acceptodds
Under review as a conference paper at ICLR 2027

Tokenscale: Training-Free High-Resolution Image Generation In Token Space

Abstract

Diffusion Transformer (DiT)-based flow models achieve strong synthesis performance at training resolution, yet their extension to high-resolution (HR) generation frequently suffers from blurring and spatial layout collapse. We attribute this degradation to the token level. As resolution increases, image tokens grow quadratically while text tokens remain fixed, thereby diluting image–text attention and disrupting prompt alignment. In addition, persistent low-resolution reference guidance can restrict the synthesis of new HR details. To address these issues, we propose TokenScale, a training-free token-level framework. TokenScale interpolates complete image-token vectors obtained from low-resolution generation, rather than individual positions on the unpacked latent grid, and re-noises the expanded representation to initialize high-resolution sampling. During an initial denoising window, selective reference fusion corrects clean predictions in spatially coherent high-deviation regions, leaving unselected tokens unchanged and imposing no further reference correction afterward. This design supports structural consistency while allowing later denoising to synthesize details beyond the low-resolution content. Extensive experiments demonstrate that TokenScale substantially improves HR image generation quality, outperforms state-of-the-art training-free methods, and can be integrated with positional-encoding-based approaches for further gains.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Flow Model, Image generation, High resolution, Diffusion Transformer, Training-free

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_eMwFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TokenScale studies training-free resolution expansion for pretrained DiT-based flow models. Its main methodological choices are to interpolate whole patch-token vectors from a completed low-resolution sample and to correct selected regions against a fixed reference during a short initial denoising window. The selection rule uses smoothed prediction–reference distances in the input-projected token space, followed by quantile thresholding and connected-component filtering. The corrected clean prediction is incorporated into the current latent state while preserving its implicit residual, after which the velocity is re-evaluated. The paper reports improved perceptual quality on FLUX.1[dev], with qualitative extensions to Qwen-Image and UltraImage.

Critical strengths

Whole-patch expansion is a concrete and implementable design choice.

The complete pipeline shows promising practical results.

Critical weaknesses
  1. The independent contribution of whole-token expansion is not established. The tokens being interpolated are obtained immediately after an affine input projection. As the paper's own equivalence shows, the projection commutes with spatial interpolation, so expansion can be implemented entirely by resampling patchified VAE latents. The methodological distinction is therefore the spatial grouping used during interpolation. This could be useful, but it does not by itself establish the claimed preservation of semantic information. The quantitative ablation mainly contrasts the proposed initialization with a collapsed configuration, while the matched comparison with ordinary latent-grid interpolation is limited to a qualitative example. It remains unclear how much of the gain comes from preserving patch coordinates and how much comes from inheriting any reasonable low-resolution initialization. A benchmark comparison of ordinary latent interpolation, pixel-upsample-and-re-encode initialization, and patch-grid expansion, using the same base sample, noise, restart schedule, and extrapolation toolkit, is needed to establish this first contribution.

  2. The selective fusion rule has not been shown to identify the regions where correction is beneficial. The mask selects relatively large, spatially coherent deviations from a fixed low-resolution-derived reference. Both structural drift and legitimate new high-resolution detail can produce such deviations, and a quantile threshold continues to select a roughly fixed fraction even when the absolute discrepancy is small. Smoothing and connected-component filtering make the mask coherent, but coherence alone does not demonstrate that those regions should be pulled toward the reference. The current ablation measures the effect of adding the complete fusion procedure; it does not separate the value of the projected-token distance, the spatial mask, and the early reference correction itself. A comparison with uniform early fusion and matched-area control masks, with correction strength and network-evaluation budget accounted for, would show whether spatial selection adds value. Comparing projected-token distance with raw patch distance would further test whether the token-space metric is an essential part of the method.

  3. The contribution of the combined sampling strategy is insufficiently distinguished from existing early-guidance cascades. DiffuseHigh already confines structural guidance to an early stage, and HiFlow reduces its guidance strength over time. The temporal release of reference constraints therefore has a direct precedent. TokenScale needs to establish the added value of coupling its patch initialization with spatial selection. The reported comparisons use the competing methods' official settings, while TokenScale also uses positional and attention-balancing techniques adopted from prior work. They do not include a simple, matched FLUX cascade with ordinary initialization and tuned early global fusion. The parameter examples show that excessive guidance can hurt, but do not establish that the proposed combination outperforms a well-tuned simpler strategy. The 4K UltraImage combination disables fusion, so that result supports expansion compatibility rather than the added value of the complete two-module design. A matched comparison that separates initialization, spatial selection, and guidance duration is necessary to determine what the overall method contributes beyond these existing ingredients.

