acceptodds
Under review as a conference paper at ICLR 2027

Patch the Path: Fixed-Budget Post-Sampling Trajectory Upgrade for Flow Matching

Abstract

We study how to improve a flow-matching sample after sampling is complete, using a small, fixed budget of additional model evaluations. We propose FixUp-FM, a training-free method that revisits the completed Euler trajectory and upgrades steps with large estimated local errors to Heun. A minimum-norm velocity adjustment derived from the trajectory propagates these corrections downstream without additional model evaluations, reserving computation for selected Heun corrections. FixUp-FM also supports incremental refinement across successive requests, with each pass building on the latest corrected trajectory rather than restarting. Under smoothness and bounded-propagation assumptions in the fixed-budget setting, we establish near-oracle defect selection and second-order endpoint agreement with a selective Euler–Heun rerun using exact velocities. Across SD3 Medium, SD3.5 Large, and FLUX.1-dev, FixUp-FM achieves higher mean ImageReward than NFE-matched direct Euler in 111 of 120 tested correction-budget configurations across the main sampling range.

Primary area:
generative models
Keywords:
flow matching, text-to-image generation, post-sampling correction, trajectory refinement, inference-time computation, numerical analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DEDHOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addersses the problem of how to further improve the same image with a small amount of additional computation after sampling has finished. It proposes to select the steps with the largest errors on a completed Euler trajectory, upgrade them to Heun, and propagate the corrections to subsequent steps using a linear approximation derived from the trajectory. On SD3, SD3.5, and FLUX, it shows that under the same total compute, ImageReward is generally higher than simply taking more Euler steps. However, it lacks comparisons with multistep solvers, and the theoretical goal of getting close to the exact solution is not directly verified.

Critical strengths

(1) The problem is clearly defined and the method is cleverly designed: the correction direction coincides with a direction that can be estimated directly from the trajectory, so propagating correction to subsequent steps requires no additional model evaluations, and the first Heun correction also reuses an existing evaluation.

(2) The experiments are relatively large-scale, which gives the results some credibility.

Critical weaknesses

(1) A large single-step error does not necessarily mean a large impact on the final image. An error at an early step may be amplified or smoothed out by many subsequent steps, so errors of the same size can have different consequences at different times. FixUp ranks steps only by single-step error and does not consider how sensitive the endpoint is to error propagation.

(2) Where the error is large, Heun does not necessarily correct it well. FixUp may well select the intervals where Heun is least reliable, and further experiments are needed to rule this out.

(3) The theory and the experiments do not match. The theory guarantees a solution closer to the exact ODE solution, but the experiments only report ImageReward and CLIP, without the error with respect to a high-accuracy reference solution.

(4) How common is the premise of “run Euler first and fix it afterwards” in practical pipelines? If a better sampler is used from the start, is this method still needed? Hence, the most direct baselines are missing. The improvement obtained without any additional model evaluations indicates that part of the gain comes from recombining existing velocities, which multistep solvers such as DPM-Solver++(2M) and UniPC already do at zero cost.

Questions for authors

Pls see weaknesses

8 comments

Thank you for the thoughtful review and for recognizing our correction mechanism and large-scale evaluation.

Does local-error ranking help control endpoint error?

Our top-KK rule exactly minimizes the remaining-defect term in the terminal-error bound of Theorem 3, Eq. (14). The bound accounts for downstream amplification and includes correction-related remainders.

We additionally evaluated two endpoint-aware alternatives. Terminal-norm propagates each local defect to the endpoint using the cached linear propagation model and ranks the resulting magnitudes. Greedy-terminal selects intervals that most reduce the squared norm of the remaining aggregate endpoint defect.

On SD3 Medium with 512 prompts, a 21-step trajectory and three corrections, all methods use 23 total NFEs.

Selection methodImageReward ↑
FixUp0.74842
Terminal-norm0.74713
Greedy-terminal0.74713

Our score achieves the highest mean ImageReward with the simplest selection rule.

Is Heun reliable on the selected high-error intervals?

We tested 64 SD3 prompts. Euler and Heun start from the same left-end state and take one step across each selected interval. We measure local MSE against an RK4 reference computed with 32 substeps.

Heun improves local accuracy on 317 of 320 selected intervals, reducing mean local MSE by 79.2%, 85.1%, and 87.3% at base steps 12, 20, and 32, respectively.

