This is the paper's strongest contribution.

  1. Simple implementation with carefully controlled experiments (Sections 3-4).

Fitted CFG requires only a modification of the guidance coefficient, without additional denoiser evaluations or retraining. The authors also provide substantial experimental controls, including matched seeds, different guidance strengths and sampling steps, closure comparisons, and local response analyses. The CIFAR-10 results show a notable FID improvement from 27.87 to 20.88, while the SDXL local-response experiments provide evidence of the predicted numerical instability.

Critical weaknesses
  1. Limited algorithmic novelty and unclear advantage over existing methods (Sections 3.2, 4.1, and 5).

The proposed update belongs to the existing deterministic CFG++ family and can also be expressed as standard CFG with an effective guidance coefficient. Therefore, the main algorithmic contribution is a particular coefficient selection rather than a fundamentally new sampling mechanism.

More importantly, the experiments do not convincingly demonstrate that contraction matching provides a practical advantage over simpler stabilization strategies. In Section 4.1, the sign cap achieves almost identical CIFAR-10 performance (FID 24.735 versus 24.733), and Table 4 reports nearly identical local-response improvements for the two methods.

Although the theoretical contraction-accuracy guarantee is novel and potentially valuable, its practical importance remains insufficiently demonstrated. The authors should establish settings in which contraction matching provides a measurable advantage over simpler clipping rules and stronger existing CFG-based methods.

  1. Restrictive theoretical assumptions and inconsistent empirical improvements (Sections 3-4).

The theoretical guarantees rely on ordered, commuting Gaussian covariance matrices and exact Gaussian denoisers. These assumptions are restrictive and do not directly establish stability or contraction accuracy for nonlinear learned diffusion models.

Furthermore, the empirical improvements are not consistent across evaluation metrics and models. On CIFAR-10 (Table 2), Fitted CFG improves FID but worsens KID from 0.0108 to 0.0177. On SDXL (Table 3), the basic Fitted CFG method worsens FID from 19.489 to 23.178, while also reducing CLIP alignment. The calibrated crossover extension improves FID only modestly (19.489 to 19.006), while still reducing text alignment.

These results suggest that stabilization and contraction matching do not necessarily translate into better overall generation quality. Stronger evidence of generalization and more favorable quality-alignment trade-offs would substantially strengthen the paper.

Questions for authors
  1. Given that Fitted CFG belongs to the existing CFG++ update family, what practical benefit does the proposed contraction-matching coefficient provide beyond the simpler sign cap? Can the authors identify experimental settings where the two methods differ meaningfully in generation quality or numerical accuracy?

  2. How applicable are the theoretical stability and relative-error guarantees to realistic learned diffusion models, where the Gaussian and commuting-covariance assumptions generally do not hold? Can the authors provide additional evidence connecting the theoretical contraction-matching property to empirical performance?

  3. How do the authors explain the deterioration in KID on CIFAR-10 and the substantially worse FID of basic Fitted CFG on SDXL? Can they demonstrate a consistent advantage over tuned CFG, CFG++, and sign-cap alternatives when both image quality and conditioning strength are considered, under matched evaluation settings?

3 comments

Thank you for asking where contraction matching makes a measurable difference. We added a CIFAR comparison in which the fitted and capped coefficients separate.

Q1. Benefit beyond the sign cap. The stabilization band leaves the contraction magnitude free. Terminal matching selects βfit=1−r1+w\beta_{\rm fit}=1-r^{1+w} within deterministic CFG++, and Theorem 3 bounds its accumulated log-relative error independently of the noise horizon under the stated Gaussian assumptions.

For the new CIFAR test, we chose r=0.85r=0.85 and nominal g=9g=9 before sampling. The pure-mode factors on untruncated steps are −0.35-0.35, 00 and 0.2316170.231617 for CFG, cap and fitted. We used two independent balanced 5k runs. Within each run, methods share checkpoints, initial noise, labels, 134 branch evaluations and conditional closure. Entries are run 1 / run 2; the g=2g=2 row transfers the earlier calibration unchanged to this mesh.

