acceptodds
Under review as a conference paper at ICLR 2027

WAVEALIGN: Wavelet-Based Alignment for Training-Free High-Resolution Image Generation

Abstract

High-resolution image generation aims to extend pretrained generator beyond their native training resolution while preserving coherent global structure and enabling finer details to emerge. Existing progressive methods transfer information across resolutions, yet parent and child stages differ in spatial scale and sampling phase, leading to mismatches in initialization, reference selection, and structural guidance. To tackle this problem, we formulate progressive high-resolution generation as a cross-resolution correspondence problem involving stochastic, temporal, and structural coherence. Based on this perspective, we introduce , a training-free framework that establishes these correspondences in a shared orthonormal wavelet representation. Specifically, couples sampled noise and constructs a mixed clean anchor for coherent initialization, derives scale-time conjugacy for phase-matched parent references, and selectively corrects shared coarse structures while progressively releasing guidance for high-resolution detail synthesis. Experiments on FLUX.1-dev demonstrate effective progressive generation, with achieving the lowest FID, KID, pFID, and pKID at among the evaluated FLUX-based pipelines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Training-Free, High-Resolution Image Generation

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

This paper introduces WAVEALIGN, a training-free, high-resolution image generation method based on wavelet decomposition. Its core mechanism involves using a low-resolution reference to constrain low-frequency structures while allowing high-frequency components to generate new details as the generation resolution is progressively increased. The paper presents a corresponding cross-scale alignment and sampling mechanism, and validates the method's effectiveness through experiments, including comparisons with baselines and ablation studies.

Critical strengths
  1. By employing wavelet decomposition to separate global structure from local details, this approach provides a clear, actionable framework for maintaining layout while enhancing detail during high-resolution generation. The method requires no additional training, making it highly practical.
  2. The ablation studies are comprehensive; beyond presenting high-resolution generation results, the paper investigates factors such as noise coupling, phase, cross-scale connections, and decomposition depth, which helps clarify the role of each individual component.
Critical weaknesses

The additional gains provided by "stage-matching" references—as opposed to fixed references—warrant more direct validation. In the proposed method, high-resolution sampling involves querying trajectories from a coarser (parent) resolution based on a scale-dependent temporal mapping; these queries are then used to refine the low-frequency structure of the current prediction. While the appendix offers valuable evidence by comparing various temporal query rules, it does not explicitly quantify the incremental improvement gained from dynamically querying parent-resolution trajectories versus consistently using the final, denoised output from the parent resolution. It is recommended to compare these two reference strategies—while keeping factors such as wavelet initialization, noise coupling, the intensity and scheduling of low-frequency correction, auxiliary inference settings, and the total sampling budget constant—to isolate the specific contribution of the stage-matching mechanism to structural preservation and detail generation.

Questions for authors

How does performance change when wavelet alignment is removed, while keeping the auxiliary reasoning setup, cascaded workflow, and computational budget constant?

4 comments
AuthorsOpenReview ↗

Response to Reviewer txhT (Part1)

We thank Reviewer txhT for recognizing the practical value of the framework and for suggesting direct controlled comparisons. Below, we report the contribution of cross-resolution alignment under a fixed budget and the incremental benefit of dynamic reference selection.

Q1: Removing wavelet alignment with auxiliary inference and budget fixed

Following your suggestion, we added an ablation with auxiliary inference, the cascaded workflow, and the model-evaluation budget fixed. The comparison shows that cross-resolution alignment improves high-resolution generation quality under these controls.

Controlled ablation.

We fix the backbone weights, cascade resolutions, sampling nodes, and attention settings. We then cumulatively disable wavelet structural correction, replace wavelet-mixed initialization with pixel-only initialization, and remove parent–child noise coupling. The final configuration retains auxiliary inference and cascaded generation without these cross-resolution alignment constraints.

The table reports final 4K results on the same 1,000 prompts with seed 42. RADIO uses 1024×1024 inputs.

Configurations: A: WAVEALIGN (full method); B: No structural correction; C: Pixel-only initialization, no structural correction; D: No cross-resolution alignment.

