acceptodds
Under review as a conference paper at ICLR 2027

Endpoint-Calibrated Pullback Interpolation for Latent Image Diffusion

Abstract

Diffusion image interpolation can lose texture and edge definition between otherwise detailed endpoints. Geometric and density-based routing criteria guide intermediate latents, but do not directly target such losses in the rendered images. We propose Endpoint-Calibrated Pullback Interpolation (ECPI), a plug-in correction for existing diffusion interpolation paths. ECPI derives position-dependent detail references from the frequency-band statistics of reconstructed endpoints. Responses below these references define the spectral deficits to correct. We differentiate the rendered responses through denoising and decoding to obtain updates to the noisy intermediate latents, allowing the generative process to realize the correction. On MorphBench and our 100-pair Animals-and-Humans evaluation set (AH100), ECPI improves no-reference image quality and reduces FID against endpoint reference sets across SLERP, Tangent, and GeoDiff routes. These gains are achieved while constraining each correction to have a non-negative first-order change in conditional log density at the correction state. This shows that detail recovery is compatible with locally probability-supported correction directions.

Primary area:
generative models
Keywords:
diffusion models, image interpolation, generative models, latent space, spectral detail, probability geometry

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nqnZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Endpoint-Calibrated Pullback Interpolation, a plug-in correction for existing diffusion interpolation paths. It constructs position-dependent references from the log-frequency-band energies of reconstructed endpoints and treats responses below these references as one-sided spectral deficits. Gradients are propagated through the remaining denoising process and decoder to adjust intermediate noisy latents. Updates are constrained by route-normal motion, latent-norm preservation, and local score compatibility, aiming to improve intermediate-frame detail while retaining the underlying interpolation framework.

Critical strengths

1.A valuable shift in problem formulation. Rather than evaluating interpolation exclusively through latent-space geometry, the method identifies deficiencies in rendered images and asks the generative process to implement their correction.

2.A restrained objective design. Reconstructed endpoints calibrate the target to the capabilities of the inversion–generation system. The one-sided objective is also less restrictive than forcing every frequency response to match a prescribed reference exactly.

3.Careful investigation of alternative explanations. RGB sharpening, fitted filtering, decoder-only refinement, objective variants, and score-constraint controls provide substantially stronger evidence than aggregate quality comparisons alone. The paper also transparently reports transition-quality trade-offs.

Critical weaknesses

1.The necessity of endpoint-specific calibration is not established. Reference-value interventions show that the controller responds to its target, but not that the current endpoint pair provides a better target than a fixed enhancement request, population-level reference, or shuffled endpoint reference.

2.Spectral deficits need not indicate missing semantically appropriate detail. Global frequency energies do not specify where texture belongs or whether an edge is structurally correct. Deformation, occlusion, and category transitions may legitimately reduce particular frequency components.

3.Sharper frames do not necessarily imply more natural interpolation. No-reference quality metrics and FID against endpoint sets do not directly measure identity transition or sequence naturalness. Table 6 also shows costs in transition statistics. Complete-sequence perceptual evaluation would help resolve this distinction. (pp. 7, 14)

4.The necessity of full continuation backpropagation remains insufficiently isolated. Each pair and route requires 1,800 denoising-step vector–Jacobian products, without reported runtime or peak memory. The decoder-only control is useful, but differs in optimization and constraints, while its mean TOPIQ score is not clearly distinguishable from ECPI’s. (p. 13; Table 13)

Questions for authors

1.How important is the current pair’s endpoint reference? Could fixed, population-level, and shuffled-endpoint references be compared under a matched latent-displacement budget?

2.How does ECPI behave when reduced intermediate-frequency energy is appropriate because of deformation or occlusion? Have you identified corrections that introduce inappropriate texture or alter structure?

3.Is there perceptual evidence evaluating complete sequences and separately assessing local detail, structural plausibility, and identity continuity, particularly given the transition-statistic trade-offs?

4.Which differences between ECPI and decoder-only refinement account for the observed advantages? Please also report correction runtime and peak memory, and clarify whether a simpler or shorter-backpropagation alternative achieves comparable results.

2 comments

We thank the reviewer for the constructive questions. We added reference-source interventions, a targeted visual audit, selected multi-frame comparisons in Appendices G and H, and a controlled simpler-gradient comparison to examine endpoint calibration, perceptual quality, temporal consistency, and computational cost.