MethodFIDKIDTarget accuracy
CFG, g=9g=945.330 / 45.0920.026692 / 0.02681087.30% / 87.40%
Sign cap, g=9g=925.091 / 25.0330.009548 / 0.00942396.20% / 96.26%
Fitted CFG, g=9g=920.071 / 20.0860.006221 / 0.00613798.22% / 98.58%
Transferred CFG, g=2g=210.485 / 10.4260.002486 / 0.00254899.02% / 98.98%

Fitted improves all three metrics over cap in both runs. The accuracy gain is 2.17 percentage points (95% paired bootstrap interval [1.75, 2.59]); amplification p95 falls from 2.239 / 2.125 to 1.463 / 1.465. The paired guided closure preserves the FID/KID ordering. This gives an empirical reason to use the matching rule: it selects a coefficient with a Gaussian accuracy guarantee and delivers joint quality and conditioning gains over the cap in this setting, with no extra denoiser evaluations.

Q2. Applicability to learned denoisers. The denoiser-error margin can be quantified. With s=r1+ws=r^{1+w}, Dc=ecxD_c=e_cx and Du=(1+eu)xD_u=(1+e_u)x, the fitted factor is s+(1−s)ec+(s−r)eus+(1-s)e_c+(s-r)e_u. It remains positive if (1−s)∣ec∣+(r−s)∣eu∣<s(1-s)|e_c|+(r-s)|e_u|<s, so strong contraction can leave a small margin.

We also integrated 200 prespecified CIFAR cases against the same learned field. Of these, 167 pass two successive refinement checks and RK4 agreement at 10−410^{-4}; 33 remain unresolved. At the last positive-noise node, median normalized endpoint errors on the accepted cases are 0.3782 / 0.5956 / 0.6832 for CFG / cap / fitted. Fitted has smaller error than cap in 6 cases and larger error in 161.

Q3. Metric disagreement and tuned alternatives. In the original 50k CIFAR comparison, FID falls from 27.8706 to 20.8797 and KID rises from 0.010827 to 0.017734. FID measures feature means and covariances through Gaussian approximations; KID uses a cubic-kernel discrepancy. They can rank distributions differently. The new comparison above improves both metrics over cap in both runs.

On SDXL, the reported trade-off persists with FP32 persistent states. On 5,000 archived pairs at g=12,N=12g=12,N=12, crossover has FID 27.1793 versus CFG's 27.5896. The KID difference is -0.0002694 (95% interval [-0.0004034, -0.0001248]), with CLIP difference -0.0030005 ([-0.0033917, -0.0025958]). These use native closures, DcD_c for crossover and DwD_w for CFG. The cap's KID/CLIP advantage over fitted also persists in the first 1,000 pairs under stable arithmetic.

We evaluated CFG, fitted, cap, crossover, CFG++ and DPM-Solver++ on 256 calibration pairs at N=12,30N=12,30, selecting by KID subject to CLIP being at least the CFG g=7.5g=7.5 value minus 0.002. All six families have an eligible point at N=30N=30; at N=12N=12, only CFG and DPM-Solver++ meet the floor. Held-out evaluation of these selected settings is pending.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts, particularly the new results demonstrating an advantage over the sign cap. However, some concerns remain:

  1. The simpler CFG baseline at g=2 still achieves substantially better results.
  2. Fitted CFG shows larger endpoint errors than the sign cap in 161 of 167 cases, raising questions about the practical relevance of its theoretical guarantees.
  3. The SDXL trade-offs remain, and the held-out comparisons with stronger baselines are still pending.

I appreciate the additional evidence and will carefully revisit the rebuttal and results to reconsider whether my original assessment and score should change.

Thank you for the thoughtful follow-up. We have now completed the further experiments. They show improved generation metrics at the CIFAR g=2g=2 setting, better terminal diagnostics on the same endpoint-error cases, and a reproducible SDXL operating point against stronger baselines at N=30N=30.

1. Comparison with CFG at g=2g=2. We compared both schemes at g=2g=2 on two fresh, independent 5k blocks, with paired seeds, the same mesh, 134 denoiser-branch evaluations, and matched DcD_c closure.