Questions for authors
  1. How does patch-grid expansion compare with ordinary latent-grid interpolation and pixel upsampling followed by VAE re-encoding when the low-resolution sample, added noise, restart schedule, positional/attention settings, and sampling budget are identical? Please report this comparison over a shared evaluation set rather than selected examples.

  2. With the same initialization and early window, does the proposed mask outperform uniform fusion and matched-area control masks? Please account for the amount of reference correction and the extra velocity evaluations. How much is gained by the projected-token distance over raw patch distance, and by smoothing/component filtering over quantile selection alone?

  3. How much of the final improvement remains over a simple FLUX cascade with tuned early global guidance, or an appropriately tuned/early-stopped HiFlow? Can initialization, spatial selection, and guidance duration be varied separately to show which combination is necessary?

  4. Do the components retain their contributions on Qwen-Image and when paired with UltraImage? In particular, can the full combination be compared with expansion alone under matched quality and total inference cost?

4 comments
AuthorsOpenReview ↗

Part 1 (W1 & Q1): A Matched Benchmark of Alternative Expansion Strategies: Isolating the Effect of Patch-Grid Token Expansion

Thank you for the suggestion. We ran this benchmark on the shared 1,000-prompt 4K evaluation, comparing unpacked latent-grid interpolation, pixel upsampling followed by VAE re-encoding, and packed patch-grid expansion. Within each fusion setting, the runs matched the source sample, injected noise, restart schedule, auxiliary positional and attention settings, and denoiser evaluation budget.

FusionMetricLatent gridPixel + VAEPatch grid (TE)
OffFID ↓61.1361.5260.45
OffFGResQ ↑0.57680.58640.6004
OffKID ↓0.0074050.0077290.007332
OffCLIP Score ↑33.8334.3033.62
OffCLIP-IQA ↑0.60540.40950.7223
OnFID ↓59.6159.8559.15
OnFGResQ ↑0.58410.58100.6024
OnKID ↓0.0066590.0074310.007200
OnCLIP Score ↑34.2034.2234.31
OnCLIP-IQA ↑0.62480.41210.7350

The comparison shows that patch-grid expansion provides an effective initialization. The advantage holds with fusion both enabled and disabled.

AuthorsOpenReview ↗

Part 2 (W2 & Q2): Ablation Study of Different Mask Strategies and Distance-Metric Diagnostics

We compared all-token fusion, an area-matched random mask, quantile thresholding without smoothing or component filtering, and the full structural mask. The runs used the same patch-grid initialization, early fusion window, blend weight per selected token, and re-evaluation protocol, one extra velocity evaluation per fusion step, so all four conditions incur the same number of network evaluations. The structural and quantile-only masks each covered about 40% of the tokens, the random control matched this area, and all-token fusion covered the full map. The mask changes only where the correction is applied, not the evaluation count.

MetricAll-tokenRandomQuantile onlyStructural
FID ↓60.5159.9259.2459.15
FGResQ ↑0.60110.59410.60030.6024
KID ↓0.0076650.0077370.0074320.007200
CLIP Score ↑34.4334.2434.2934.31
CLIP-IQA ↑0.72660.71660.72990.7350

The results show that the structural mask performs best on most metrics. It corrects only the spatially coherent regions with the largest deviations, which stabilizes the structure while leaving the remaining regions free to synthesize new detail. New detail can also produce large deviations, so we use the deviation map only as a heuristic for candidate regions, requiring coherence on top of magnitude.

On 100 prompts, the mean Pearson correlations between raw and projected distances were 0.912 before smoothing and 0.937 after smoothing, and the mean final-mask IoU was 0.880, with 39.6% coverage using raw patch distances and 39.7% using projected distances. The two distances are highly correlated, so the choice has a very small impact on generation. For computational efficiency, we compute the distances directly on the patched latents.