Base stepsEuler local MSE ↓Heun local MSE ↓MSE reductionIntervals improved by Heun
120.0049980.00103879.2%63 / 64
200.00083820.000125085.1%127 / 128
320.00025430.0000323987.3%127 / 128

Repeating the comparison with 8- and 16-substep RK4 references gives exactly the same interval-wise winners. These checks support applying Heun precisely where FixUp selects corrections, addressing the concern that selection targets unreliable intervals. We have included the full results and reference-refinement checks in Appendix N (Table 15).

What does the theory guarantee, and does reference error improve?

FixUp is a post-sampling method, not a standalone solver. Corollary 1 and Eq. (16) bound the gap to an ideal selective rerun; they do not establish a smaller error relative to the exact ODE solution than a fresh NFE-matched solver. The ideal rerun uses the same grid and selected corrections but re-evaluates the true velocity at every step. Under the assumptions and a fixed budget, the gap is second order in the maximum step size, preserving the rerun's leading first-order accuracy while avoiding downstream model calls. We apologize for the ambiguity and have clarified the comparison target after Eq. (16).

To assess reference accuracy directly, we measured latent MSE against a 1,000-step Euler reference on 1,024 SD3 prompts, using the three main-paper operating points.

a=12a = 12a=20a = 20a=32a = 32
Original Euler sample0.4478(0)0.4478^{(0)}0.3511(0)0.3511^{(0)}0.2299(0)0.2299^{(0)}
FixUp-FM0.4391(+0)\mathbf{0.4391}^{(+0)}0.3459(+1)\mathbf{0.3459}^{(+1)}0.2275(+1)\mathbf{0.2275}^{(+1)}
Restart0.4587(+1)0.4587^{(+1)}0.3612(+1)0.3612^{(+1)}0.2392(+1)0.2392^{(+1)}
SDEdit0.4682(+1)0.4682^{(+1)}0.3706(+1)0.3706^{(+1)}0.2487(+1)0.2487^{(+1)}
Fresh NFE-matched Euler0.4478(+0)0.4478^{(+0)}0.3399(+1)\mathbf{0.3399}^{(+1)}0.2205(+1)\mathbf{0.2205}^{(+1)}

Entries are latent MSE (↓). Superscripts denote total NFE minus aa; fresh Euler is rerun from noise.

FixUp is the only evaluated post-sampling method that reduces reference MSE at all three points, while remaining close to fresh NFE-matched Euler. A practical distinction is that FixUp achieves this accuracy without restarting sampling. Improving an existing 20-step sample takes one additional evaluation, whereas fresh 21-step Euler requires 21 new evaluations.

Is Euler-first refinement useful when stronger samplers are available?

FixUp is an optional post-sampling method, not a standalone solver, allowing users to improve a sample they already have. We develop this approach on Euler trajectories for their simplicity and broad applicability, with practical gains measured from the existing aa-step sample.

Our contribution is to turn a completed trajectory into a reusable resource, revisiting selected intervals and propagating corrections to the output without re-evaluating every downstream velocity. We hope this trajectory-reuse perspective can inspire future work on post-sampling extensions to higher-order samplers, for example by using cached Runge–Kutta stages and embedded defect estimates to guide retrospective corrections.

We additionally evaluated fresh Euler, Heun, DPM-Solver++(2M), and UniPC on FLUX.1-dev with 1,024 prompts, using the three main-paper operating points and matched initial noise. FixUp remains competitive with strong fresh solvers.

a=12a = 12a=20a = 20a=32a = 32
FixUp-FM0.8781(+2)0.8781^{(+2)}0.9072(+3)\mathbf{0.9072}^{(+3)}0.9151(+7)\mathbf{0.9151}^{(+7)}
Euler0.8768(+2)0.8768^{(+2)}0.8812(+3)0.8812^{(+3)}0.9007(+7)0.9007^{(+7)}
Heun0.8566(+3)0.8566^{(+3)}0.8859(+3)0.8859^{(+3)}0.9050(+7)0.9050^{(+7)}
DPM-Solver++(2M)0.9040(+2)\mathbf{0.9040}^{(+2)}0.8997(+3)0.8997^{(+3)}0.8980(+7)0.8980^{(+7)}
UniPC0.9002(+2)0.9002^{(+2)}0.9003(+3)0.9003^{(+3)}0.8898(+7)0.8898^{(+7)}

Entries are ImageReward. Superscripts denote total NFE minus aa; fresh solvers are rerun from noise. Heun uses 15 NFEs in the first column to accommodate its two-stage update.