MethodggFID, run 1FID, run 2Pooled KIDPooled target accuracy
CFG210.34910.6840.00257799.26%
Fitted29.69710.0320.00226899.15%

Fitted improves FID and KID in both runs. Its pooled KID difference is −3.09×10−4-3.09\times10^{-4} (95% paired/shared-reference interval [−3.66,−2.50]×10−4[-3.66,-2.50]\times10^{-4}). The target-accuracy difference is −0.11-0.11 percentage points (95% paired interval [−0.25,0.03][-0.25,0.03]). At g=2g=2, fitted lowers FID and KID in both held-out runs, with 99.15% target accuracy and unchanged sampling cost.

2. Practical relevance on the endpoint-error cases. Theorem 3 motivates matching through relative-contraction accuracy in the ordered Gaussian model. To examine practical usefulness directly, we evaluated terminal outputs on the same 167 accepted cases underlying your endpoint-error comparison, at g=9g=9 with matched DcD_c closure:

EndpointTarget accuracyMedian pre-clamp out-of-range fraction
Converged reference81.44%37.43%
Sign cap94.61%10.81%
Fitted98.80%2.31%

The out-of-range fraction counts scalar image values outside [−1,1][-1,1]. Both reference integrators agree on all 167 target-classification outcomes. On these same cases, fitted combines the larger endpoint distances you noted with higher target accuracy and substantially less out-of-range output. Distance to the learned-field endpoint and terminal performance therefore rank the methods differently.

3. Completed SDXL comparisons with stronger baselines. At N=30N=30, where all six methods met the calibration CLIP floor, we evaluated each on two independent 5k blocks. Parameters were fixed using separate calibration data, with an equal five-candidate budget and KID selection subject to the CLIP floor. Methods share prompts, seeds, arithmetic and 60 denoiser-branch evaluations. Pooled 10k results:

MethodParameterFIDKIDCLIP
CFGg=7.5g=7.522.0330.0066770.31828
Fittedg=12g=1221.6530.0063410.31726
Sign capg=7.5g=7.522.0120.0065230.31800
Crossoverg=7.5g=7.521.7820.0066760.31769
CFG++β=1\beta=122.2800.0065840.31835
DPM-Solver++g=7.5g=7.521.6330.0065760.31741

The table uses native closures: DwD_w for CFG/DPM-Solver++ and DcD_c for the others. Fitted improves FID/KID over CFG and the sign cap in both blocks; these improvements also hold with matched DcD_c closure. Against CFG, its KID difference is −3.36×10−4-3.36\times10^{-4} (pointwise 95% interval [−4.25,−2.50]×10−4[-4.25,-2.50]\times10^{-4}), and CLIP changes by −0.001016-0.001016 ([−0.001224,−0.000797][-0.001224,-0.000797]), within the prespecified 0.002 tolerance. Fitted achieves the lowest point KID among the six selected settings and FID close to DPM-Solver++. This provides held-out evidence for fitted's practical use in a latent diffusion model at the 30-node budget.

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

This paper studies the mismatch between classifier-free guidance (CFG) and coarse deterministic DDIM updates. For exact Gaussian denoisers with ordered, commuting covariances, it shows that guidance can cause coordinate reversal and normalized-residual amplification at distinct step-size thresholds. It derives stabilization bands containing simple coefficient caps, then selects fitted CFG by matching the terminal contraction factor without additional denoiser evaluations. The main theoretical contribution is a horizon-independent bound on accumulated log-relative contraction error, distinguishing fitted CFG from caps that share first-order absolute accuracy but can accumulate unbounded relative distortion. The paper also develops a finite-scale crossover extension and relates its updates to deterministic CFG++ and effective guidance schedules.

Validation includes exact Gaussian calculations, CIFAR-10 and SDXL generation experiments, and local-response comparisons against refined integration of the same learned guided field. Paired coefficient controls and terminal-closure ablations isolate shared stabilization effects. On CIFAR-10, fitted CFG reduces FID and residual amplification, with benefits largely matched by the sign cap and mixed results on other metrics. On SDXL, calibrated crossover fitting improves FID, KID, and saturation at some cost in text alignment, while the local-response studies distinguish response improvements from endpoint accuracy.