MetricABCD
RADIO ↑34.201634.246433.790833.7030
FID ↓64.36464.49965.34567.091
KID×10³ ↓4.4094.3425.0776.563
pFID ↓42.67943.18663.40973.392
pKID×10³ ↓18.32418.43336.44552.561
IS ↑20.26918.98618.43818.488
pIS ↑7.7877.6276.0445.149

Removing all cross-resolution alignment increases FID from 64.364 to 67.091, pFID from 42.679 to 73.392, and pKID from 18.324 to 52.561, while reducing RADIO from 34.2016 to 33.7030. Under the same auxiliary inference settings and model-evaluation budget, these results show that cross-resolution alignment further improves high-resolution generation quality.

The intermediate ablations further illustrate the components' roles. Disabling structural correction alone increases pFID from 42.679 to 43.186 and pKID from 18.324 to 18.433, while RADIO and KID improve slightly. Replacing mixed initialization with pixel-only initialization in this configuration increases pFID to 63.409; additionally removing noise coupling increases it to 73.392. These results support the contribution of mixed initialization and noise coupling to high-resolution detail generation.

AuthorsOpenReview ↗

Response to Reviewer txhT (Part2)

Q2: Dynamic stage matching versus a fixed endpoint reference

Thank you for suggesting this direct comparison. The fixed-endpoint-reference experiments show a patch-level benefit from dynamic stage matching while retaining the other settings. We report the controlled image-quality comparison and a separate structural-consistency measurement below.

1. Dynamic versus fixed references under matched settings.

A fixed endpoint reference is conceptually related to I-Max's use of a low-resolution endpoint for guidance. In our experiment, we replace the dynamically queried parent clean prediction with the parent's last-step clean prediction, retaining wavelet-mixed initialization, noise coupling, the strength and schedule of low-frequency correction, auxiliary inference settings, and the model-evaluation budget. We test fixed references at 4K only and at both 2K and 4K. The former shares the same 2K parent trajectory and 4K initialization as the full method, directly isolating the effect of reference selection at 4K.

The table reports final 4K results on the same 1,000 prompts with seed 42 and includes the complete I-Max method as a reference. RADIO uses 1024×1024 inputs to evaluate global image–text consistency.

Configurations: I-Max: I-Max (full method); Fixed 4K: WAVEALIGN: fixed reference at 4K only; Fixed 2K/4K: WAVEALIGN: fixed references at 2K and 4K; Full: WAVEALIGN (full method).

MetricI-MaxFixed 4KFixed 2K/4KFull
RADIO ↑34.031134.186934.180134.2016
FID ↓65.54264.37364.66564.364
KID×10³ ↓6.0334.3584.2314.409
pFID ↓83.44844.33146.28942.679
pKID×10³ ↓56.26518.60019.13718.324
IS ↑19.16419.25019.04720.269
pIS ↑5.7797.6897.6787.787

Dynamic stage matching improves patch-level distributional quality. Relative to fixing the reference only at 4K, dynamic querying reduces pFID from 44.331 to 42.679 and pKID from 18.600 to 18.324, with a small increase in RADIO. Using fixed references at both 2K and 4K increases final 4K pFID to 46.289. Both fixed-reference variants still achieve better patch-level distributional metrics than the complete I-Max method, indicating contributions from WAVEALIGN's other design choices. The internal controlled comparisons directly quantify the incremental gain from dynamic stage matching.

2. Structural preservation with a fixed parent trajectory. To further examine structural preservation, we conduct a separate paired experiment on 20 prompts, fixing the 2K parent trajectory and 4K initialization. We measure the shared-LL2 structural residual between the final 4K output and the parent endpoint reference. Dynamic querying reduces the mean residual from 13.353% to 12.490% and yields a lower residual on 18/20 prompts. This indicates that dynamic stage matching improves consistency with the parent's coarse structure alongside patch-level distributional quality.

Thank you again for your thoughtful comments and valuable suggestions. Your feedback has helped us better explain the role of cross-resolution alignment and dynamic parent-reference selection. We will incorporate the corresponding analyses and ablation results into the revised manuscript.

Reviewer_txhTOpenReview ↗

Response to Authors

Thank you for the clarification. The response resolves my concerns.

Reviewer_LdFVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper mainly investigates training-free high-resolution image generation. For the existing cascaded diffusion inference framework, it proposes alignment on noise, timestep and structure, so that low-resolution generation results can be better directly injected into the high-resolution generation process for sequential generation. Based on the pre-trained FLUX model, generation experiments are conducted at 2K and 4K resolutions, demonstrating superior generation performance.