Q1. Importance of the current pair's endpoint reference

We compared current-pair, population-fixed, and three preselected shuffled endpoint references on GeoDiff MorphBench pairs. The population-fixed reference uses mean reconstructed endpoint log-band energies from a separate 12-pair calibration cohort, excluded from the primary recipient analysis.

For active corrections, we use the same actual spherical displacement (0.0025 times the endpoint chord). Shuffled-reference results are averaged within each recipient before pair-level inference.

The updated reference-source comparison shows consistent advantages for the current-pair reference (Appendix I.1, Table 15):

Own − ControlΔTOPIQ [95% bootstrap CI]ΔMUSIQ [95% bootstrap CI]
Population-fixed+0.00400 [0.00278, 0.00520]+0.220 [0.153, 0.290]
Shuffled (3 donors)+0.00460 [0.00329, 0.00584]+0.290 [0.220, 0.357]

All four confidence intervals are above zero. These results provide direct evidence that using the current pair's reconstructed endpoints as the spectral calibration reference improves perceptual-quality scores relative to both population-level and unrelated endpoint references.

The displacement is matched per active correction; reference-dependent activation rates may differ, so aggregate correction dose is not necessarily identical. The population-fixed reference is a constant log-energy target, distinct from a fixed enhancement amount relative to each intermediate image.

Q2. Semantically appropriate spectral reductions and structural changes

In 12 high-transition cases selected solely from baseline adjacent-frame variation, the median ECPI RGB change was 10.3% of the local baseline transition, and most showed no conspicuous isolated texture artifacts or contour breakage. These cases were selected independently of ECPI outputs and IQA scores following quantitative screening of 4,050 matched-position comparisons (Appendix I.6). Thus, substantial baseline transitions can coexist with modest ECPI interventions in the inspected cases.

The image comparisons show what the corrections change: facade and vehicle details become more distinct, while some corrections also adjust object contours (Appendix G, Figs. 15–19). The separate ten-case high-change audit flagged two cases for possible over-restoration. We also show two selected unfavorable cases in Figs. 20–21, with mixed brow/eye strokes and displaced facial features under correction. Appendix I.6 distinguishes these visual observations from the spectral flags.

This behavior is consistent with ECPI's calibrated, bounded intervention: reduced spectral energy during deformation or occlusion need not indicate missing detail, and the one-sided objective does not require exact matching to the endpoint reference. The score and latent-movement constraints restrict the correction, while the images reveal its actual perceptual effect.

Q3. Perceptual evidence across interpolation paths and temporal consistency

We provide complete 17-frame sequences for the 16 selected manuscript examples, with synchronized baseline/ECPI playback and matched FreeMorph outputs for Morph4Data.

The printed figures guide inspection of the three dimensions raised by the reviewer: aligned crops and ear contours show local detail (Figs. 15–19 and 22), vehicles show structural evolution (Figs. 23–25), and portraits show identity-related changes (Figs. 17 and 27). Appendices G and H display both endpoints and five consecutive intermediate positions; the viewer extends each example to its complete path. Appendix F also retains the two printed 17-frame sequences (Figs. 11–12).

For single-frame perceptual quality, ECPI improves TOPIQ, MUSIQ, and MANIQA across all three routes on Morph4Data without retuning, with all nine paired 95% confidence intervals above zero (Appendix I.2, Tables 16–17). These established no-reference predictors have been validated against human subjective image-quality ratings, providing perceptual-quality evidence beyond the optimized spectral objective.

To address the reported increase in unsmoothed PPL and PDV, we evaluated a predefined temporal-smoothing grid on all 90 MorphBench pairs. For each route, we used the unsmoothed baseline's mean PDV as a common ceiling and selected the smallest feasible smoothing strength independently for each arm, without optimizing for IQA. At these operating points, ECPI retains positive improvements:

RouteΔTOPIQΔMUSIQBaseline alphaECPI alpha
GeoDiff+0.01031+1.046800.10
Tangent+0.01399+1.954200.25

All four paired 95% confidence intervals are above zero. Selection is repeated within each pair bootstrap. Appendix I.4 and Table 19 report the protocol, gains, and achieved PDV values.