Critical strengths
  • The paper clearly separates coordinate sign preservation, normalized-residual control, absolute endpoint accuracy, and relative contraction accuracy. In particular, it demonstrates that clipping-based methods can remain stable and achieve first-order absolute accuracy while accumulating unbounded log-relative distortion. This provides a useful criterion for evaluating guidance-aware discretizations beyond stability alone. (§§2–3; Appendix B.3)

  • Theorem 3 is the strongest technical contribution, I would say. For exact Gaussian denoisers with ordered, commuting covariances, it bounds accumulated log-relative contraction error independently of the noise horizon, covariance eigenvalues, and dimension, at fixed guidance and maximum log step. The proof accommodates nonuniform meshes and controls accumulated error through integrable crossover profiles. The central derivations hold under independent reconstruction, and the Gaussian calculations in Tables 5 and 6 reproduce the reported precision. (§3.3; Appendix B.2)

  • Fitted CFG follows directly from matching the terminal contraction factor, with an additional semigroup characterization explaining its uniqueness within the specified scalar update family. It requires no extra denoiser evaluations or covariance estimation. The explicit connections to deterministic CFG++ and effective guidance schedules clarify both its implementation and its relationship to existing methods. (§3.2; Appendix A)

  • The paper includes independent CIFAR-10 grid replications, a separate 50k-sample comparison, sign-cap and calibrated lower-guidance controls, and a factorial experiment separating finite-step updates from terminal closure. These controls substantiate the shared stabilization effect and address important alternative explanations for the observed gains. (§4.1; Appendices D.5 and D.8)

  • The authors disclose metric trade-offs, near equivalence with the sign cap, unresolved references, and unsuccessful experimental criteria. They also distinguish exact-model guarantees from learned-model observations and document floating-point cancellation and a factored implementation. This transparency makes the scientific claims easier to assess and the experiments easier to reproduce. (§6; Appendices A–E)

Critical weaknesses
  • Fitted CFG is a particular mesh-dependent parameter choice within the deterministic CFG++ family and is exactly equivalent to ordinary DDIM+CFG with an effective guidance schedule. The relative-error analysis also extends the crossover-based argument acknowledged in reference [19]. Consequently, the contribution rests primarily on the guided theoretical extension and its explanatory value; the coefficient formula alone provides limited methodological novelty. (§3.2–3.3; §5)

  • The sign cap essentially matches the fitted CFG in the main CIFAR control, with FID 24.735 versus 24.733, and both repair 11/12 verified SDXL response reversals. Moreover, the separately calibrated g=2g=2 CFG control achieves better FID, KID, and target accuracy under conditional closure. These results support shared stabilization and guidance attenuation, but leave the practical reason to prefer exact fitting unresolved. (§4.1; Tables 4 and 20)

  • The guarantees require exact denoisers with ordered, commuting Gaussian covariances. The paper itself shows that finite-label conditioning can have a different terminal exponent, while all verified SDXL reversals occur at early finite-noise steps. No perturbation analysis establishes robustness to departures from the Gaussian assumptions. The crossover variant’s failure to repair any of the twelve verified reversals further separates its generation benefits from the tested repair mechanism. (§4.3; Appendices B.4 and C.1)

  • The 50k CIFAR FID improvement accompanies worse KID, from 0.0108 to 0.0177. Terminal fitting worsens SDXL FID, KID, and CLIP, while calibrated crossover fitting yields modest distribution-metric gains at the cost of alignment. Those modest gains lack experiment-level uncertainty estimates. A matched quality–conditioning comparison across tuned guidance choices, together with an equal-budget guided-solver baseline, would better establish practical significance. (Tables 2–3, 16, and 22)