Critical strengths

The paper has a good representation, which making it easy for reader to read and follow.

Critical weaknesses
  1. Experiments need to be conducted on more backbones. Current experiments are only carried out on FLUX, which limits the generalizability of the proposed method. Experiments should be extended to additional backbones such as SDXL or SD3. Experiments on recent architectures like Qwen would be highly desirable.

  2. There is a lack of comparative experiments for temporal alignment. SD3 and FreCaS adopt different formulas for cross-resolution timestep conversion within the cascaded inference framework. The authors are asked to analyze and compare these approaches and conduct corresponding experiments.

  3. Prior literature has demonstrated a notable advantage of the cascaded inference framework: it can substantially speed up the overall inference process. For instance, HiDiffusion and FreCaS have reported approximately 3× inference acceleration for 4K generation. However, this paper appears to omit relevant discussion. The authors should perform corresponding experiments and report the results.

Questions for authors

See above.

Overall, the innovation and experiments of this paper are mediocre. If the authors supplement the discussions and experiments required above, the generalizability of its conclusions can be further improved, which will help the reviewer raise the score.

5 comments
AuthorsOpenReview ↗

Response to Reviewer LdFV (Part1)

We sincerely thank Reviewer LdFV for the constructive suggestions on backbone generalization, temporal alignment, and inference efficiency. We address these points with the additional evaluations and clarifications below.

Q1: Cross-backbone evaluation

Following your suggestion, we extended WAVEALIGN to SDXL and Qwen-Image. We report quantitative comparisons on SDXL and describe the visual evaluation on Qwen-Image separately below.

1. Quantitative evaluation on SDXL. We use the same SDXL base weights, 1,000 prompts, and seed 42 for evaluation at 2K and 4K. Among the nine compared methods, WAVEALIGN ranks first in RADIO, FID, KID, pFID, and pKID at 2K. At 4K, it ranks first in RADIO and FID and second in KID and pFID. For the evaluations below, RADIO uses a fixed image input size of 1024×1024 to measure alignment between the high-resolution generated images and their text prompts. Bold indicates the best result for each metric.

2K results

KID and pKID values are multiplied by 10³.

MethodRADIO ↑FID ↓KID ↓pFID ↓pKID ↓
WAVEALIGN37.940163.7512.46725.3745.878
FreCaS36.934564.2422.66328.8008.288
DemoFusion36.820666.5643.83828.9248.275
DiffuseHigh36.576866.5913.08627.4016.663
FreeScale36.344371.2285.64032.84411.373
AccDiffusion v236.232968.1474.04043.53316.740
ScaleCrafter35.289670.6593.88233.1487.545
HiDiffusion35.217470.6134.08935.5718.293
SDXL-Direct33.6739101.98715.65547.75817.131

4K results

KID and pKID values are multiplied by 10³.

MethodRADIO ↑FID ↓KID ↓pFID ↓pKID ↓
WAVEALIGN37.869064.8322.64933.49814.703
FreCaS36.532265.0362.12434.5088.876
DemoFusion36.504068.1664.70239.90314.677
DiffuseHigh36.472466.9643.32440.37115.217
FreeScale36.347668.9784.06033.24210.341
AccDiffusion v235.581869.3504.62968.93432.215
ScaleCrafter32.569484.5349.17359.40420.752
HiDiffusion28.6335100.08417.935122.13473.679
SDXL-Direct27.9161186.57272.70261.31621.714

2. Visual evaluation on Qwen-Image. We implemented 1K→2K→4K cascaded generation, retaining cross-resolution noise coupling, mixed initialization, phase-matched references, and coarse structural correction. We adapted the method to Qwen-Image's native sampling coordinates, latent normalization, positional encoding, and attention computation. The revision will include generated images and enlarged crops to demonstrate WAVEALIGN's high-resolution generation on this backbone.

AuthorsOpenReview ↗

Response to Reviewer LdFV (Part2)

Q2: Relationship to the time mappings in SD3 and FreCaS

Thank you for suggesting a closer comparison with SD3 and FreCaS. WAVEALIGN, SD3, and FreCaS derive cross-resolution time mappings from SNR relationships. After unifying the noise coordinates and mapping direction, their mappings can be written in the same form; their distinction lies in how the formula is used during generation.