FixUp achieves the highest mean ImageReward among all compared methods at both a=20a = 20 and a=32a = 32, and exceeds fresh Euler and Heun at all three operating points. Users can thus choose to improve a sample they already have with a small additional budget, without committing to a full sampling rerun.

We hope these clarifications and additional experiments address your concerns.

Thanks for the rebuttal and additional experiments. I still have a few concerns:

(1) Evaluation: The results rely heavily on ImageReward. I also ran tests using Diffusers on FLUX.1-dev: at 10 NFE, Euler achieved 0.63 GenEval, compared with 0.58 for Heun (0.62 with an Euler ending). Although this is not a direct test of FixUp, it suggests that higher-order corrections do not necessarily improve compositional quality. A GenEval evaluation would be helpful.

(2) Reference accuracy: FLUX.1-dev typically use around 28–50 sampling steps, as in the official Diffusers documentation and published studies. Some studies also report no quality improvement beyond 50 steps. The authors use 1000-step Euler as a reference without showing convergence. Does lower MSE against this reference actually correlate with better image quality?

(3) Generality: The method mainly focsues on Euler trajectories. Its effectiveness across different schedules, few-step budgets, and stronger solvers remains unclear.

Overall, I'm still not fully convinced about the practical benefits and generality. I'll keep my original score for now.

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

FixUp-FM improves a finished Euler sample from a flow-matching model with a few extra model calls. It selects the steps with the largest estimated errors, redo them with Heun, and adjusts later steps without new model calls. The first fix usually reuses a saved value, so K fixes typically cost K-1 calls. The paper provides numerical guarantees reports higher ImageReward than NFE-matched Euler in 111 of 120 settings on SD3, SD3.5, and FLUX

Critical strengths
  1. The proposed method is practical and nearly free. On SD3/SD3.5 at a=12, one fix with zero extra calls gains nearly as much as four extra Euler steps (Table 1), with under 0.7% aded latency. The method needs no training and plugs into existing Euler pipelines.
  2. The saved trajectory reveals the direction in which the first fix moves the sample, making the next-step adjustment exact for locally affine fields (Theorem 2). Ablations show that both step selection and the adjustment improve ImageReward significantly.
Critical weaknesses
  1. The theory concerns numerical error, but on real models the paper reports only ImageReward and CLIP, not error against a high-accuracy reference. These scores may not track accuracy -- FLUX Euler stops improving beyond ~29 NFE (Table 1). In addition, the main text shows correction-count sweeps only for a>=24 (Fig. 4) and does not explain the 9 losing settings; the appendix shows that at a=8, adding fixes beyond the first falls below plain Euler (Tables 5-6).

  2. Multistep samplers (DPM-Solver++ (2M) [1], UniPC [2]) reach second-order accuracy at no extra cost and are the natural alternative to the f free first fix, but they are not compared.

[1] Lu et al., Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2022 [2] Zhao et al., UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models, 2023

Questions for authors
  1. On real models, how far is the output from a high-accuracy reference (e.g., 1000 steps), compared with NFE-matched Euler?
  2. How does FixUP-FM compare with DPM-Solver++(2M) and UniPC at equal total NFE?
  3. Where are the 9 losing settings, and why do extra fixes hurt at small a?
  4. On FLUX, why do repeated small requests outperform a single request of the same total size (Fig. 5)?
4 comments

Thank you for the thoughtful review and for recognizing the practical value and trajectory-based correction mechanism of FixUp-FM.

What does the theory guarantee, and does reference error improve? (Weakness 1 and Question 1)

FixUp is an optional post-sampling refinement method, not a standalone solver. Corollary 1 and Eq. (16) bound the gap to an ideal selective rerun; they do not establish a smaller error relative to the exact ODE solution than a fresh NFE-matched solver. The ideal rerun uses the same grid and selected corrections but re-evaluates the true velocity at every step. Under the assumptions and a fixed budget, the gap is second order in the maximum step size, preserving the rerun's leading first-order accuracy while avoiding downstream model calls. We apologize for the ambiguity and have clarified the comparison target after Eq. (16).