Thus, ECPI's measured frame-quality improvements persist at operating points satisfying a common mean-PDV ceiling on this cohort. The unsmoothed examples in Appendices F–H complement these measurements with direct visual inspection of detail and transition behavior.

Q4. Full continuation and computational cost

To address the confounds in the earlier decoder-only control, we added a same-state, same-controller comparison with a one-step predicted-clean-plus-decoder Jacobian approximation. Both arms update the same noisy latent, use the same actual baseline deficits, endpoint references, route-normal/score constraints, and controller rules, and are evaluated through the complete continuation. The substituted component is the readout Jacobian; weights and doses are recomputed under the same rules using each Jacobian.

On the frozen 12-pair mechanism cohort across three routes, full pullback achieves 1.74–2.11 times the mean spectral-deficit recovery at equal midpoint displacement. On complete, perceptually resampled 17-frame paths, the gains over the approximation are (Appendix I.3, Table 18):

RouteTOPIQ gainMUSIQ gain
GeoDiff+0.00299+0.281
Tangent+0.01044+0.936
SLERP+0.01024+0.747

All six paired 95% confidence intervals are above zero. The full-path comparison uses the same controller and movement cap, with realized displacement determined separately in each arm. Full pullback also yields higher spectral excess and higher PDV on SLERP.

On an RTX 4090, in FP32 at 512 x 512, correction takes approximately 143 s versus 31 s, and existing-route-to-final-output time is 173 s versus 60 s, for full versus approximate gradients. Peak allocated memory, including resident models, is 11.23 versus 12.59 GiB for these implementations. Both timings include the actual-baseline forward computation; inversion, text-condition optimization, route construction, model loading, I/O, and IQA scoring are excluded.

The controlled comparison demonstrates stronger target recovery and measured perceptual-quality gains from full-continuation gradients relative to the tested approximation, with an explicit runtime cost. Appendix I.5, Tables 20–21, provides the measured-stage costs and separately labeled pipeline estimates.

Reviewer_d2YJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Endpoint-Calibrated Pullback Interpolation (ECPI), a plug-in correction method designed to improve image interpolation in latent diffusion models. The authors identify a recurring issue in existing latent routing methods (like SLERP, Tangent, and GeoDiff): intermediate frames often suffer from "spectral deficits," losing high-frequency details (textures and edges) present in the endpoint images. To address this, ECPI derives position-dependent frequency-band references from the reconstructed endpoints. It then computes the spectral deficits of the intermediate frames and differentiates these rendered responses backward through the remaining denoising and decoding steps. This "pullback" mechanism yields bounded updates to the noisy intermediate latents, effectively instructing the generative process to recover missing details while constraining the updates to maintain local probability support (non-negative first-order change in conditional log density). Experiments on MorphBench and a custom AH100 dataset show that ECPI consistently improves no-reference image quality metrics (TOPIQ, MUSIQ, MANIQA) and reduces FID across different base routes.

Critical strengths
  1. Elegant and Theoretically Grounded Formulation: Framing the loss of detail in intermediate frames as a "spectral deficit" and correcting it via a differentiable pullback through the frozen denoising/decoding chain is a highly innovative approach. It avoids the pitfalls of direct pixel-space optimization (which often breaks the diffusion prior) by keeping the corrections anchored to the model's native generative process and local probability density.

  2. Rigorous Mechanistic Ablations: The paper goes beyond simple metric reporting by conducting extensive diagnostic experiments. The authors convincingly demonstrate that ECPI is not merely an elaborate post-hoc sharpening filter (via RGB filtering comparisons and best-fit reconstructions) and provide a deep analysis of how the corrections interact with the model's score function and conditional log-density.

Critical weaknesses
  1. Inadequate Visualization of the Interpolation Process: Image interpolation and morphing are fundamentally about the temporal coherence and smoothness of the transition, not just the quality of isolated frames. However, the main text visualizations (e.g., Figures 1, 4, and 8) predominantly focus on single intermediate frames (like s=9/16) or localized crops, alongside the endpoints. While the appendix contains some sequence strips, the main text fails to visually demonstrate the full temporal dynamics of the corrected paths. More alarmingly, the quantitative transition metrics (Table 6) reveal that ECPI actually increases PPL (Perceptual Path Length) and PDV (Path Distance Variance). This indicates that while single-frame detail is recovered, the overall smoothness and uniformity of the morphing sequence are degraded. Without comprehensive visual evidence of the full sequences in the main text, it is impossible to judge whether the "detail recovery" introduces temporal flickering, jarring structural shifts, or semantic inconsistencies in the middle frames.