AuthorsOpenReview ↗

Part 3 (W3 & Q3): Simple-Cascade Baseline, Factorized Decomposition, and Early-Stopped HiFlow

DiffuseHigh uses wavelet-based structural guidance early in denoising, and HiFlow gradually reduces its flow-alignment weights. TokenScale combines patch-grid initialization with spatially selective correction toward a fixed reference. To test the added value of this combination, we ran the simple matched cascade you describe, ordinary latent-grid initialization with early all-token fusion:

MetricSimple cascade (Latent + All-token)TokenScale
FID ↓60.8359.15
FGResQ ↑0.59210.6024
KID ↓0.0073210.007200
CLIP Score ↑33.2134.31
CLIP-IQA ↑0.67320.7350

The simple cascade is worse on all five metrics under matched settings. We also varied the three components separately around the TokenScale configuration.

Initialization, with structural fusion fixed:

MetricLatent + TFPixel + TFTE + TF (ours)
FID ↓59.6159.8559.15
FGResQ ↑0.58410.58100.6024
KID ↓0.0066590.0074310.007200
CLIP Score ↑34.2034.2234.31
CLIP-IQA ↑0.62480.41210.7350

Spatial selection, with patch-grid initialization fixed:

MetricAll-tokenRandomQuantile onlyStructural
FID ↓60.5159.9259.2459.15
FGResQ ↑0.60110.59410.60030.6024
KID ↓0.0076650.0077370.0074320.007200
CLIP Score ↑34.4334.2434.2934.31
CLIP-IQA ↑0.72660.71660.72990.7350

Guidance duration, with initialization and selection fixed, measured on the 200-prompt cohort; columns show the fraction of the post-restart schedule during which fusion is applied:

Metric6.7%10%13.3%26.7%40%60%80%
FID ↓171.50166.35171.86168.52168.63172.59171.92
FGResQ ↑0.60210.60330.59870.59040.57420.55450.5623
KID ↓0.0072230.0072120.0079980.0075970.0078910.0078760.007874
CLIP Score ↑34.2434.2634.2134.0634.1434.1034.10
CLIP-IQA ↑0.72090.73420.75120.71360.60110.52970.4958

These comparisons estimate separate effects under the tested settings.

A HiFlow variant that stops guidance early while completing the denoising schedule reached, at 4K, KID 0.0086, CLIP Score 34.66, CLIP-IQA 0.602, and FGResQ 0.5934, showing no significant change relative to the official-settings metrics we measured in the main text. Early stopping does not materially improve HiFlow, because its setting resamples after 66% of the schedule, leaving little room for released guidance to benefit the model. FID is omitted since it is a biased estimator at this sample size. Appendix G additionally compares attention spectra using official and restart-matched HiFlow settings.

AuthorsOpenReview ↗

Part 4(Q4) Component Contributions on Qwen-Image and UltraImage

Qwen-Image. On Qwen-Image, we found during experiments that, because the model is strong, TE alone already generates good images, so TF's contribution there is limited; the paper includes the corresponding clarification.

UltraImage. In the coupling experiment with FLUX as the base model, UltraImage alone cannot generate 4K images; adding TE greatly extends its capability. TE also accelerates generation: instead of 40 steps, the super-resolution stage can restart from step 10 or 20 with the low-resolution upsampled initialization, saving about 25%–50% of the super-resolution time steps. TF's effect is weakened after coupling, and we will report this honestly in the paper.

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

This paper introduces TokenScale, a training-free method for high-resolution image generation. The approach initializes high-resolution sampling via token expansion, constrains the structure during the early stages of denoising through selective reference fusion, and continues the generation process by maintaining consistent implicit noise across state updates. The authors evaluate the method at 2K and 4K resolutions and provide additional visual examples along with inference code.