To assess reference accuracy directly, independently of ImageReward and CLIP, we measured latent MSE against a 1,000-step Euler reference on 1,024 SD3 prompts, using the three main-paper operating points.

a=12a = 12a=20a = 20a=32a = 32
Original Euler sample0.4478(0)0.4478^{(0)}0.3511(0)0.3511^{(0)}0.2299(0)0.2299^{(0)}
FixUp-FM0.4391(+0)\mathbf{0.4391}^{(+0)}0.3459(+1)\mathbf{0.3459}^{(+1)}0.2275(+1)\mathbf{0.2275}^{(+1)}
Restart0.4587(+1)0.4587^{(+1)}0.3612(+1)0.3612^{(+1)}0.2392(+1)0.2392^{(+1)}
SDEdit0.4682(+1)0.4682^{(+1)}0.3706(+1)0.3706^{(+1)}0.2487(+1)0.2487^{(+1)}
Fresh NFE-matched Euler0.4478(+0)0.4478^{(+0)}0.3399(+1)\mathbf{0.3399}^{(+1)}0.2205(+1)\mathbf{0.2205}^{(+1)}

Entries are latent MSE (↓). Superscripts denote total NFE minus aa; fresh Euler is rerun from noise.

FixUp is the only evaluated post-sampling method that reduces reference MSE at all three points, while remaining close to fresh NFE-matched Euler. The practical distinction is that FixUp achieves this accuracy without restarting sampling. Improving an existing 20-step sample takes one additional evaluation, whereas fresh 21-step Euler requires 21 new evaluations.

How does FixUp compare with strong multistep samplers? (Weakness 2; Question 2)

FixUp is a post-sampling refinement method, not a standalone solver. FixUp reuses the completed trajectory, revisiting selected intervals and propagating corrections to the output without re-evaluating every downstream velocity.

We additionally evaluated fresh Euler, Heun, DPM-Solver++(2M), and UniPC on FLUX.1-dev with 1,024 prompts, using the three main-paper operating points and matched initial noise. FixUp remains competitive with strong fresh solvers.

a=12a = 12a=20a = 20a=32a = 32
FixUp-FM0.8781(+2)0.8781^{(+2)}0.9072(+3)\mathbf{0.9072}^{(+3)}0.9151(+7)\mathbf{0.9151}^{(+7)}
Euler0.8768(+2)0.8768^{(+2)}0.8812(+3)0.8812^{(+3)}0.9007(+7)0.9007^{(+7)}
Heun0.8566(+3)0.8566^{(+3)}0.8859(+3)0.8859^{(+3)}0.9050(+7)0.9050^{(+7)}
DPM-Solver++(2M)0.9040(+2)\mathbf{0.9040}^{(+2)}0.8997(+3)0.8997^{(+3)}0.8980(+7)0.8980^{(+7)}
UniPC0.9002(+2)0.9002^{(+2)}0.9003(+3)0.9003^{(+3)}0.8898(+7)0.8898^{(+7)}

Entries are ImageReward. Superscripts denote total NFE minus aa; fresh solvers are rerun from noise. Heun uses 15 NFEs in the first column to accommodate its two-stage update.

FixUp achieves the highest mean ImageReward among all compared methods at both a=20a = 20 and a=32a = 32, and exceeds fresh Euler and Heun at all three operating points. Users can thus choose to improve a sample they already have with a small additional budget, without committing to a full sampling rerun.

Where are the nine losing settings, and why can extra corrections hurt at small base budgets? (Weakness 1; Question 3)

Tables 5–7 in Appendix K already report the complete 5,000-prompt sweeps, including the low-aa settings. The nine losses against samplewise NFE-matched Euler are listed below.

ModelLosing configurations (a,K)(a, K)
SD3 Medium(8, 2), (8, 3), (24, 8)
SD3.5 Large(8, 2), (8, 3), (12, 3)
FLUX.1-dev(8, 1), (8, 2), (8, 3)

Seven losses occur at a=8a = 8; the other two IR gaps are −0.00024 at SD3 (24,8)(24,8) and −0.00336 at SD3.5 (12,3)(12,3). Across all 120 settings, 119 achieve higher mean IR than the uncorrected Euler output.

A possible explanation is that too many additional corrections on coarse grids can move the trajectory farther from the cached trajectory, increasing the error of the local velocity approximation. We have added this low-budget analysis in Appendix K.1, after Table 7.