1. Relationship and uses of the time mappings. In RF noise coordinates, the SNR is

S(σ)=(1−σσ)2.S(\sigma)=\left(\frac{1-\sigma}{\sigma}\right)^2.

Converting FreCaS's VP SNR scaling to these coordinates gives the child-to-parent mapping

\qquad a=r^{\gamma/2},$$ where $r$ is the side-length ratio. For $\gamma=2$, this mapping agrees with the inverse of SD3's geometric resolution-scaling formula. For adjacent stages with a twofold side-length increase, it also agrees with WAVEALIGN's $\Phi_2$. The formula serves different purposes: SD3 uses it to adjust resolution-dependent sampling times, FreCaS uses it to determine time correspondence at a cascade transition, and WAVEALIGN uses it to select a cached prediction from the completed parent trajectory that matches the child's current SNR. WAVEALIGN does not change the child's own sampling time grid. **2. Cross-resolution signal and noise correspondence.** Under ideal signal consistency, Haar analysis gives a child low-frequency signal with twice the amplitude of the corresponding parent signal, while standard Gaussian noise retains unit variance after analysis. Noise coupling further ensures that the parent and child low-frequency noise realizations match. We can therefore derive the parent reference time by matching the signal-to-noise ratio in shared coarse coordinates. **3. Controlled comparison of query rules.** Our existing ablations compare the following three query rules while keeping initialization, noise coupling, structural correction, auxiliary inference settings, and the model-evaluation budget fixed. For adjacent stages with a twofold side-length increase and $\gamma=2$, FreCaS's VP mapping and the inverse of SD3's geometric scaling formula agree with our(Φ₂). When these mappings are all used for parent-reference querying within our pipeline, these parameter settings produce identical query times and are therefore represented by the same experimental results. The table reports 4K results on 1,000 prompts with seed 42; RADIO uses 1024×1024 inputs. **Configurations:** Same σ: Same-noise-level query ($a=1$); Half shift: Half-shift query ($a=\sqrt{2}$); our(Φ₂): our(Φ₂) ($a=2$). | Metric | Same σ | Half shift | our(Φ₂) | | --- | ---: | ---: | ---: | | RADIO ↑ | **34.2365** | 34.2191 | 34.2016 | | FID ↓ | 64.694 | 64.427 | **64.364** | | KID×10³ ↓ | 4.703 | 4.564 | **4.409** | | pFID ↓ | 44.338 | 43.164 | **42.679** | | pKID×10³ ↓ | 20.205 | 19.033 | **18.324** | | IS ↑ | 19.400 | 19.279 | **20.269** | | pIS ↑ | 7.781 | 7.765 | **7.787** | Compared with same-noise-level querying, our(Φ₂) reduces pFID from 44.338 to 42.679 and pKID from 20.205 to 18.324 in the table's scaled units. Half-shift querying uses the coefficient in the FreCaS RF helper function we examined; converting FreCaS's VP rule with $\gamma=2$ instead yields our(Φ₂). Together with the actual-trajectory correspondence experiment, these results support the effectiveness of this query rule. **4. Interaction with noise initialization and re-noising.** At each resolution transition, we initialize the child by re-noising the mixed clean anchor with the preconstructed, cross-resolution coupled Gaussian noise. This makes the child's random component match the parent's in shared coarse coordinates while retaining independent randomness in the complementary detail subspace. The ablation visualizations show that removing low-frequency noise coupling, while retaining reference querying and structural guidance, degrades the layering of the eagle's feathers and the bark texture; 4K pFID increases from 42.679 to 53.942. Fine-detail generation thus depends on both reference-time selection and coordinated cross-resolution noise initialization and re-noising.
AuthorsOpenReview ↗

Response to Reviewer LdFV (Part3)

Q3: Inference efficiency and memory

Thank you for highlighting the need to quantify the inference cost. We have added comparisons of inference time, speedup, and peak memory on FLUX and SDXL.

Measurement protocol. All tests use a single NVIDIA H200 GPU with batch size 1. Each configuration receives one warm-up run, followed by measurement of average time over 10 prompts covering 10 semantic categories. Timing uses CUDA synchronization and includes text encoding, all generation stages, cross-resolution transitions, and final VAE decoding. Peak memory includes resident models. Each method retains its inference configuration, and speedups are computed using the mean Direct time at the same resolution in the corresponding table.