Critical strengths
  1. This method requires no additional training and can leverage pre-trained models to perform progressive, high-resolution generation, offering practical utility.
  2. Experiments cover a range of baselines and evaluation metrics, with strong performance in terms of perceptual quality; close-up views in the supplementary material further demonstrate clear details and well-preserved overall structure.
  3. The paper includes component ablation studies, controlled qualitative comparisons, and examples using different random seeds, while also explaining the equivalence between token manipulation and patch-based latent representation.
Critical weaknesses
  1. The quantitative evidence isolating the benefit of the proposed expansion strategy remains limited. The equivalence to patchified-latent interpolation in Appendix C.3 is not itself a concern. However, a matched quantitative comparison against standard latent-grid interpolation, with the restart schedule, fusion procedure, and auxiliary extrapolation settings held fixed, would help establish the specific benefit of the proposed initialization. The controlled qualitative comparison is useful but does not fully resolve this question.
  2. There are discrepancies between the submitted code and the descriptions provided in the paper that require clarification. On page 5 and in Algorithm 1, the paper describes constructing the fusion reference via a "VAE decoding → bilinear upscaling in pixel space → VAE encoding" pipeline, whereas the code performs bicubic upscaling directly on the latents. Furthermore, the paper constructs the fusion mask based on the distances between tokens after input projection, while the code utilizes the distances between unprojected patch latents. Additionally, the default settings in the code—specifically the number of fusion iterations, the fusion strength, and the stages at which fusion is enabled—do not entirely align with the experimental setup described in the paper. Differences in default hyperparameters may be explained by experiment-specific configurations. However, the differences in reference construction and mask computation leave it unclear whether the submitted implementation reproduces the method used for the reported results.
Questions for authors
  1. Could you provide additional quantitative baseline results—using strictly matched sampling and auxiliary settings—to help disentangle the contributions of token expansion itself from those of other inference techniques?
  2. There are discrepancies between the paper’s description and the supplementary code regarding the construction of the fused reference, the calculation of the fusion mask, and the fusion parameters. Please clarify which specific implementation was used to obtain the results presented in Table 1.
4 comments
AuthorsOpenReview ↗

Part 1 (W1 & Q1): Matched initialization ablation with fusion on and off

Thank you for your suggestion.We compared three initialization operators on the same 1,000-prompt 4K evaluation: unpacked latent-grid interpolation, pixel upsampling followed by VAE re-encoding, and packed patch-grid expansion. Within each fusion setting, the runs matched the input sample, injected noise, restart schedule, positional encoding, attention settings, and denoiser evaluation budget; disabling fusion isolates the contribution of token expansion itself from the other inference techniques. The comparison shows that patch-grid expansion provides an effective initialization under fixed extrapolation and sampling settings.

FusionMetricLatent gridPixel + VAEPatch grid (TE)
OffFID ↓61.1361.5260.45
OffFGResQ ↑0.57680.58640.6004
OffKID ↓0.0074050.0077290.007332
OffCLIP Score ↑33.8334.3033.62
OffCLIP-IQA ↑0.60540.40950.7223
OnFID ↓59.6159.8559.15
OnFGResQ ↑0.58410.58100.6024
OnKID ↓0.0066590.0074310.007200
OnCLIP Score ↑34.2034.2234.31
OnCLIP-IQA ↑0.62480.41210.7350

The comparison shows that patch-grid expansion provides an effective initialization under fixed extrapolation and sampling settings.

AuthorsOpenReview ↗

Part 2 (W2 & Q2): Clarification of the Table 1 implementation

Thank you for this careful inspection. We clarify each point below.

Which implementation produced Table 1. Table 1 used the reference construction of the paper: the low-resolution latent is decoded by the VAE, bilinearly upsampled in pixel space, and re-encoded. The fusion-mask distances were computed between unprojected patch latents, as in the submitted code.

Why unprojected patch-latent distances. The projected-token and raw patch-latent distances are highly correlated and barely change the mask. Measured on 100 prompts:

StatisticMean95% CI
Pearson correlation, raw distances0.9118[0.9066, 0.9168]
Pearson correlation, smoothed distances0.9369[0.9315, 0.9419]
Threshold-mask IoU0.8406[0.8222, 0.8593]
Final-mask IoU0.8797[0.8604, 0.8987]
Mask coverage, latent vs. token39.6% vs. 39.7%—

The two choices produce nearly identical fusion regions, so the reported results are unaffected. We kept the raw patch-latent distance because it skips the input projection at every fusion step, which is faster. The paper uses the projected-token form because the fusion rule is defined in token space (Eq. (6)); the input projection is affine, so the two views are equivalent up to this metric choice. The revision will state both forms.

The reference construction and default settings in the code. The bicubic-on-latents reference and the differing fusion defaults are leftover variants from our earlier design exploration, not the configuration behind Table 1, which follows Sec. 4.1. We apologize for the confusion and will update the released code so its defaults match the paper.

Reviewer_zMnFOpenReview ↗

Response to Author

Thanks for your detailed response, which has addressed my concerns. I will raise my score to 3.

Thank you for your careful reading and positive assessment. We are glad that our clarification of the implementation details underlying Table 1 addressed your concerns. We will ensure that the released code is consistent with the revised manuscript.

Reviewer_rJgsOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a training-free method called TokenScale.