  2. Lack of Modern Baselines and Dataset Diversity: The empirical evaluation is narrow. The chosen baselines are primarily geometric or latent-space routing methods (SLERP, Tangent, GeoDiff) and basic RGB filtering. The paper completely ignores recent, state-of-the-art diffusion image morphing and interpolation techniques that explicitly handle semantic correspondence or use test-time optimization (e.g., DiffMorpher, IMPUS, FreeMorph). Comparing only against routing heuristics makes it impossible to assess ECPI's true standing in the current landscape. Furthermore, the evaluation relies on very small, domain-specific datasets (MorphBench's 90 pairs and AH100's 100 pairs, heavily skewed towards faces, animals, and stylized art). Datasets that used in the previous works are missing.

Questions for authors

1.Can you provide full sequence visualizations in the rebuttal to demonstrate the temporal coherence of the ECPI-corrected paths compared to baselines?

  1. Why were recent state-of-the-art diffusion morphing methods (such as DiffMorpher, IMPUS, FreeMorph, or optimization-based approaches) excluded from the empirical baselines?
4 comments

We thank the reviewer for highlighting the need for sequence visualizations and modern morphing baselines. We have added an evaluation on all 76 Morph4Data pairs, a FreeMorph comparison, and selected multi-frame visualizations in Appendices G and H. We also evaluate whether ECPI's frame-quality gains persist under a common temporal-variation budget.

Q1: Sequence visualizations and temporal consistency. We provide complete 17-frame sequences for the 16 selected manuscript examples, with synchronized baseline/ECPI playback and matched FreeMorph outputs for Morph4Data.

Appendix F includes complete printed sequences for the Van Gogh and animal examples (Figs. 11–12). Appendices G and H provide legible endpoint-plus-five-position panels: Appendix G includes matched detail crops and two failure examples (Figs. 20–21), and Appendix H adds comparisons with FreeMorph (Figs. 22–28). The online sequences extend these local windows to all 17 positions for each displayed case.

To examine the reported increase in unsmoothed PPL and PDV, we tested whether ECPI retains its frame-quality gains under a common mean-PDV ceiling. On all 90 MorphBench pairs, we evaluated the same smoothing grid {0, 0.10, 0.25, 0.50} for baseline and ECPI. Using each route's unsmoothed baseline mean PDV as the ceiling, we independently selected the smallest feasible strength for each arm, without selecting by IQA. The selected baseline/ECPI strengths were 0/0.10 for GeoDiff and 0/0.25 for Tangent. At these operating points, ECPI retains the following gains:

RouteTOPIQ gain [paired 95% CI]MUSIQ gain [paired 95% CI]
GeoDiff+0.01031 [0.00774, 0.01357]+1.0468 [0.7710, 1.3629]
Tangent+0.01399 [0.01061, 0.02331]+1.9542 [1.5531, 2.6558]

All four paired confidence intervals are above zero, showing that ECPI retains measured frame-quality gains at operating points satisfying the common mean-PDV ceiling on this cohort. Appendix I.4 and Table 19 detail the selection rule and achieved PDV. The unsmoothed examples in Appendices F–H complement this post-processing experiment with direct visual context.

Q2: Modern morphing baselines and dataset coverage. SLERP, Tangent, and GeoDiff were originally chosen to isolate ECPI's contribution across different routing criteria while holding the generator and evaluation settings fixed. We have now expanded the evaluation to all 76 Morph4Data pairs and added the recent tuning-free morphing method FreeMorph (Cao et al., ICCV 2025). The new evaluation covers all four classes (25/21/15/15 pairs), including cross-semantic and cross-layout transitions as well as vehicle, face, and animal examples.

We use FreeMorph's official implementation, retaining its method-specific model and guidance settings while adapting resolution and output sampling to a common 512×512, 17-frame protocol. All methods use the same input images and uniformly spaced output positions. This provides an application-level comparison alongside the controlled baseline/ECPI comparisons.