FLUX.1-dev

2K

MethodTime (s) ↓Speedup ↑Memory (GiB) ↓
Direct37.0731.000×33.878
I-Max49.8990.743×33.911
HiFlow34.4031.078×35.875
ScaleDiff36.9951.002×31.528
DyPE33.2291.116×33.875
UltraImage43.5360.852×33.974
SEGA33.2241.116×33.875
HRDiT29.6751.249×36.275
WAVEALIGN29.7171.248×35.786

4K

MethodTime (s) ↓Speedup ↑Memory (GiB) ↓
Direct342.0701.000×37.605
I-Max427.5600.800×40.670
HiFlow170.7602.003×39.849
ScaleDiff127.3622.686×36.061
DyPE310.9691.100×37.977
UltraImage305.2761.121×40.837
SEGA310.8581.100×37.977
HRDiT88.6373.859×45.801
WAVEALIGN149.7312.285×41.541

WAVEALIGN reduces FLUX 4K generation time from 342.070 to 149.731 seconds, a 56.23% reduction and a 2.285× speedup over Direct. It is also 2.856× and 1.140× faster than I-Max and HiFlow, respectively.

SDXL

Direct uses 50 Euler sampling steps at the target resolution with CFG set to 5.

2K

MethodTime (s) ↓Speedup ↑Memory (GiB) ↓
Direct12.4321.000×11.426
ScaleCrafter15.0040.829×11.427
HiDiffusion8.8411.406×11.426
FreCaS5.9792.079×11.553
DiffuseHigh18.3280.678×11.560
FreeScale21.6060.575×11.477
AccDiffusion v261.3200.203×20.315
DemoFusion36.3550.342×11.524
WAVEALIGN8.8681.402×11.456

4K

MethodTime (s) ↓Speedup ↑Memory (GiB) ↓
Direct95.4121.000×11.687
ScaleCrafter182.6350.522×13.075
HiDiffusion35.2942.703×11.687
FreCaS37.9702.513×11.823
DiffuseHigh98.9520.964×12.189
FreeScale187.6160.509×13.511
AccDiffusion v2488.6750.195×20.315
DemoFusion279.5750.341×11.895
WAVEALIGN37.1012.572×11.733

On SDXL, WAVEALIGN takes 37.101 seconds for 4K generation, reducing time by 61.11% relative to Direct and providing a 2.572× speedup. Its peak memory is 11.733 GiB, close to Direct's 11.687 GiB, and its generation time is comparable to FreCaS. Together with the quality results above, these experiments show that WAVEALIGN improves high-resolution generation quality on both backbones while substantially reducing generation time relative to Direct.

Reviewer_U712OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces WAVEALIGN, a training-free framework for extending pretrained text-to-image rectified flow models to higher resolutions through progressive generation. The method formulates cross-resolution coordination as three complementary alignment problems: stochastic, temporal, and structural. Using an orthonormal wavelet representation, WAVEALIGN constructs coupled Gaussian noise across resolutions and combines latent- and pixel-space information for initialization. It further derives a scale-time correspondence to select phase-matched predictions from the parent trajectory and applies selective structural corrections in shared coarse wavelet subspaces, allowing finer details to develop at higher resolutions. The framework is evaluated using frozen FLUX.1-dev for progressive 1K-to-2K-to-4K generation. Experiments compare WAVEALIGN against existing training-free high-resolution generation methods using image-level and patch-level metrics, while ablation studies investigate the contributions of noise coupling, phase matching, initialization, and structural guidance.

Critical strengths

The paper has two main strengths that support its potential contribution to training-free high-resolution image generation.

Methodological design. The paper provides a systematic approach to cross-resolution coordination by addressing stochastic, temporal, and structural alignment within a unified framework. The mathematical formulation is clearly connected to the implementation, and the integration of coupled noise, phase matching, and selective structural guidance represents a technically sound and well-designed solution to a relevant problem in training-free high-resolution generation.

Ablation results. WAVEALIGN achieves strong results against several recent training-free high-resolution generation methods, particularly on patch-level distributional metrics at 4K. The ablation studies provide useful evidence of the contributions of noise coupling, mixed initialization, and phase matching. These results support the practical effectiveness of the proposed pipeline, even though the broader generality and methodological novelty remain to be established.