The authors first investigate why DiT-based flow models like FLUX.1-dev break down past their training resolution. Specifically this failure happens at the token level. Increasing the resolution from 1k to 4k results in 16 times more tokens, while a fixed prompt keeps the same number of text tokens. In joint attention each image token therefore puts much less weight on the prompt, so per-word attention maps dissolve into noise at 4k resolution.

To solve the problem, TokenScale adopts a cascading way from 1k 2k to 4k. At each stage, the final low-resolution latent is enlarged by directly interpolating the packed patch tokens so that the values from different positions inside each patch will not be mixed. Then they renoised the enlarged latent and high-resolution sampling starts a quarter of the way into the schedule. The first few highres steps is also guided by the low-res result upscaled in pixel space. The guidances only applies in large connected regions where the prediction deviates most. The authors smooths the deviation map with a 3x3 filter with certain thresholds to find the most deviated positions, then let the remaining denoise without guidance.

The results are good over multiple metrics and beats training free methods like DemoFusion DiffuseHigh. I-Max and HiFlow etc.

Critical strengths
  1. The method has faster per-step time than other methods i.e. I-Ma and HiFlow which shows its efficiency

  2. Interpolating whole patch tokens is a useful observation for patchified DiTs. The comparison with unpacked interpolation in Appendix B shows a clear difference in sharpness.

  3. The method is universal across models. In the paper the authors also shows the results on Qwen-Image which is also good and it can serve as the initialization for UltraImage.

Critical weaknesses
  1. I think the authors final results is somewhat similar to a diluted image-text attention that might be the reason why the CLIP score is low. I hope the authors can address this, and I would like to know if simply duplicate the text like I-Max will help to increase the CLIP score.

  2. The packed latent upscale is studied in "Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers", if you take a closer pass of their code. Nevertheless the discussion here is good

  3. I think the authors should do comparisons with some training based method like Diffusion 4K, UltraImageGen, URAE, and attention based method like CLEAR etc.

Questions for authors
  1. I would like to se some results on timestep distilled models (i.e. "turbo" version, 4/8 step generation)

  2. If image–text attention dilution is the main cause, why is TokenScale’s CLIP Score the lowest at both resolutions? (similar to the first point in Weaknesses section)

I would incline to support the paper to be accepted if my questions are addressed.

8 comments
AuthorsOpenReview ↗

Part 1 (Q1): Results on time-distilled models

Thank you for this question. We have evaluated TokenScale on two time-distilled models under the same protocol: FLUX.1-Turbo-Alpha and Qwen-Image-2.1-Turbo, both with 8-step generation, the 4K stage restarting from the third step, and fusion applied in the early post-restart steps.

MetricFLUX.1-Turbo-AlphaQwen-Image-2.1-Turbo
FGResQ ↑0.59580.6083
KID ↓0.0094980.007098
CLIP Score ↑34.6535.12
CLIP-IQA ↑0.63130.7453

End-to-end inference takes 253.4 s per 4K image with 28.7 GB peak memory on FLUX.1-Turbo-Alpha, and 196.8 s with 24.8 GB on Qwen-Image-2.1-Turbo. The pipeline transfers to distilled schedules without other changes. We have added a corresponding part in Sec. 4.5 of the paper to present these distilled-model results.

AuthorsOpenReview ↗

Part 2 (W1 & Q2): Attention dilution, CLIPScore, and text-duplication control

Thank you for raising this point.

Our analysis focuses on image-to-text attention dilution within MMDiT during generation (Fig. 2). TE directly addresses this issue, as shown by the attention maps in Figs. 2 and 10. In contrast, CLIPScore is computed using a separate CLIP model on the generated image and text [1]. A lower CLIPScore therefore does not necessarily indicate that attention dilution persists within MMDiT.

These small differences may partly reflect the limited sensitivity of CLIPScore to high-resolution details, as its image preprocessing involves resizing and center cropping [2]. The differences are only 0.11 compared with I-Max and 0.02/0.23 compared with HiFlow (Table 1). Since CLIPScore does not directly evaluate high-frequency details, we also report CLIP-IQA (0.742/0.735) and FGResQ to provide a more comprehensive assessment of high-resolution image quality.

Regarding text duplication (I-Max style): we agree it is a relevant control for the image/text balance in joint attention, and we ran it under our matched settings. CLIP Score changed from 34.61 without duplication to 34.44 with duplication, a difference of −0.17, so text duplication does not improve CLIPScore.

