acceptodds
Under review as a conference paper at ICLR 2027

Grow, Don't Prune: Compute-Efficient Training of Diffusion Transformers

Abstract

Diffusion Transformers (DiTs) achieve strong generation quality but are expensive to train and evaluate. Existing dynamic-width methods reduce inference cost by pruning a pretrained dense model, leaving its pretraining cost unchanged. We propose GrowDiT, a dynamic-growth training strategy that starts from a narrow network and progressively activates attention heads and MLP channel groups across diffusion noise levels. A gradient-based need score selects which units to activate under an increasing compute budget. By reducing the FLOPs of early training updates, GrowDiT reallocates the saved computation to additional optimization steps. On ImageNet-256 and FFHQ-256, under a fixed training-FLOP budget and approximately matched inference FLOPs, GrowDiT consistently outperforms fixed-width, random-growth, and dense-training baselines across DiT-S/2, DiT-B/2, and DiT-L/2. On the 130M-parameter DiT-B/2, it improves FID by 7% over fixed-width training while using the same nominal training and inference FLOPs. Matching budgets by wall-clock rather than FLOPs gives a smaller 5.3% on DiT-B/2 and 1.1% on DiT-S/2. GrowDiT also outperforms depth-growth configurations and a DyDiT-inspired train-then-prune baseline. These results suggest that allocating width during training is a promising approach to improving the training efficiency of diffusion transformers.

Primary area:
generative models
Keywords:
diffusion transformers, efficient training, network growing, dynamic architectures

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cm3QOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes GrowDiT, a method to improve the training efficiency of Diffusion Transformers. The method starts with a smaller model and gradually activates more attention heads and MLP groups. It uses gradient-based scores to decide which units to activate at different noise levels. Under similar compute budgets, GrowDiT achieves better FID scores than fixed-width training, random growth, and train-then-prune. However, the improvement is relatively small when comparing actual training time rather than FLOPs.

Critical strengths

GrowDiT avoids the computational waste of training a full model before pruning it and achieves better generation quality under a fixed training budget. This idea has potential practical value for training generative models with high computational costs. The paper finds that different noise levels require different amounts of model capacity, with lower noise levels needing more computational resources. The capacity allocation learned by the model is also consistent with the theoretical requirements. This suggests that the gradient-based score can identify where more computation is needed to some extent, rather than simply increasing the model width.

Critical weaknesses

The novelty of this method is relatively limited, as gradient-guided network growth and timestep-dependent compute allocation have both been explored in previous work. In the experiments, Random Growth already achieves good results, while the additional improvement from the Need Score is relatively small. Also, the improvement in actual training time is much smaller than the FLOPs-based improvement.

Questions for authors

The Need Score is computed using first-order gradients at gate = 0, but many inactive Attention Heads and MLP Groups may not have been sufficiently trained. Why can their gradient magnitudes reflect how useful they will be after training? Also, gradient magnitudes may be affected by the scale of unit outputs. How do the authors account for this?

GrowDiT combines progressive growth with noise-dependent capacity allocation, but the paper does not clearly separate their contributions. What if we directly use the final allocation of Attention Heads and MLP Groups learned by GrowDiT and train a fixed architecture from scratch? Could it achieve similar results? If so, what is the benefit of progressive growth itself?

3 comments

We thank the Reviewer for their careful evaluation and for their questions, which have helped us state more precisely what the need score contributes.

Question 1 (usefulness after training). Weights are shared across noise levels and only the gates are per noise level, so most inactive candidates are units already trained at other noise levels. Most growth decisions therefore ask whether a trained unit should also serve another noise level. Although the need score is a local, first-order estimate (Proposition 1), our experiments show that it provides a useful signal for learning noise-dependent capacity allocations over the course of training. Two findings support this conclusion. First, without access to a fully trained model, GrowDiT recovers the same overall allocation trend as the routers learned by DyDiT on a pretrained dense model (Fig. 4a). Second, the learned allocation consistently outperforms a hand-designed allocation, with the advantage becoming more pronounced for DiT-B (table below). The choice of units, in contrast, matters little. At low noise levels, where fine details are refined, most units end up active, so the choice of units mainly decides when a unit is activated, not whether it is (Appendix E, Table 11). The activation time of a unit, and hence the locality of the need score, has little effect on the final FID (Sec. 5.3, Table 5).