  • Only 64 of 96 selected cases have accepted numerical references; thirteen selected robust coarse-negative cases remain unresolved. Bootstrap intervals characterize the selected population but do not resolve this attrition. The independent unscreened study strengthens the response-error finding, but fitted CFG increases center-state error in all sixteen cases. Thus, response improvements cannot be interpreted as general improvements in trajectory accuracy. (Appendix C; Tables 7 and 11–13)

  • The paper documents a cancellation that can erase representable fitted contraction factors and provides a factored remedy. However, the main generation studies use the additive formulation, including FP16 persistent states in the large SDXL experiments. A representative paired comparison using factored updates and persistent FP32 states is needed to determine whether the reported executions retain the numerical property motivating the method. (Appendix A; Table 25)

Questions for authors
  1. What is the essential theoretical advance over the closest prior work? Given the exact connection to deterministic CFG++ and the acknowledged extension of reference [19], could you isolate the new argument required for the guided horizon-independent relative-error bound? Which aspects of the result cannot be obtained directly from the existing unguided analysis? (§§3.2–3.3)

  2. Can exact matching provide a measurable benefit beyond sign capping? The sign cap nearly matches fitted CFG in the CIFAR controls and repairs the same eleven SDXL events. Do your existing records contain a setting where the coefficients differ materially and fitted CFG improves contraction magnitude or endpoint accuracy? A targeted comparison with matched closure, evaluation count, and a converged reference would help establish the practical value of matching. (Tables 4 and 20)

  3. How robust is the stabilization mechanism to denoiser error? On the pure test mode, perturbing Du=xD_u=x to Du=(1+ϵ)xD_u=(1+\epsilon)x changes the fitted factor to r1+w+(r1+w−r)ϵr^{1+w}+(r^{1+w}-r)\epsilon, which can become negative when the intended contraction is small. Can you quantify this sensitivity or provide learned-model diagnostics indicating whether the relevant stability margins are sufficient? (Theorem 2; Appendix B)

  4. How much does the stable implementation change the reported results? On a representative paired subset, how do the principal generation comparisons change when using the factored update with persistent FP32 states? In particular, are the fitted-versus-cap differences and crossover gains robust to this implementation change, and how frequently does the original arithmetic lose the small fitted contribution? (Appendix A; Table 25)

  5. Does crossover fitting improve the quality–conditioning trade-off? At comparable CLIP or GenEval2 alignment and equal denoiser budget, does the calibrated crossover coefficient retain its FID/KID or saturation advantage over tuned ordinary CFG and caps? Could you provide uncertainty for the modest distribution-metric gains beyond the reported KID subset-resampling ranges? (Table 3; Appendix D.7)

  6. How sensitive are the response conclusions to unresolved references? Thirteen selected robust coarse-negative cases remain unresolved. Can additional refinement resolve a representative subset, or can you otherwise bound how their outcomes would affect the repair comparison? This would clarify how far the results extend beyond the twelve verified events. (Appendix C.1–C.3)

2 comments

Thank you for the detailed review. We have completed the coefficient-separation and precision tests you suggested.

Q1. Guided theoretical advance. Terminal matching selects βfit=1−r1+w\beta_{\rm fit}=1-r^{1+w} within deterministic CFG++. The new proof step extends [19] by controlling error across two covariance scales. For g=1+wg=1+w, A=σ2/(a+σ2)A=\sigma^2/(a+\sigma^2) and B=σ2/(b+σ2)B=\sigma^2/(b+\sigma^2), with 0≤a≤b0\le a\le b, the fitted factor is