References

[1] Hessel et al., CLIPScore: A Reference-free Evaluation Metric for Image Captioning, EMNLP 2021. https://aclanthology.org/2021.emnlp-main.595/

[2] OpenAI CLIP, official preprocessing implementation. https://github.com/openai/CLIP/blob/main/clip/clip.py

AuthorsOpenReview ↗

Part 3 (W2): Relation to RALU's packed-latent upsampling

Thank you very much for pointing this out, and for the kind note on our discussion. We acknowledge that we overlooked RALU (Jeong et al., CVPR 2026) [1], which performs packed-latent upsampling by interpolating the patchified latent sequence along the token axis. We will add the citation in our paper。

Our motivation for adopting this operation, however, differs from that of RALU.

  • RALU focuses on reducing the inference cost of DiTs while addressing two artifacts introduced by naive latent upsampling: aliasing and distribution mismatch. Its packed-latent nearest-neighbor duplication is selected so that the noise-covariance computation of NT-Matching stays tractable (their Tab. S5 compares bilinear/bicubic/Lanczos/NN and chooses NN).
  • TokenScale instead focuses on attention dilution at high resolutions, where image-to-text attention weakens and per-word attention becomes spatially dispersed (Fig. 2). This motivates transferring the organized semantic–spatial structure carried by the low-resolution tokens into high-resolution sampling through whole-token interpolation, followed by selective early reference correction.

Although both methods interpolate whole patch vectors rather than individual latent positions, their specific mechanisms differ:

  1. Interpolant. RALU duplicates patch vectors with identity weights (nearest along the token axis; in the unpacked view this is the 4×4-ones block covariance of their Eq. (4)). TE mixes neighboring patch vectors with bilinear weights on the token grid.
  2. Scope and trajectory. RALU upsamples only the edge-selected top-r patches mid-trajectory and continues the same trajectory with matched correlated noise. TE expands the complete low-resolution token grid and re-noises it to a restart noise level, initializing a new high-resolution trajectory.
  3. Purpose and evaluation. RALU targets spatial acceleration (up to 7× on FLUX.1[dev]) with quality preservation; TokenScale targets perceptual quality at 2K/4K and is evaluated against both training-free and training-based methods.
  4. Coupled module. Our selective token fusion (TF) uses spatially coherent, adaptive-threshold correction during early denoising, which is not present in RALU.

Token expansion is defined in token space as bilinear mixing of whole token vectors, and Appendix C.3 presents an equivalent packed-latent implementation of this operation. Both methods perform token-wise upsampling, but they differ in their motivation, interpolation strategy, and sampling procedure. We will cite RALU in the revised manuscript and clarify these differences, and we will distinguish the token-space formulation from its packed-latent implementation in Appendix C.3.

Reviewer_oW4aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TokenScale, which is a training-free framework for high-resolution image generation that combines expansion of low-resolution image tokens with selective reference fusion during early denoising. The motivation, method, and experiment results are clear to show the improvement compared with others.

Critical strengths

The proposed method is simple, requires no additional training. It integrates with pretrained DiT-based models. Experiments demonstrate improved perceptual quality and local detail at 2K and 4K with effective comparisons.

Critical weaknesses
  1. The claimed advantage of token-space expansion is insufficient. As the author showed that token expansion is equivalent to interpolating patchified latents. I wonder whether the gains are from better semantic preservation as the author claimed, or only from a simple different spatial resampling. Appendix B provides only a case comparison with conventional latent-grid interpolation. I think a quantitative comparison across the evaluation dataset is necessary.
  2. The whole method relies on many manually set hyperparameters, like some thresholds pp and a{min}a_\{min\} in Algorithm 2, and time window nsn_s, FF in Algorithm 1. This makes the method dependent on multiple tuning to achieve the best results. Thus the end-to-end time cost comparison is more valuable than per-step time.
Questions for authors
  1. Can you provide the comparison between the proposed method and latent-grid interpolation across the whole dataset or more cases?
  2. Is there an adaptive strategy to determine the hyperparameters included in the algorithm?
5 comments
AuthorsOpenReview ↗

Part1 (W1 & Q1): Dataset-Level Comparison of Initialization Strategies against Latent-Grid Interpolation

We agree that interpolation on a packed patch grid should be evaluated against ordinary latent-grid resampling across the dataset. On the shared 1,000-prompt 4K evaluation, we compared unpacked latent-grid interpolation, pixel upsampling followed by VAE re-encoding, and patch-grid expansion. Within each fusion setting, the initialization comparison matched the source sample, noise, restart schedule, fusion settings, positional and attention settings, and denoiser evaluation budget.