over random growthFID gainrelative gain
DiT-S, hand-set allocation0.751.1%
DiT-S, need-score allocation1.442.2%
DiT-B, hand-set allocation0.531.2%
DiT-B, need-score allocation1.714.0%

Gain over random growth on ImageNet-256 (1× budget), averaged over seeds. All three methods choose the units at random. Relative gain: gain divided by the FID of random growth.

Question 1 (scale of unit outputs). The need score measures the local sensitivity of the loss to gate activation under the current parameterization (Eq. 3). This sensitivity naturally depends on both the magnitude of the unit's output and its alignment with the loss gradient, since both determine the first-order loss variation when the gate is activated. We acknowledge that the score is parameterization-dependent. However, our objective is not to estimate an intrinsic, scale-invariant importance of each unit, but to identify beneficial gate activations within the current model. The empirical results support its effectiveness for this purpose.

Question 2. We trained the baseline suggested by the Reviewer (final allocation without growth; Sec. 5.3, Table 5). It takes the final allocation of heads and MLP groups from a GrowDiT run with the same seed and keeps it fixed throughout training. This architecture is trained from random initialization with the same budget. With the full budget its FID is still higher than that of GrowDiT (65.0 against 64.3 on DiT-S; table below), and progressive growth has two further benefits. (1) It finds the allocation within one run, whereas the fixed architecture needs a GrowDiT run first. (2) With a limited budget it also gives a lower FID: growth is ahead until about 75% of the budget (Appendix E, Fig. 6), because the early narrow model takes more steps for the same FLOPs.

DiT-SFIDgain
random growth65.6–
final allocation without growth65.00.8%
GrowDiT64.32.0%

FID-50k on DiT-S (ImageNet-256, 1× budget), averaged over seeds. Gain: FID reduction relative to random growth.

On the Novelty (Weakness). While GrowDiT shares the goal of noise-dependent capacity allocation with methods such as DyDiT, it addresses a fundamentally different learning problem. DyDiT determines which units to retain after training a full model, whereas GrowDiT learns how to allocate capacity across noise levels during training, without access to a fully trained network. Crucially, because weights are shared across noise levels, GrowDiT can evaluate whether units already trained at one noise level should be activated at another. This enables adaptive capacity allocation without the cost of first training a dense model. Gradient-guided growth such as GradMax (Evci et al., 2022) and Firefly (Wu et al., 2020) uses gradients to decide which units to add to a network that serves every input. GrowDiT instead uses its gradient-based need score to decide which noise levels each unit should serve, which gives an adaptive, noise-dependent allocation of capacity. This allocation yields a measurable advantage: in seed-matched experiments, it consistently outperforms fixed or unguided growth under constrained training budgets. The advantage is also interpretable: the need score automatically finds an allocation with more capacity at low noise levels, the same overall pattern as one learned from a fully trained model. The same mechanism also improves depth growth. We believe this gives a new way to train diffusion models and a reference for later work.

On the Gain over Random Growth (Weakness). The improvement over random growth is consistent across all nine seed-matched comparisons (Appendix D, Table 9) and across all three model scales. While the absolute FID reduction varies with model size, the relative improvement increases from 2.3% on DiT-S to 3.9% on DiT-L. More importantly, the additional ablations isolate the source of this improvement: the benefit arises primarily from the noise-dependent capacity allocation learned by the need score, rather than from the selection of individual units (Sec. 5.3).

GrowDiT over random growthFID gainrelative gain
DiT-S1.522.3%
DiT-B1.273.0%
DiT-L0.933.9%

Gain of GrowDiT over random growth on ImageNet-256 (1× budget), averaged over seeds. Relative gain: gain divided by the FID of random growth.