Why do repeated small requests outperform one large request on FLUX? (Question 4)

Splitting requests can change both selection and correction order. A single request selects all intervals upfront and corrects them in trajectory order. Smaller requests re-score the remaining intervals after previous corrections, so later requests can revisit earlier parts of the trajectory. Even when the final selected set is unchanged, Heun stages can therefore be evaluated at different states, producing different outputs at the same total NFE.

We suspect the FLUX–SD difference reflects how terminal and upstream corrections interact. Smaller requests can apply a high-scoring terminal correction early and retain it during later upstream updates, whereas a single request evaluates it last. The effect on image quality depends on the model's velocity field, so finer-grained requests need not benefit all models equally.

We hope these clarifications and additional experiments address your concerns.

Dear Reviewer cxbf,

We would like to follow up and ask whether our responses and additional results have addressed your concerns. We would be happy to discuss any remaining questions or provide further clarification. If your concerns have been resolved, we would be grateful if you could update your assessment accordingly.

Thank you for your time and consideration.

Thank you for the author to address my question. I would keep my current score.

Reviewer_62tgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to improve a completed few-step Euler flow-matching sample with a small additional NFE budget. FixUp-FM scores each interval by a trace-based Heun-minus-Euler defect estimate, upgrades the top-K intervals to Heun, and propagates the resulting state corrections downstream with a rank-one minimum-norm velocity adjustment. The adjustment is built from chord residuals of the reference trace, so propagated steps need no new model evaluations. Reusing the first selected endpoint stage brings the cost of K upgrades down to K−1 NFEs, and the procedure can be repeated incrementally. The theory establishes near-oracle defect selection, exact recovery of the propagated velocity under locally affine dynamics, and second-order agreement with an ideal selective Euler–Heun rerun. Experiments on SD3 Medium, SD3.5 Large, and FLUX.1-dev (5,000 prompts each) show higher ImageReward than NFE-matched Euler in 111/120 configurations. The paper also includes stable anytime-continuation results and ablations of the selection and propagation rules.

Critical strengths
  • Reusing the completed trajectory is a new and practical angle. Endpoint-stage reuse and the chord–correction alignment (Eq. 7) give gains at zero added NFE (+0.050/+0.058 ImageReward at a=12 on SD3/SD3.5).
  • The theory is coherent and complete: near-oracle selection, O(h²) propagation error, and second-order agreement with an ideal selective Euler–Heun rerun.
  • The experiments are large and consistent: three production-scale models, 5,000 prompts each, wins in 111/120 configurations, under 0.7% latency overhead, and stable anytime continuation where SDEdit and Restart degrade.
Critical weaknesses
  • There is no comparison with a fresh fixed-step Heun, DPM-Solver++, or UniPC run at the same total NFE. The solver baselines used (DPM-adaptive, DOPRI5, SRDS) are known to be weak at 8–40 NFEs.
  • ODE accuracy is never measured on the real models (only on affine examples, Appendix Table 2). It is therefore unclear whether the gains come from reduced discretization error.
  • The evaluation relies mainly on ImageReward. FLUX gains are small (0.01–0.02, negative at a=8), CLIP is flat, and Table 1 has no confidence intervals.
Questions for authors
  1. How does FixUp-FM at a+b total NFE compare with a fresh fixed-step Heun / DPM-Solver++ / UniPC run at the same a+b NFE on SD3 and FLUX, both in ImageReward and in distance to a high-accuracy reference solution?
3 comments

Thank you for the thoughtful review and for recognizing the practical value of trajectory reuse, our theoretical analysis, and the large-scale evaluation.

How does FixUp compare with strong fixed-step solvers? (Weakness 1; Question)

FixUp is an optional post-sampling refinement method, not an ODE solver. It refines an already completed sample using its trajectory with a small additional budget. Users can choose to improve an existing sample with a small additional budget, whereas using a different solver would require a fresh sampling run from noise.

FixUp remains competitive with strong fresh solvers in ImageReward. We additionally evaluated fresh Euler, Heun, DPM-Solver++(2M), and UniPC on SD3 Medium and FLUX.1-dev at 512×512512 \times 512, using the same 1,024 prompts and shared settings (a,K)=(12,1),(20,2),(32,2)(a,K)=(12,1),(20,2),(32,2). Methods within each model use matched initial noise.