\kappa\sim(1-A)\delta_0+B\delta_1+(A-B)\delta_g,$$ where $\delta_t$ is a point mass at $t$ and $h=-\log r$. The mean is $gA-wB$, and $$\operatorname{Var}(\kappa)=g^2A(1-A)+w^2B(1-B)-2gwB(1-A) \le g^2A(1-A)+w^2B(1-B).$$ Exponential tilting bounds the frozen-exponent defect by $h^2e^{gh}\operatorname{Var}(\kappa)/2$. Each crossover profile has log-noise integral at most $1/2$; bounded variation controls the changing exponent on nonuniform meshes. Together these give the horizon-independent $O_w(H)$ log-relative bound, $H=\max_n h_n$. The unguided proof controls a single profile. Guidance introduces the second profile and the changing mean $gA-wB$, requiring both estimates. The resulting Gaussian bound controls relative contraction even as the noise floor tends to zero. Diagonalization and coupling are standard reductions. **Q2. Matching versus capping.** We selected $r=0.85$, nominal $g=9$, common conditional closure and 134 branch evaluations before sampling. The pure-mode factors on untruncated steps are $-0.35$, $0$ and $0.231617$ for CFG, cap and fitted. Two independent balanced 5k runs, paired across methods, give: | Method | FID, run 1 / 2 | KID, run 1 / 2 | |---|---:|---:| | CFG, $g=9$ | 45.330 / 45.092 | 0.026692 / 0.026810 | | Cap, $g=9$ | 25.091 / 25.033 | 0.009548 / 0.009423 | | Fitted, $g=9$ | 20.071 / 20.086 | 0.006221 / 0.006137 | | Transferred CFG, $g=2$ | 10.485 / 10.426 | 0.002486 / 0.002548 | Fitted gains 2.17 accuracy percentage points over the cap (95% paired bootstrap interval [1.75, 2.59]); the guided closure preserves the FID/KID ordering. The coefficient choice now makes a measurable difference to generation at the same nominal guidance and denoiser cost. The $g=2$ row transfers the earlier calibration unchanged. Of 200 prespecified same-field references, 167 pass two successive Heun refinements and RK4 agreement at $10^{-4}$; 33 remain unresolved. At the last positive-noise node, median normalized endpoint errors are 0.3782 / 0.5956 / 0.6832 for CFG / cap / fitted on the accepted cases. Fitted has smaller error than cap in 6 cases and larger error in 161. The error differences exceed twice the reference disagreement in magnitude, and RK4 preserves all 167 orderings. The paired median difference is +0.0611 (95% interval [0.0554, 0.0696]). **Q3. Denoiser sensitivity.** With $s=r^{1+w}$, $D_c=e_cx$ and $D_u=(1+e_u)x$, the fitted factor is $$F=s+(1-s)e_c+(s-r)e_u.$$ The sufficient positivity condition is $(1-s)|e_c|+(r-s)|e_u|<s$. For your perturbation, $e_c=0$, $e_u\ge0$ and $w>0$, the exact threshold is $e_u<s/(r-s)$. Stronger intended contraction leaves a smaller margin for denoiser error. **Q4. Stable arithmetic.** On 5,000 archived SDXL pairs at $g=12,N=12$, switching from FP16 to FP32 persistent states changes CFG FID from 27.5210 to 27.5896 and crossover FID from 27.1432 to 27.1793. The KID/CLIP change intervals include zero. The cap's KID/CLIP advantage also persists on the first 1,000 pairs with factored fitted updates. In 488 cases with matched diagnostics, among components with nonzero FP64 fitted-cap contrast, computed updates coincide in 59.59%, 9.11% and 6.65% under legacy, additive FP32 and factored FP32 arithmetic, respectively. These rates are measured along each execution's own trajectory. **Q5. Quality and conditioning.** In the FP32-state 5k comparison, crossover-minus-CFG KID is $-0.0002694$ (95% interval $[-0.0004034,-0.0001248]$), with CLIP difference $-0.0030005$ ($[-0.0033917,-0.0025958]$). KID uses a paired two-sample influence-multiplier interval including the shared 10k real bank; CLIP uses paired bootstrap. This uses native closures ($D_w$ for CFG, $D_c$ for crossover) and remains a quality/alignment trade-off. We also tested 60 calibration configurations on 256 paired cases across six families, including CFG++ and DPM-Solver++, at $N=12,30$. Selection minimizes KID with CLIP at least the CFG $g=7.5$ value minus 0.002. CFG and DPM-Solver++ meet this floor at both budgets; fitted, cap, crossover and constant-parameter CFG++ do so at $N=30$. Held-out evaluation of the selected settings is pending. **Q6. Unresolved references.** Both fitted and cap exceed the positive-response threshold on all 13 unresolved coarse-negative cases. If $k$ meet the verified-reversal criteria, both repair rates become $(11+k)/(12+k)$, $0\le k\le13$. Their tie therefore holds for every possible $k$. The verified result remains 11/12.