On the Wall-clock Time (Weakness). We recomputed the wall-clock comparison using FID averaged over seeds, consistently with Table 2, and with all growth-related operations included in the training time (Table 6; Appendix G). As expected, the practical gains are smaller than those measured under equal FLOP budgets. Nevertheless, GrowDiT achieves lower FID at equal wall-clock time on DiT-B and DiT-L, with improvements of 3.8% and 7.1%, respectively, while matching uniform training on DiT-S. Importantly, the gap between FLOP-based and wall-clock improvements becomes smaller for larger models, suggesting that the benefits of adaptive growth become increasingly relevant at scale. These results demonstrate practical training-time advantages beyond theoretical compute savings.

FID gain over uniformDiT-SDiT-BDiT-L
GrowDiT, same FLOPs2.1%6.0%9.6%
GrowDiT, same wall-clock time0.0%3.8%7.1%
wall-clock gain / FLOP gain0%63%74%

FID improvement over uniform on ImageNet-256 (Table 6), averaged over seeds as in Table 2. Same wall-clock time: every method trains for the time that uniform needs for the 1× FLOP budget.

On Presentation. We acknowledge the Reviewer's concerns about clarity. We have revised the presentation to better distinguish progressive growth from noise-dependent capacity allocation, clarify the role of the need score, and make the relationship with prior approaches more explicit. We have also reorganized the ablation discussion to isolate the contribution of each component.

Thank you for the detailed response. The additional experiments address my main concerns. I will raise my rating to 3.

Reviewer_6b7mOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GrowDiT starts with a narrow DiT and gradually activates attention heads and MLP groups across noise levels. Gradient-based scores guide growth, allowing more training steps within a fixed compute budget.

Critical strengths
  1. The method consistently improves FID across two datasets and three model sizes.
  2. Scoring overhead is included, and gains remain under matched wall-clock budgets.
Critical weaknesses
  1. The experiments do not fully separate progressive growth from noise-dependent allocation. A simple low-noise-first growth baseline would help show whether the gradient-based score is necessary.
  2. Similarly, the depth-growth comparison changes both the growth axis and the allocation strategy: only GrowDiT uses noise-dependent allocation. This limits the conclusion that width growth is preferable to depth growth.
Questions for authors
  1. The random-growth baseline supports the need score. How much of the gain comes from allocating more capacity to low-noise bins, versus selecting specific units within each bin? (Does simply allocate more capacity to low-noise bins?)
  2. Do higher-scoring heads actually give larger improvements after activation and a fixed number of training steps? (The proposed scoring is short-term metric, but does it help making decision for long term training?)
2 comments

We thank the Reviewer for their careful evaluation and for the suggested baselines, which have made the conclusions of the paper clearer.

Question 1 (also Weakness one). To answer this question, we added two baselines (Sec. 5.3, Table 5). Together, they show that the allocation of capacity across noise levels and layers is substantially more important than the identities of the individual units being activated. (1) The baseline need-score allocation with random units keeps the number of units that the need score gives to each layer at each noise level and chooses the units at random. Its FID is no higher than that of GrowDiT (64.1 against 64.3 on DiT-S), so with the same allocation, the identities of the individual units being activated matter little. (2) The baseline hand-set allocation tests whether simply giving more capacity to low-noise bins is enough: in every layer, the width falls linearly from full at the lowest noise level to about half at the highest, with random units. Both allocations improve on random growth in every seed-matched pair, but the need-score allocation gains about twice as much as the hand-set allocation on DiT-S and about three times as much on DiT-B (table below). The need score therefore identifies a more effective allocation than the tested low-noise-first heuristic, with a larger improvement observed on DiT-B.

over random growthFID gainrelative gain
DiT-S, hand-set allocation0.751.1%
DiT-S, need-score allocation1.442.2%
DiT-B, hand-set allocation0.531.2%
DiT-B, need-score allocation1.714.0%

Gain over random growth on ImageNet-256 (1× budget), averaged over seeds. All three methods choose the units at random. Relative gain: gain divided by the FID of random growth.