FusionMetricLatent gridPixel + VAEPatch grid (TE)
OffFID ↓61.1361.5260.45
OffFGResQ ↑0.57680.58640.6004
OffKID ↓0.0074050.0077290.007332
OffCLIP Score ↑33.8334.3033.62
OffCLIP-IQA ↑0.60540.40950.7223
OnFID ↓59.6159.8559.15
OnFGResQ ↑0.58410.58100.6024
OnKID ↓0.0066590.0074310.007200
OnCLIP Score ↑34.2034.2234.31
OnCLIP-IQA ↑0.62480.41210.7350

The comparison shows that patch-grid expansion provides an effective initialization under fixed extrapolation and sampling settings.

AuthorsOpenReview ↗

Part 2 (W2 & Q2): Hyperparameter robustness and end-to-end inference cost(1)

(1) No per-prompt tuning

Thank you for raising this concern. All 1,000 prompts in our evaluation (the Diffusion-4K prompt set) use exactly the same hyperparameter configuration, without manual adjustment or grid search for individual prompts. Specifically, each stage uses 32 sampling steps, the high-resolution stage restarts at a ratio of 1/4, and Token Fusion is applied only during the first 3 steps after restart with a fixed strength of β=0.2\beta=0.2. The same configuration is used throughout the 1K→2K→4K generation process.

(2) Adaptive selection and hyperparameter robustness

The adaptivity of TokenScale mainly lies in spatial token selection. Algorithm 2 smooths the token-deviation map with a 3×33\times3 box filter, thresholds at the 60th percentile of the current deviation distribution, and removes connected components smaller than 8 tokens. The actual threshold and resulting mask thus adapt to the current generation, while the quantile, minimum area, restart position, fusion steps, and fusion strength are fixed across prompts. As shown in Fig. 13, the fused-token ratio remains around 40% across cases.

Due to time and compute constraints, the rebuttal-stage sensitivity sweeps were run on 200 prompts, varying the quantile, the fusion strength, the fusion steps, and the restart position with all other settings held fixed, to observe trends and robustness.

Quantile pp:

Metricp=0p=0.4p=0.6p=0.8p=1
FID ↓170.17168.32166.35167.78168.12
FGResQ ↑0.60110.59980.60330.59090.5803
KID ↓0.0076650.0076230.0072000.0076710.007292
CLIPScore ↑34.4334.4034.2634.2834.22
CLIPIQA ↑0.72660.74360.73420.71990.6857

Fusion strength β\beta:

Metricβ=0.05β=0.1β=0.2β=0.4β=0.5β=0.7β=1
FID ↓167.33168.18166.35165.82165.17168.29169.78
FGResQ ↑0.58550.58300.60330.60160.59630.59550.5802
KID ↓0.0067130.0079330.0072000.0067400.0064590.0085580.008381
CLIPScore ↑34.2234.1934.2634.3134.1534.0634.04
CLIPIQA ↑0.71000.71610.73420.76180.75230.69800.6558

Fusion steps (fraction of the post-restart schedule with fusion):

Metric6.7%10%13.3%26.7%40%60%80%
FID ↓171.50166.35171.86168.52168.63172.59171.92
FGResQ ↑0.60210.60330.59870.59040.57420.55450.5623
KID ↓0.0072230.0072120.0079980.0075970.0078910.0078760.007874
CLIPScore ↑34.2434.2634.2134.0634.1434.1034.10
CLIPIQA ↑0.72090.73420.75120.71360.60110.52970.4958

Restart position at 4K (40-step schedule; fraction of the schedule skipped):

Metric10%20%25%40%50%80%
FID ↓172.34165.73166.35172.87170.93180.66
FGResQ ↑0.58150.59240.60330.60190.58330.4472
NFE ↓115.00107.00103.0091.0083.0059.00
KID ↓0.0077130.0068190.0072000.0102060.0089870.018499
CLIPScore ↑33.5334.5734.2634.0233.8132.81
CLIPIQA ↑0.57320.70110.73500.72580.72100.4543

The default configuration and the recommended ranges reported in the paper both lie within the stable regions of these sweeps.