MethodFID ↓TOPIQ ↑MUSIQ ↑MANIQA ↑PPL ↓PDV ↓ISTD ↓
GeoDiff62.350.565866.050.45061.11320.05950.0331
GeoDiff + ECPI61.440.581267.140.46591.14090.06090.0349
Tangent66.140.512759.970.39001.17580.05910.0274
Tangent + ECPI65.180.540662.760.41881.27230.06440.0302
SLERP65.350.535861.840.41271.09080.05110.0239
SLERP + ECPI64.990.560164.170.43741.15500.05590.0268
FreeMorph115.960.660169.520.54352.64970.10160.0761

Without retuning, ECPI improves all three IQA metrics across all three routes, with all nine paired 95% confidence intervals above zero. Appendix I.2 reports the comparison protocol, absolute results (Table 16), and paired intervals (Table 17). This demonstrates consistent frame-quality benefits on the added benchmark.

Appendix H provides seven selected route-level comparisons across six input pairs (Figs. 22–28), covering character, vehicle, and cross-category transitions. Each shows baseline, ECPI, and FreeMorph at the same five consecutive positions. Figures 23 and 25 revisit the same car pair with different base routes.

FreeMorph achieves higher IQA scores, while the corrected geometric routes have lower endpoint-reference FID and path-variation measures under this protocol. FID is reported as a point estimate against the endpoint reference set. The quantitative table and aligned visual examples jointly contextualize ECPI-enhanced routes alongside a modern morphing system.

Reviewer_d2YJOpenReview ↗

New experimental results show the weakness more clearly

From the new experimental results, it is very clear to me that the proposed method is beaten both quantitatively and qualitatively by the simple, trivial method SLERP or by FreeMorph, which the authors did not compare against in their initial submission (whether accidentally or purposely)

Normally, I would not decrease the score during rebuttal, and I really appreciate that the authors submitted a rebuttal despite all scores being 2. However, I would like to send a clear message to the AC that this paper should be clearly rejected.

AuthorsOpenReview ↗

Clarifying ECPI's contribution and joint quality–FID improvements

Thank you for examining the additional results. The conclusion appears to rest on a misunderstanding of ECPI's role and the comparison being evaluated: it conflates the incremental benefit of correcting an existing interpolation route with whether the resulting system outperforms every alternative on every metric. We respectfully disagree that the additional results establish ECPI's overall inferiority to SLERP or FreeMorph.

1. ECPI improves existing latent interpolation routes.

SLERP, GeoDiff, and Tangent construct paths connecting endpoint representations in the diffusion model's latent space. These are the base latent interpolation routes to which ECPI is applied. Our focus is a specific problem: intermediate renderings can lose texture and edge definition even when both endpoints contain clear detail.

ECPI is a plug-in correction using endpoint-calibrated spectral references to guide bounded updates within the generative process. Its primary controlled evaluation therefore compares each base route with the same route plus ECPI, holding the underlying generator and route settings fixed.

The central empirical result is simultaneous improvement in image-quality scores and endpoint-reference FID. On all 76 Morph4Data pairs, Table 16 reports:

MetricSLERPSLERP + ECPI
TOPIQ ↑0.53580.5601
MUSIQ ↑61.8464.17
MANIQA ↑0.41270.4374
Endpoint-reference FID ↓65.3564.99

All three paired IQA improvement intervals are above zero. The same joint improvement holds for GeoDiff and Tangent: ECPI improves all three IQA metrics and lowers the reported endpoint-reference FID on every tested base route, with all nine paired IQA intervals above zero (Tables 16–17). These results directly demonstrate ECPI's contribution over the underlying routes, including SLERP. SLERP retains lower unsmoothed transition variation, but it does not outperform SLERP + ECPI on the reported image-quality and distributional measures.