Question 2. The answer depends on the level at which the need score is used. (1) For individual units, a higher need score does not give a lasting advantage: with the same allocation, units chosen at random reach a similar final FID (Question 1). (2) For the allocation, the need score does guide the long-term decision: without access to a fully trained model, the allocation it finds is still better than both the hand-set allocation and random growth at the end of training (table above), shows the same overall pattern as the DyDiT routers of train-then-prune, which are learned from the whole trained model (Fig. 4a), and also works for noise-dependent depth growth (Weakness two).

Weakness two. The depth baseline grows depth as in SANA 1.5 (Xie et al., 2025), and its blocks serve every noise level, so it differs from GrowDiT in both the growth axis and the allocation. To separate the two, we added a variant, depth growth with the need-score allocation, which activates blocks per noise level (Sec. 5.3, Table 5). On DiT-S, its FID is 64.1, against 69.2 for depth growth without noise dependence (table below), so the gap comes from the allocation found by the need score, not from the growth axis. We therefore refine our claim: the need score also helps growth along the depth axis, not only along the width axis.

DiT-SFIDgain
random growth65.6–
depth growth without noise dependence69.2-5.6%
depth growth with need-score allocation64.12.3%

FID-50k on DiT-S (ImageNet-256, 1× budget), averaged over seeds. Gain: FID reduction relative to random growth.

Thanks for the response, I will keep my rating.

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

GrowDiT trains a diffusion transformer by starting with fewer active attention heads and MLP groups, then progressively activating additional units across noise-level bins. A gradient-based need score ranks candidate activations by their estimated benefit per unit of added compute. Experiments on ImageNet-256 and FFHQ-256 compare this approach with fixed-width training, random growth, depth growth and train-then-prune. The paper reports improvements under fixed training-FLOP budgets and approximately matched inference costs, with additional wall-clock-matched comparisons.

Critical strengths
  1. GrowDiT consistently improves FID over fixed-width training and random growth across three ImageNet model sizes and two FFHQ model sizes. The wall-clock-matched comparisons also show improvements over fixed-width training, although the gains are smaller than under FLOP matching.

  2. The growth procedure is clearly specified, and the random-growth baseline provides a useful comparison under the same growth schedule. The analysis of capacity allocation across noise levels helps explain the learned architectures. The paper also acknowledges limitations, including the gap between FLOP savings and practical runtime savings.

Critical weaknesses
  1. Uncertainty in the incremental benefit of need-score selection. At the 1× ImageNet budget, GrowDiT improves over random growth by 1.4, 1.0 and 0.7 FID on DiT-S/B/L. These consistent but modest gains are reported as two-seed means without variability estimates, making their robustness difficult to assess. The default DiT-S FID also differs between Tables 2 and 7 (63.8 versus 64.6). Figure 6 measures agreement with a higher-sample need-score estimate, rather than actual downstream training gains. The random-growth comparison supports the overall selection strategy, but an allocation-matched random baseline would help distinguish gains from capacity allocation across bins and layers from those due to particular unit choices.

  2. Compute accounting and the scope of baseline comparisons need clarification. For the stated DiT-S uniform mask, using the attention/MLP cost shares of 0.40/0.60 given in Section 6 changes normalized cost from 0.77778 to 0.78333, a relative increase of approximately 0.714%. The relationship between this calculation and the claimed maximum difference of 0.25% needs explanation. Scoring is explicitly charged to the FLOP budget, but Table 5 lacks absolute runtimes and does not specify whether wall-clock budgets include all growth-related work, including any recompilation. The pruning and depth-growth comparisons are also sensitive to the chosen schedules and capacity allocation. Conclusions about these alternatives should remain specific to the configurations tested.

Questions for authors
  1. What explains the default DiT-S FID of 63.8 in Table 2 versus 64.6 in Table 7? Please clarify any differences in settings or seeds, and provide per-seed results for GrowDiT and random growth in the main comparison.