Thank you again for your detailed questions. Our response clarifies the new guided-error argument, quantifies denoiser sensitivity, and reports the paired FP32/factored-update checks. The held-out SDXL comparisons are now complete: our follow-up to EroU includes the N=30N=30 stronger-baseline results with conditioning measurements and uncertainty, alongside terminal diagnostics on the same 167 reference cases. We would welcome your feedback on these clarifications and additional experiments.

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

This paper studies the mismatch between classifier-free guidance (CFG) and coarse DDIM updates. Using an ordered Gaussian model with commuting covariances, the authors characterize a stabilization band that prevents coordinate reversal and bounds normalized-residual growth on discriminative coordinates with zero conditional variance. They propose “fitted CFG” by matching the exact contraction factor of the pure terminal mode and prove a horizon-independent upper bound on accumulated log-relative contraction error. Experiments on CIFAR-10 and SDXL show stabilization benefits shared with the sign cap. A calibrated finite-scale extension improves SDXL FID and KID scores at some cost to text alignment.

Critical strengths
  1. Clear distinction between stabilization and contraction accuracy. In the ordered Gaussian model, fitted CFG and simple coefficient caps provide stability guarantees and both achieve first-order absolute accuracy. Unlike the caps, fitted CFG also has a horizon-independent upper bound on accumulated log-relative contraction error. This gives a concrete example of why preventing unstable updates does not necessarily ensure accurate contraction.

  2. Meaningful experimental controls. On CIFAR-10, the authors separate finite-step effects from the final closure. On SDXL, they compare coarse updates with refined integration of the same guided field and check sensitivity to numerical precision, perturbation size, and reference integration. Within the stated tolerances, these checks provide evidence of response-sign discrepancies between coarse updates and refined integration. They support the discretization interpretation without certifying the exact-flow error.

Critical weaknesses
  1. The practical benefit of contraction matching beyond simpler guidance adjustments remains unclear. In Table 20, fitted CFG and the sign cap have nearly identical generation metrics, but the authors also explain that their coefficients are very close on this mesh, limiting this comparison’s ability to distinguish matching from capping. More importantly, under the same conditional closure, calibrated CFG at g = 2 has lower point FID/KID and higher classification accuracy than fitted CFG at nominal g = 9. These methods target different guided fields, so this is not a comparison of integration accuracy for a fixed field, but it limits the evidence for a practical generation advantage. In the independent SDXL study, fitted CFG also improves response-vector accuracy over CFG while increasing center-state error in all sixteen cases (Tables 12–13). The theoretical guarantee is valuable, but its additional practical benefit is not established by these results.

  2. Limited evaluation across SDXL sampling settings. The main SDXL comparison of fitted CFG and CFG focuses on one high-guidance, low-step setting, g = 12, N = 12. Here, fitted CFG reduces the reported saturation metric but has worse FID/KID and lower CLIP alignment. The additional setting at g = 5, N = 30 shows small, mixed differences in FID/KID, while CLIP and saturation are unreported (Table 15). Since the proposed coefficients depend directly on guidance strength and step size, broader coverage would help establish when fitted CFG offers a useful trade-off. The calibrated finite-scale extension improves FID/KID in the reported SDXL comparisons, but those generation evaluations also use g = 12, N = 12 (Tables 3, 18, 22).

Questions for authors

Beyond the Gaussian guarantees, what reported empirical evidence shows an additional benefit of terminal contraction matching over the sign cap or calibrated lower-guidance CFG?

2 comments

Thank you for suggesting a comparison where the fitted and capped coefficients separate!

Benefit over the sign cap. In two independent CIFAR-10 runs, fitted improves FID, KID and target accuracy over the cap at nominal g=1+w=9g=1+w=9. We chose a mesh with ratio r=0.85r=0.85 from the pure-mode factors before sampling: −0.35-0.35 for CFG, 00 for the cap and 0.2316170.231617 for fitted. The last positive interval is truncated to the noise floor. Each run has 5,000 class-balanced cases, paired across methods, with identical checkpoints, 134 branch-equivalent denoiser evaluations and conditional closure.