2. FreeMorph has substantially higher endpoint-reference FID and transition variation.

FreeMorph achieves higher image-quality scores. However, its endpoint-reference FID and all three transition measures are substantially higher than those of SLERP + ECPI under the reported comparison protocol:

MetricSLERP + ECPIFreeMorph (ratio to SLERP + ECPI)
Endpoint-reference FID ↓64.99115.96 (1.78×)
PPL ↓1.15502.6497 (2.29×)
PDV ↓0.05590.1016 (1.82×)
ISTD ↓0.02680.0761 (2.84×)

FreeMorph's endpoint-reference FID is 1.78× that of SLERP + ECPI, indicating a larger discrepancy between its generated intermediate-image distribution and the original endpoint reference distribution in the evaluation feature space. Its PPL, PDV, and ISTD are also 2.29×, 1.82×, and 2.84× as high. Corrected GeoDiff and Tangent likewise have lower values on all four measures than FreeMorph.

ECPI improves IQA while reducing endpoint-reference FID relative to every base route. An assessment based only on image-quality scores overlooks this joint benefit and does not support the blanket claim of quantitative inferiority.

The original paired-route experiments were designed to isolate this plug-in contribution. We added FreeMorph and the complete 76-pair Morph4Data evaluation to provide modern system-level context. The shared-input uniform17 comparison protocol and method-specific model and guidance settings are documented in Appendix I.2.

The accompanying response addresses the transition trade-off and visual evidence.

AuthorsOpenReview ↗

Transition trade-offs and complete-sequence evidence

3. The same quality–transition trade-off is already present among uncorrected base routes.

On Morph4Data, GeoDiff improves endpoint-reference FID over SLERP (62.35 vs. 65.35) and achieves higher TOPIQ, MUSIQ, and MANIQA, while increasing PPL (1.1132 vs. 1.0908) and PDV (0.0595 vs. 0.0511). Tangent also has higher PPL (1.1758) and PDV (0.0591) than SLERP. Thus, ranking methods solely by PPL or PDV would place SLERP ahead of both routes, while overlooking GeoDiff's gains in image quality and endpoint-distribution agreement.

The FID–PDV trade-off is also documented in prior work: on MorphBench Metamorphosis, both GeoDiff and Tangent achieve lower FID but higher PDV than SLERP (Saito and Matsubara, Table 2).

This is precisely why we report quality, distributional, and transition measures jointly. Lower transition variation is a desirable property, but it is not the sole objective of image morphing. ECPI addresses the complementary problem of detail attenuation along an existing route, and the paired comparisons show that its IQA improvements accompany lower endpoint-reference FID on all three routes.

4. ECPI's quality gains persist when its transition cost is controlled.

We explicitly report ECPI's increase in unsmoothed PPL, PDV, and ISTD relative to its base routes. We also directly investigate this cost on all 90 MorphBench pairs.

We apply the same candidate smoothing grid to baseline and ECPI, and independently select each arm's smallest feasible strength under a common ceiling defined by the corresponding unsmoothed baseline's mean PDV. Selection does not use IQA. Modest smoothing of ECPI outputs retains:

  • GeoDiff, α = 0.10: TOPIQ +0.01031, MUSIQ +1.0468.
  • Tangent, α = 0.25: TOPIQ +0.01399, MUSIQ +1.9542.

The baseline uses α = 0 in both comparisons, and all four paired 95% confidence intervals are above zero (Table 19). The evidence therefore goes beyond acknowledging a trade-off: ECPI's frame-quality benefits persist at operating points satisfying the corresponding baseline's mean-PDV ceiling.

5. The visual improvements are available for direct inspection.

Our anonymous viewer provides complete 17-frame sequences for the 16 selected manuscript examples, including synchronized baseline/ECPI comparisons and matched FreeMorph outputs for Morph4Data.

Relative to the base routes, the examples show concrete changes: more distinct ear contours in Fig. 22, less diffuse headlights and vehicle boundaries in Fig. 24, and clearer facial linework in Fig. 27. The complete sequences allow these changes to be examined throughout the transitions, alongside the endpoint-plus-five-position panels in Appendices G–H. Selected unfavorable cases are also retained in Appendix G and the viewer.

The paired-route experiments isolate ECPI's incremental contribution. The added benchmark, FreeMorph comparison, smoothing analysis, and complete visual examples broaden that evaluation. Together, they support simultaneous improvements in image-quality scores and endpoint-reference FID across existing interpolation routes, with temporal and computational costs explicitly evaluated.

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

The paper is well organized and shows consistent improvements in image quality metrics. However, the limited novelty, lack of human evaluation, and unresolved computational and temporal trade-offs weaken its contribution. Stronger evidence is needed to support acceptance. I feel the same only.