Critical weaknesses

The paper presents two main weaknesses that limit the strength of its contribution.

Unclear methodological novelty. The proposed framework primarily relies on established properties of orthogonal projections, multiresolution decompositions, and progressive generative guidance. It remains unclear which benefits specifically arise from wavelet representations rather than from more general orthogonal multiresolution operators. Furthermore, the main paper does not sufficiently distinguish WAVEALIGN from closely related training-free approaches, particularly HiFlow and I-Max, relegating much of the relevant discussion to the appendix. Since the contribution appears to lie primarily in the integration of existing techniques, a clearer characterization of what is fundamentally new is essential to assess its significance. Related work must be discussed in the main body.

Incomplete experimental validation. Although the ablations support the effectiveness of individual components, they do not directly establish the accuracy of the theoretically derived phase correspondence on actual generation trajectories or isolate the advantages of wavelet-based alignment over alternative multiresolution formulations. Moreover, the strongest improvements concern patch-level distributional metrics, which do not necessarily imply better semantic consistency or global image quality, while CLIP Scores remain below several baselines. The evaluation is also limited to a single generative backbone and a single generation seed. These limitations make it difficult to determine how robust and generalizable the reported benefits are beyond the specific experimental setting.

Additional concerns. Beyond these central issues, some aspects of the evaluation could be strengthened. The relatively small, curated evaluation set and the absence of a systematic perceptual assessment limit the conclusions that can be drawn about overall visual quality. Moreover, the main paper lacks a comprehensive comparison of inference time and memory requirements against competing methods, which is particularly relevant for practical high-resolution generation. Also, the sensitivity of the results to wavelet decomposition depth and guidance settings suggests that the choice of configuration involves empirical trade-offs that deserve clearer discussion.

Questions for authors

The following questions concern aspects of the methodological contribution and experimental validation that could help clarify the significance of the proposed approach.

Q1. Role of the wavelets. Could the authors clarify the key methodological differences between WAVEALIGN and closely related approaches, particularly HiFlow and I-Max? More specifically, which properties or benefits of the proposed framework depend on wavelet representations rather than on general orthogonal multiresolution projections?

Q2. Validity of scale-time conjugacy. The proposed phase correspondence is derived under ideal cross-resolution consistency assumptions. Is there empirical evidence that this correspondence accurately reflects the relationship between actual parent and child generation trajectories? How do the authors assess the validity of these assumptions beyond the observed improvements in downstream image quality metrics?

Q3. Interpretation of image quality improvements. How do the authors interpret the substantial improvements in patch-level distributional metrics alongside the lower CLIP Scores compared with several baselines? Is there evidence that the generated fine details are not only statistically more realistic but also semantically appropriate and consistent with the global image structure?

10 comments
AuthorsOpenReview ↗

Response to Reviewer U712 (Part1)

We sincerely thank Reviewer U712 for the careful review and constructive comments. We address each point below.

Q1: Methodological differences and the role of wavelets

1. Methodological differences from HiFlow and I-Max.

I-Max uses a low-resolution endpoint to guide a projected flow and adapts positional encoding and attention for high-resolution inference. HiFlow upsamples the clean predictions at individual low-resolution sampling steps to construct a virtual reference flow, combining re-noising initialization, Fourier low-frequency residual correction, and acceleration alignment between adjacent sampling steps. WAVEALIGN further addresses the correspondence of noise and reference times across resolutions: how to construct coupled random initializations at resolution transitions, and how to select a prediction from the parent trajectory that matches the child's current signal-to-noise ratio (SNR).

Noise coupling at re-noising initialization. When initializing the high-resolution stage, we match the coarse wavelet coefficients of its noise to the parent noise while retaining new randomness in the complementary detail coefficients. The noise therefore satisfies

$ D\epsilon_H=\epsilon_L, $

while its marginal distribution at each resolution remains standard Gaussian. Mixed initialization also uses wavelet decomposition to combine the coarse components of the latent pathway with the complementary components of the pixel pathway. Thus, the child enters denoising with an explicit cross-resolution noise correspondence, rather than relying solely on subsequent reference guidance to establish it.