AuthorsOpenReview ↗

Part 2 (W2 & Q2): Hyperparameter robustness and end-to-end inference cost(2)

(3) End-to-end inference cost

We agree that per-step inference time alone does not fully reflect the computational cost of multi-stage generation, particularly because each Token Fusion operation requires an additional velocity evaluation. We therefore measured the complete 1K→2K→4K cascade on the same [GPU model]:

MethodEnd-to-end 4K (s/image)
I-Max832
HiFlow423
Ours694
Ours (FLUX.1-Turbo-Alpha, 8-step)253.4
Ours (Qwen-Image-2.1-Turbo, 8-step)196.8

The measurement includes the complete cascade, VAE operations, and additional velocity evaluations, while excluding model loading. Specifically, TokenScale requires 32 velocity evaluations for the initial 1K stage, and 24 regular evaluations plus 3 additional evaluations for each subsequent high-resolution stage, 86 evaluations in total for 1K→2K→4K.

Our end-to-end time is longer than HiFlow's, and we see this as a deliberate quality-speed trade-off. Our per-step time remains the lowest of the FLUX-based methods, 26.53 s vs. 30.23 s for HiFlow and 35.76 s for I-Max (Table 3); the end-to-end difference comes from the number of evaluations. Fig. 24 shows these extra evaluations are well spent: increasing sampling steps consistently improves TokenScale's quality, whereas HiFlow degrades as persistent image-domain guidance accumulates. For speed-sensitive scenarios, the time-distilled variants above are faster than both HiFlow and I-Max.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary Comment to All Reviewers

We thank all reviewers for their constructive feedback. We have revised the manuscript and conducted additional experiments and analyses to address the reviewers' concerns. The main updates are summarized below.

Clarified the contribution and scope of token expansion. We added a controlled comparison of three initialization strategies on the same 1,000 prompts at 4K resolution, both with and without token fusion. (oW4a: W1/Q1; zMnF: W1/Q1; eMwF: W1/Q1)

Added mask ablations. We compared all-token fusion, an area-matched random mask, quantile-only thresholding, and the full structural mask, controlling for mask coverage where applicable and using the same evaluation budget. We also conducted a diagnostic experiment on 100 prompts, showing that projected-token and raw patch distances are highly correlated. (eMwF: W2/Q2; zMnF: W2/Q2)

Added comparisons with simpler alternatives. We evaluated a matched cascade using conventional latent-grid initialization with early all-token fusion, as well as an early-stopped HiFlow variant. The simple cascade performs worse on all five metrics, while early stopping provides little improvement in HiFlow's performance. (eMwF: W3/Q3)

Extended the comparison with training-based methods. We added Diffusion-4K, UltraImageGen, URAE, and CLEAR to Table 1 under the same evaluation protocol, along with qualitative comparisons in Fig. 18. (rJgs: W3)

Evaluated TokenScale on distilled models. We added results on FLUX.1-Turbo-Alpha and Qwen-Image-2.1-Turbo, both using 8 sampling steps, in Sec. 4.5. (rJgs: Q1; eMwF: Q4)

Clarified the relationship between attention dilution and CLIPScore. We conducted a matched text-token duplication experiment, showing that I-Max-style duplication does not improve CLIP similarity. We also explained why internal image-to-text attention dilution and the external CLIPScore metric measure different aspects of generation. (rJgs: W1/Q2)

Clarified hyperparameter selection and sensitivity. All 1,000 prompts are evaluated using the same fixed hyperparameters, without per-prompt tuning. We also conducted sensitivity analyses on 200 prompts, varying the quantile threshold, fusion strength, fusion duration, and restart position. The results show that our default settings and recommended ranges fall within stable regions. (oW4a: W2/Q2)

Clarified implementation details. We specified the pixel-space reference construction and the use of unprojected patch-latent distances in the experiments reported in Table 1. We will also ensure that the released code uses the same default settings as the manuscript. (zMnF: W2/Q2)

Added end-to-end inference times. We reported the total inference time for 4K generation, including results for the distilled models. (oW4a: W2/Q2)

Added a comparison with RALU. We included the missing citation and clarified the differences in motivation, interpolation strategy, and sampling procedure. (rJgs: W2)

Clarified the combination with UltraImage. We specified that token fusion is disabled when TokenScale is combined with UltraImage. (eMwF: Q4)

We sincerely appreciate the time and effort the reviewers have devoted to evaluating our work, and we hope the revised manuscript and additional results address your questions and concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.