Critical strengths

The paper addresses a clear problem and proposes a training-free correction that can be applied to existing interpolation methods. The same is a useful thing only.

The method is well defined, with a clear spectral objective and a score-based constraint. The formulation itself is neat.

The experiments cover multiple interpolation methods and include useful comparisons and ablation studies. The same gives good confidence.

Critical weaknesses

The technical novelty is limited. The method is closely related to existing diffusion guidance approaches, and its advantages over simpler guidance methods are not fully established. I have a doubt here itself.

The perceptual benefits remain uncertain. No human evaluation is provided, and Table 3 shows that simple RGB filtering achieves higher no-reference quality scores than ECPI, although its FID is worse. This is a worrying point only.

The method reduces transition consistency and requires substantial computation. Table 6 shows increased temporal variation, while runtime and GPU memory costs are not reported. The authors should have given the same itself.

Questions for authors

N/A

2 comments

We thank the reviewer for recognizing the problem formulation and controlled experiments. We have added a simpler-gradient control, an evaluation on all 76 pairs of Morph4Data, introduced in FreeMorph [1], selected multi-frame comparisons in Appendices G and H, a common-PDV-budget smoothing analysis, and computational profiling. We address the concerns below.

1. Technical contribution and simpler guidance. Our contribution is an endpoint-calibrated correction framework for existing interpolation routes. Reconstructed endpoint spectra provide scale-specific references without ground-truth intermediate images; full-continuation pullback translates these requests into bounded latent updates. The contribution lies in this correction formulation and its constrained realization through the generative process.

To test the value of continuation gradients, we added a same-state, same-controller comparison with a one-step predicted-clean-plus-decoder Jacobian approximation. Both arms update the same noisy latent and use the same actual rendered deficits, endpoint references, constraints, and movement caps. Weights and doses are recomputed under the same rules using each Jacobian. Every corrected state is evaluated through the complete remaining denoising process.

On the frozen 12-pair mechanism cohort across three routes, full pullback achieves 1.74–2.11 times the mean spectral-deficit recovery at equal midpoint displacement. On complete, perceptually resampled 17-frame paths, it improves predicted image-quality scores relative to the approximation:

RouteTOPIQ gain [paired 95% CI]MUSIQ gain [paired 95% CI]
GeoDiff+0.00299 [0.00039, 0.00629]+0.281 [0.100, 0.476]
Tangent+0.01044 [0.00634, 0.01477]+0.936 [0.465, 1.473]
SLERP+0.01024 [0.00444, 0.01693]+0.747 [0.401, 1.131]

All six paired intervals are above zero. The full-path comparison uses matched movement caps, with realized displacement determined separately in each arm. It supports the practical value of full continuation relative to the tested approximation; full pullback also has higher spectral excess and SLERP PDV. The protocol and results are now in Appendix I.3, Table 18.

2. Perceptual quality evidence and RGB filtering. TOPIQ, MUSIQ, and MANIQA are learned image-quality predictors evaluated against human subjective quality judgments on IQA benchmarks. They provide perceptually motivated evidence beyond sharpness or frequency-energy measurements alone.

On the new 76-pair Morph4Data evaluation, ECPI improves all three metrics across all three routes without retuning (Appendix I.2, Tables 16–17):

RouteTOPIQ gainMUSIQ gainMANIQA gain
GeoDiff+0.01538+1.0856+0.01530
Tangent+0.02797+2.7910+0.02881
SLERP+0.02433+2.3306+0.02468

All nine paired 95% intervals are above zero. The new aligned examples complement these gains: Appendix G shows matched detail crops (Figs. 15–19), and Appendix H shows five-position comparisons across all three routes alongside FreeMorph (Figs. 22–28). Figure 22 highlights a more distinct ear contour, while Figs. 24 and 27 expose changes in vehicle detail and facial boundaries. These provide visual evidence alongside the measured frame-quality gains.

The tested fixed RGB filter achieves higher IQA in the submitted comparison, alongside worse endpoint-reference FID than the unprocessed baseline. This is a measured quality–distribution trade-off. Our claim is consistent paired improvement through constrained, generator-mediated correction; direct human preference over RGB filtering remains untested.