SD3 Mediuma=12a = 12a=20a = 20a=32a = 32
FixUp-FM0.6688(+0)0.6688^{(+0)}0.7376(+1)0.7376^{(+1)}0.7590(+1)0.7590^{(+1)}
Euler0.6148(+0)0.6148^{(+0)}0.6943(+1)0.6943^{(+1)}0.7447(+1)0.7447^{(+1)}
Heun0.6079(+1)0.6079^{(+1)}0.7159(+1)0.7159^{(+1)}0.7365(+1)0.7365^{(+1)}
DPM-Solver++(2M)0.6982(+0)0.6982^{(+0)}0.7446(+1)0.7446^{(+1)}0.7579(+1)0.7579^{(+1)}
UniPC0.6898(+0)0.6898^{(+0)}0.7368(+1)0.7368^{(+1)}0.7674(+1)0.7674^{(+1)}
FLUX.1-deva=12a = 12a=20a = 20a=32a = 32
FixUp-FM0.8682(<1)0.8682^{(<1)}0.9045(+1)0.9045^{(+1)}0.9082(+1)0.9082^{(+1)}
Euler0.8676(<1)0.8676^{(<1)}0.8861(+1)0.8861^{(+1)}0.8929(+1)0.8929^{(+1)}
Heun0.8322(+1)0.8322^{(+1)}0.8904(+1)0.8904^{(+1)}0.9117(+1)0.9117^{(+1)}
DPM-Solver++(2M)0.8925(<1)0.8925^{(<1)}0.8997(+1)0.8997^{(+1)}0.8918(+1)0.8918^{(+1)}
UniPC0.9017(<1)0.9017^{(<1)}0.8970(+1)0.8970^{(+1)}0.8898(+1)0.8898^{(+1)}

Entries are ImageReward (↑). Superscripts denote total NFE minus aa; fresh solvers are rerun from noise. Heun uses 13 NFEs in the first column to accommodate its two-stage update.

We also hope that FixUp's trajectory-reuse principle will inspire future work on post-sampling refinement for higher-order samplers, for example through cached Runge–Kutta stages and embedded defect estimates.

Does FixUp reduce numerical error on real models? (Weakness 2; Question)

Corollary 1 and Eq. (16) bound the gap to an ideal selective rerun; they do not establish a smaller error relative to the exact ODE solution than a fresh NFE-matched solver. The ideal rerun uses the same grid and selected corrections but re-evaluates the true velocity at every step. Under the assumptions and a fixed budget, the gap is second order in the maximum step size, preserving the rerun's leading first-order accuracy while avoiding downstream model calls.

FixUp reduces the mean reference MSE of the completed Euler samples at all three tested points on both models. We measured latent MSE against a 1,000-step Euler reference on the same 1,024 prompts per model, using the shared (a,K)(a,K) settings above.

SD3 Mediuma=12a = 12a=20a = 20a=32a = 32
Original Euler sample0.4478(0)0.4478^{(0)}0.3511(0)0.3511^{(0)}0.2299(0)0.2299^{(0)}
FixUp-FM0.4391(+0)0.4391^{(+0)}0.3459(+1)0.3459^{(+1)}0.2275(+1)0.2275^{(+1)}
Restart0.4587(+1)0.4587^{(+1)}0.3612(+1)0.3612^{(+1)}0.2392(+1)0.2392^{(+1)}
SDEdit0.4682(+1)0.4682^{(+1)}0.3706(+1)0.3706^{(+1)}0.2487(+1)0.2487^{(+1)}
Fresh Euler0.4478(+0)0.4478^{(+0)}0.3399(+1)0.3399^{(+1)}0.2205(+1)0.2205^{(+1)}
Fresh Heun0.5242(+1)0.5242^{(+1)}0.4684(+1)0.4684^{(+1)}0.3726(+1)0.3726^{(+1)}
Fresh DPM-Solver++(2M)0.4141(+0)0.4141^{(+0)}0.3159(+1)0.3159^{(+1)}0.2067(+1)0.2067^{(+1)}
Fresh UniPC0.4301(+0)0.4301^{(+0)}0.3075(+1)0.3075^{(+1)}0.1804(+1)0.1804^{(+1)}
FLUX.1-deva=12a = 12a=20a = 20a=32a = 32
Original Euler sample0.4319(0)0.4319^{(0)}0.3020(0)0.3020^{(0)}0.2117(0)0.2117^{(0)}
FixUp-FM0.4303(<1)0.4303^{(<1)}0.2999(+1)0.2999^{(+1)}0.2103(+1)0.2103^{(+1)}
Restart0.4628(+1)0.4628^{(+1)}0.3315(+1)0.3315^{(+1)}0.2405(+1)0.2405^{(+1)}
SDEdit0.4499(+1)0.4499^{(+1)}0.3193(+1)0.3193^{(+1)}0.2287(+1)0.2287^{(+1)}
Fresh Euler0.4069(<1)0.4069^{(<1)}0.2906(+1)0.2906^{(+1)}0.2073(+1)0.2073^{(+1)}
Fresh Heun0.5482(+1)0.5482^{(+1)}0.4216(+1)0.4216^{(+1)}0.2791(+1)0.2791^{(+1)}
Fresh DPM-Solver++(2M)0.4106(<1)0.4106^{(<1)}0.2872(+1)0.2872^{(+1)}0.1959(+1)0.1959^{(+1)}
Fresh UniPC0.3823(<1)0.3823^{(<1)}0.2542(+1)0.2542^{(+1)}0.1627(+1)0.1627^{(+1)}