Parent-reference selection through cross-resolution SNR matching. Under ideal cross-resolution clean-signal consistency, orthonormal Haar analysis gives a child low-frequency signal with twice the amplitude of the corresponding parent signal, whereas standard Gaussian noise retains unit variance after analysis. Consequently, the projected child state and the parent state do not have the same SNR at the same numerical noise level. Matching their signal-to-noise coefficient ratios yields

τ=Φ2(σ)=σ2−σ.\tau=\Phi_2(\sigma)=\frac{\sigma}{2-\sigma}.

We query the parent trajectory using this mapping, rather than always using its endpoint or the same numerical time.

2. General orthogonal properties and the role of Haar.

Preservation of standard Gaussian noise and the minimum-norm correction that achieves a specified coarse-coordinate adjustment follow from row orthonormality of the analysis operator; they are not exclusive to wavelets. Haar provides a concrete construction with explicit spatial correspondence and cross-scale amplitude relations, allowing noise coupling, mixed initialization, and reference querying to be formulated in consistent coarse coordinates.

Haar preserves spatial correspondence during recursive coarse-scale analysis. On an aligned grid, each level-LL LL coefficient is a normalized aggregation of a corresponding 2L×2L2^L\times2^L spatial block. Further decomposition of the LL component associates each new coarse coefficient with a larger local region while preserving its spatial correspondence. Noise coupling, mixing of the two initialization pathways, and parent–child structural matching can therefore use coarse coordinates with explicit spatial locations.

AuthorsOpenReview ↗

Response to Reviewer U712 (Part2)

Q1: Methodological differences and the role of wavelets (continued)

3. Controlled comparisons of the structural representation.

For structural correction, we further test whether its effectiveness depends on the subspace defined by Haar. We replace the inner Haar structural projection with a global DCT low-frequency projection of equal rank, retaining the outer Haar cross-resolution operator, coupled noise, mixed initialization, reference-query rule, and guidance schedule. The replacement results are reported below. Both configurations use the same 1,000 prompts and seed 42; RADIO uses 1024×1024 inputs.

4K

MetricHaar (full method)DCT, replaced at 4K only
RADIO ↑34.201634.2007
FID ↓64.36464.355
KID×10³ ↓4.4094.398
pFID ↓42.67942.764
pKID×10³ ↓18.32418.403
IS ↑20.26919.175
pIS ↑7.7877.763

The two configurations have similar FID, KID, pFID, pKID, and RADIO results. This indicates that structural correction remains effective in the tested DCT subspace of equal rank and is not restricted to the Haar basis.

HiFlow with Haar structural guidance. We also test whether replacing the structural guidance of an existing method with Haar is sufficient to obtain WAVEALIGN's performance. We use the official Haar option in HiFlow to replace its Fourier structural guidance, retaining its initialization, time rule, acceleration term, and sampling settings. Results on 1,000 prompts are shown below.

MetricHiFlow-Haar (2K)WAVEALIGN (2K)HiFlow-Haar (4K)WAVEALIGN (4K)
RADIO ↑34.238634.238934.086534.2016
FID ↓65.31264.56165.46964.364
KID×10³ ↓5.1785.3464.7624.409
pFID ↓27.29423.77653.80542.679
pKID×10³ ↓8.6806.35128.28218.324
IS ↑19.54920.42019.29220.269
pIS ↑11.37412.1726.6327.787

With Haar structural operations in both pipelines, WAVEALIGN achieves lower FID, pFID, and pKID at both resolutions, indicating that its gains do not arise solely from replacing the structural representation.

We will clarify these methodological differences in the main text and move the relevant discussion from the appendix into the main paper.

AuthorsOpenReview ↗

Response to Reviewer U712 (Part3)

Q2: Empirical correspondence between actual parent and child trajectories

We appreciate the request for direct trajectory-level evidence. Our measurements show that SNR-matched querying identifies closer parent states even with structural correction disabled, and that its predicted query times are close to the empirical best matches. We examine these findings in three steps below.

1. Actual parent–child state discrepancies. We project the noisy parent and child states into shared wavelet coordinates, apply consistent amplitude normalization, and compare same-noise-level querying, half-shift querying, and our(Φ₂). To avoid structural correction directly enforcing the measured agreement, the experiment also includes trajectories with structural correction fully disabled while retaining coupled noise and mixed initialization.