Entries are run 1 / run 2. The g=2g=2 row transfers the earlier calibration unchanged to this mesh.

MethodFIDKIDTarget accuracy
CFG, g=9g=945.330 / 45.0920.026692 / 0.02681087.30% / 87.40%
Sign cap, g=9g=925.091 / 25.0330.009548 / 0.00942396.20% / 96.26%
Fitted CFG, g=9g=920.071 / 20.0860.006221 / 0.00613798.22% / 98.58%
Transferred CFG, g=2g=210.485 / 10.4260.002486 / 0.00254899.02% / 98.98%

Fitted gains 2.17 accuracy percentage points over the cap (95% class/block-stratified paired bootstrap interval [1.75, 2.59]). Residual-amplification p95 falls from 2.239 / 2.125 to 1.463 / 1.465. Using the paired guided closure preserves the FID/KID ordering. This answers the practical question left open by Table 20: with clearly separated coefficients, the matching rule improves generation quality over the cap at unchanged denoiser cost.

Same-field endpoint accuracy gives a different ordering: 167 of 200 references pass the 10−410^{-4} checks (33 unresolved), with median normalized endpoint error 0.5956 for cap and 0.6832 for fitted at the last positive-noise node. The gain above concerns generation quality.

SDXL calibration. We evaluated six families (CFG, fitted, cap, crossover, CFG++ and DPM-Solver++) on 256 prompt/seed pairs, with five parameter choices at each of two budgets, N=12,30N=12,30. Selection minimizes KID subject to mean CLIP being at least the CFG g=7.5g=7.5 value minus 0.002. At N=30N=30, all six families have an eligible point: fitted selects g=12g=12 with CLIP 0.320915 and KID 0.005637, versus 0.322251 and 0.005887 for selected CFG. At N=12N=12, only CFG and DPM-Solver++ meet the floor. These comparisons use native closures; held-out evaluation of the selected settings is pending.

Thank the authors for the responses. After reading the responses and other reviews, I prefer to maintain the original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers and Area Chair for their time and constructive feedback. We have revised the manuscript to incorporate the completed rebuttal experiments, with the main additions summarized below.

1. CIFAR comparisons with the sign cap and guidance-matched CFG (Section 4.1, Table 3). On the prespecified mesh that separates the fitted and sign-cap coefficients, fitted CFG improves FID, KID and target accuracy in both independent 5k blocks, with a pooled accuracy gain of 2.17 percentage points. A separate comparison at the same g=2g=2 improves FID from 10.349 to 9.697 and 10.684 to 10.032 across two fresh 5k blocks. KID also improves in both blocks, with pooled target accuracy of 99.15% versus 99.26%, at matched sampling cost and closure.

2. Terminal diagnostics on the endpoint-error cohort (Section 4.1, Appendix F.2). On the same 167 accepted reference cases, fitted CFG achieves 98.80% target accuracy versus 94.61% for the sign cap, while reducing median pre-clamp clipping from 10.81% to 2.31%. These measurements clarify why learned-field endpoint distance and terminal performance can rank the methods differently. We also report the reference-convergence checks and refinement results.

3. Completed SDXL comparisons with stronger baselines (Section 4.2, Table 4). At N=30N=30, six methods receive equal calibration and denoiser-evaluation budgets, followed by two independent held-out 5k blocks. Fitted CFG improves FID/KID over CFG and the sign cap in both blocks, achieves the lowest point KID among the six selected settings, and has FID close to DPM-Solver++. Its mean CLIP decrease relative to CFG is 0.001016, within the prespecified 0.002 tolerance. The FID/KID improvements over CFG and the sign cap also hold under matched conditional closure.

Appendix F provides the calibration procedures, both SDXL budgets, additional controls and statistical details.

We appreciate your consideration of the revised manuscript and these completed results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.