Entries are latent MSE (↓). Superscripts denote mean total NFE minus aa; fresh solvers are rerun from noise. Heun uses 13 NFEs in the first column. For FLUX, <1<1 denotes 0.993 additional NFEs on average.

Across both models, FixUp is the only evaluated post-sampling method that reduces mean reference MSE, and it also achieves lower MSE than fresh Heun. FixUp obtains these gains by refining the completed sample with a small additional budget, whereas fresh solvers require a full sampling rerun from noise.

How reliable are the reported gains, and what does CLIP show? (Weakness 3)

Only FLUX at a=8a=8 has a paired 95% confidence interval spanning zero; the remaining 20 intervals lie entirely above zero.

Base steps aaSD3 MediumSD3.5 LargeFLUX.1-dev
8[+0.06404, +0.07801][+0.05983, +0.07398][−0.00938, +0.00648]
12[+0.04565, +0.05427][+0.05310, +0.06230][+0.00883, +0.02325]
16[+0.03029, +0.03632][+0.03807, +0.04466][+0.00717, +0.01910]
20[+0.01597, +0.02896][+0.02532, +0.04009][+0.00911, +0.02110]
24[+0.01377, +0.02857][+0.01588, +0.03225][+0.00954, +0.02155]
28[+0.01817, +0.02833][+0.01424, +0.03141][+0.01331, +0.02493]
32[+0.01202, +0.02202][+0.01652, +0.03392][+0.00841, +0.01935]

Entries are pointwise paired 95% CIs for IR gain over samplewise NFE-matched Euler, using a normal approximation. We have added these intervals in Appendix Q (Table 19).

Actually, in practical use, users refine the existing aa-step sample, so their quality improvement can be substantially larger than the gain over fresh NFE-matched Euler reported in our paper.

The CLIP score shows that FixUp's ImageReward gains are accompanied by stable or improved text–image alignment. Table 1 reports comparable mean CLIP scores on the SD models and higher mean scores at all seven FLUX operating points, with gains of 0.0512–0.1397 over NFE-matched Euler.

We additionally evaluated GenEval on FLUX.1-dev at the paper's three FLUX operating points, using 1,106 images per setting (all 553 official prompts, two seeds each), 512×512512\times512 resolution and guidance 3.5. Methods share prompts and initial noise. We report the official evaluator's six-task macro-average.

Base + additional NFEFresh EulerFixUp
12 + 20.64980.6327
20 + 30.65100.6591
32 + 70.64840.6576

Entries are GenEval scores (↑). Budgets show base + additional NFEs for FixUp; fresh Euler uses the same total.

FixUp exceeds NFE-matched Euler on GenEval at 23 and 39 NFEs. In the separate MSE comparison above, FixUp also reduces the original Euler samples' mean reference error at all three tested points on both SD3 and FLUX. These evaluations extend the evidence beyond ImageReward to compositional correctness and numerical fidelity. We have added the GenEval setup and results in Appendix R (Table 20) and will expand the evaluation with additional samples if the paper is accepted.

We hope these clarifications and additional experiments address your concerns, and we welcome any further questions.

Dear Reviewer 62tg,

We would like to follow up and ask whether our responses and additional results have addressed your concerns. We would be happy to discuss any remaining questions or provide further clarification. If your concerns have been resolved, we would be grateful if you could update your assessment accordingly.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.