3. Temporal consistency and sequence visualizations. To examine the reported increase in unsmoothed PPL and PDV, we evaluated the same candidate smoothing grid {0, 0.10, 0.25, 0.50} for baseline and ECPI on all 90 MorphBench pairs for GeoDiff and Tangent.

Each arm selects its smallest feasible strength under the unsmoothed baseline's mean-PDV ceiling, without using IQA; selection is repeated within each pair bootstrap. The selected baseline/ECPI strengths are 0/0.10 for GeoDiff and 0/0.25 for Tangent. The retained ECPI gains are (Appendix I.4, Table 19):

RouteTOPIQ gain [paired 95% CI]MUSIQ gain [paired 95% CI]
GeoDiff+0.01031 [0.00774, 0.01357]+1.0468 [0.7710, 1.3629]
Tangent+0.01399 [0.01061, 0.02331]+1.9542 [1.5531, 2.6558]

All four paired confidence intervals are above zero, showing that measured frame-quality gains persist at operating points satisfying a common mean-PDV ceiling. Appendix F provides complete 17-frame comparisons for two examples (Figs. 11–12). The new Appendices G and H show selected unsmoothed five-position windows with endpoint context, including the unfavorable corrections in Figs. 20–21, for direct inspection of local transition behavior.

The anonymous sequence viewer extends the 16 selected manuscript examples to complete 17-frame paths, with synchronized baseline/ECPI playback and matched FreeMorph outputs for Morph4Data.

4. Computational cost. Appendix I.5 reports the measured-stage costs (Table 20) and separate pipeline estimates (Table 21). To include the work preceding rendering, we composed route-level runtime estimates from conditioning/inversion, route construction, ECPI correction, and final rendering:

RouteBaselineWith ECPI
SLERPapproximately 105 sapproximately 246 s
GeoDiffapproximately 138 sapproximately 279 s
Tangentapproximately 793 sapproximately 934 s

These stage-composed estimates combine measurements from all 15 Class C pairs with separately measured RTX 4090 rendering times of approximately 29 s for perceptually resampled outputs. They exclude model loading, caption generation, I/O, and metrics. Rendering and correction use FP32 at 512 x 512; Tangent optimization uses bf16. Appendix I.5 distinguishes these estimates from directly measured stages.

Separately, the controlled full-versus-approximate comparison on an RTX 4090 in FP32 measures correction at approximately 143 s versus 31 s, or 173 s versus 60 s including final output from an available route. Both include actual-baseline forward computation.

Independent profiling on three pairs per route, with two repetitions, gives peak allocated memory of 9.72 GiB for baseline rendering and 11.24 GiB for ECPI correction plus output, including resident models. These peaks cover the measured stages. The results quantify the added computation associated with stronger recovery and measured quality gains.

Reference

[1] Yukang Cao, Chenyang Si, Jinghao Wang, and Ziwei Liu. FreeMorph: Tuning-Free Generalized Image Morphing with Diffusion Model. ICCV, 2025, pp. 18111–18120.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Manuscript and Additional Results

Dear Reviewers and Area Chair,

We have revised the manuscript in response to your feedback. Additions and revisions are highlighted in blue. All section, table, and figure references in our responses refer to the revised manuscript.

The revision includes:

  • Endpoint-reference comparisons: Appendix I.1, Table 15.
  • Morph4Data evaluation and FreeMorph comparison: Appendix I.2, Tables 16–17.
  • Full-continuation versus one-step gradient comparison: Appendix I.3, Table 18.
  • Temporal smoothing under a common mean-PDV ceiling: Appendix I.4, Table 19.
  • Runtime and memory results: Appendix I.5, Tables 20–21.
  • Selected qualitative comparisons and failure cases: Appendices G–H, Figures 15–28. Complete 17-frame sequences for the 16 selected manuscript examples are available in our anonymous sequence viewer, with synchronized baseline/ECPI playback and FreeMorph comparisons for Morph4Data.

We thank you for your constructive feedback and hope these additions help clarify the method’s benefits and trade-offs.

The viewer provides all 17 frames for the 16 selected manuscript examples, with synchronized baseline/ECPI playback and additional FreeMorph comparisons for Morph4Data. Please access the complete interpolation sequences here: https://interpolation-sequences-969765a0.pages.dev/#F11

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.