We use 100 held-out prompts, with 10 prompts from each of 10 semantic categories, and generation seeds 42, 123, and 2026. Runs with and without structural correction are paired. Results with structural correction disabled are:

Query rule1K→2K Rel. RMS (%) ↓2K→4K Rel. RMS (%) ↓
Same noise level29.61829.453
Half shift18.91018.037
our(Φ₂)10.7938.930

Relative to same-noise-level querying, our(Φ₂) reduces relative RMS by 63.56% and 69.68% at the two resolution transitions. We measure discrepancies in shared LL2 coordinates over the original guidance interval, first averaging over the three seeds and sampling nodes for each prompt and then aggregating across prompts. These results show that SNR-matched querying identifies parent states closer to the child states even without structural correction.

2. Agreement with the empirically best-matching time. For each child state, we search the parent cache for the time minimizing the same state discrepancy, denoted by τ∗\tau^*, and measure its absolute deviation from the predicted time:

∣τ∗−Φ2(σ)∣.\left|\tau^*-\Phi_2(\sigma)\right|.

Results for the two transitions are:

TransitionMean abs. time error ↓95% CI
1K→2K0.0082[0.0071, 0.0093]
2K→4K0.0053[0.0044, 0.0063]

Confidence intervals are computed from 2,000 bootstrap resamples clustered by prompt, resampling all three seeds and sampling nodes for each prompt together. The small time errors indicate that the predicted query times are close to the parent times that minimize actual parent–child state discrepancies.

3. Control with shared noise and a different parent prompt. We further preserve the noise correspondence but pair each child trajectory with a parent trajectory from a different prompt and repeat the search. The mean absolute time errors increase to 0.1891 and 0.1584, respectively, showing that shared noise alone does not explain the temporal correspondence observed for correctly paired trajectories.

Together, these results show that our(Φ₂) yields lower parent–child state discrepancies than same-noise-level querying and predicts times close to the empirically closest parent states. This provides direct trajectory-level evidence for cross-resolution SNR matching and supports its use for parent-reference selection.

Comments on the paper

2 comments
Area_Chair_cYU6OpenReview ↗

Please read the response from authors and discuss

Dear reviewer fellows,

The author rebuttal has been posted. The rebuttal period is a critical part of our review process, allowing authors to clarify misunderstandings and address specific concerns.

Please take a moment to read the author response, engage in the discussion by replying to the authors' responses, and update your review if the rebuttal changes your assessment.

Your active participation is vital to ensuring a fair and thorough evaluation of the submissions. A constructive discussion helps us make the most informed decisions.

Thank you for your time and valuable contributions to the review process.

Best regards,

Your AC

AuthorsOpenReview ↗

Method Clarifications and Additional Experiments in the Revised WAVEALIGN Manuscript

Dear Area Chair and Reviewers,

Thank you for your careful reading of WAVEALIGN and for the questions that guided our revisions. The method, equations, and sampling procedure are unchanged from the original submission. The main changes, marked in blue, clarify how the method relates to prior work and bring the additional experiments into the paper.

The Introduction now distinguishes endpoint guidance in I-Max and trajectory guidance in HiFlow from our choices of how to couple the noise at successive resolutions and when to query the parent trajectory. Sections 2.3 and 2.4 explain the SNR basis for the query and the local spatial correspondence provided by Haar analysis.

Section 3.4 brings the controlled comparisons together. Table 2 measures parent–child state discrepancies on sampled trajectories. Table 3 compares a stage-matched reference with a fixed endpoint, and an inner Haar projector with a same-rank DCT projector. Its final three rows show what happens when we successively remove structural correction, mixed initialization, and low-frequency noise coupling at the same sampling budget. The 4K pFID rises from 42.679 for the full method to 73.392 after all three removals.

Tables 1 and 4 report latency, speedup, and peak memory for FLUX and SDXL. Relative to direct 4K sampling on the same backbone, WAVEALIGN achieves 2.285× and 2.572× speedups, respectively. WAVEALIGN ranks third in full-image RADIO at both FLUX resolutions (Table 1). Table 5 reports a blinded preference study with 19 participants. For examples generated with WAVEALIGN, please see the gallery in the supplementary material (Qwen-Image on pages 2–3 and FLUX on page 4).

We appreciate the time you have devoted to the paper and the discussion.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.