  2. How much of the gain over random growth comes from allocating capacity across bins and layers, versus selecting particular units? A diagnostic comparison could randomize unit identities while preserving the numbers of newly activated heads and MLP groups in each bin and layer at each growth event.

  3. How is the maximum cost difference of 0.25% calculated, and which masks and normalization does it refer to? For Table 5, please report absolute training times and clarify whether scoring, mask updates and any recompilation are included, along with the treatment of initial compilation and warm-up.

  4. At the same total train-then-prune budget of 1.1×, how sensitive are the results to the split between dense pretraining and recovery? Would allocating more of this budget to recovery improve the comparison?

4 comments

We thank the Reviewer for their careful evaluation and valuable suggestions, which have helped us further strengthen and clarify our work.

Question 1 (also Weakness one). We thank the Reviewer for pointing out the inconsistency. The submitted version mixed single-seed values with means over seeds. We have re-checked every value, report all seeds in Appendix D (Table 9), and increased GrowDiT, random growth and uniform to three seeds at every scale. GrowDiT has a lower FID than random growth in all nine seed-matched pairs, and the relative gain increases with model size (table below).

GrowDiT over random growthFID gainrelative gain
DiT-S1.522.3%
DiT-B1.273.0%
DiT-L0.933.9%

Gain of GrowDiT over random growth on ImageNet-256 (1× budget), averaged over seeds. Relative gain: gain divided by the FID of random growth.

Question 2 (also Weakness one). We added the diagnostic baseline suggested by the Reviewer, need-score allocation with random units: it keeps the number of units that the need score gives to each layer at each noise level and chooses the units at random (Sec. 5.3, Table 5). Our ablations show that the allocation of capacity across noise levels and layers is substantially more important than the identities of the individual units being activated, and that gradient-based signals computed during training can effectively guide this noise-dependent allocation without access to a fully trained model. (1) Unit identities matter little: the FID of need-score allocation with random units is no higher than that of GrowDiT (64.1 against 64.3 on DiT-S). (2) The allocation found by the need score matters and has a long-term effect. First, it shows the same overall pattern as the DyDiT routers of train-then-prune, which are learned on a fully trained dense model (Fig. 4a). Second, at the end of training, its gain over random growth is about twice that of a hand-set allocation (active units decreasing linearly across noise levels, random units) on DiT-S and about three times on DiT-B, so its advantage grows with model size (table below).

over random growthFID gainrelative gain
DiT-S, hand-set allocation0.751.1%
DiT-S, need-score allocation1.442.2%
DiT-B, hand-set allocation0.531.2%
DiT-B, need-score allocation1.714.0%

Gain over random growth on ImageNet-256 (1× budget), averaged over seeds. All three methods choose the units at random. Relative gain: gain divided by the FID of random growth.

Question 3 (also Weakness two). We removed the 0.25% statement, which was computed incorrectly; we thank the Reviewer for pointing this out. Appendix F now derives the cost correction when the token interactions of attention are counted: +0.71% for the DiT-S uniform mask and at most 0.94% for any method and scale. Table 13 (Appendix G) reports the absolute training times used in Table 6; they include the initial compilation and warm-up and, for the growth methods, the need-score passes, the mask updates and the compilation of new widths.

Question 4 (also Weakness two). To answer this question, we tested 0.8× of dense training with 0.3× of recovery, and 0.6× with 0.5× (DiT-S and DiT-B), within the same 1.1× budget (table below; Appendix H, Table 14). Allocating more of the budget to recovery narrows the gap between train-then-prune and dense training: the FID improves from 92.4 to 71.8 on DiT-S (68.1 for dense) and from 74.4 to 48.5 on DiT-B (47.4 for dense). Nevertheless, train-then-prune remains inferior to GrowDiT for all three budget divisions (63.8 and 42.2 with the same seed).

dense trainingrecoverytotal budgetDiT-SDiT-B
train-then-prune1.0×0.1×1.1×92.474.4
train-then-prune0.8×0.3×1.1×72.153.2
train-then-prune0.6×0.5×1.1×71.848.5
dense1.1×–1.1×68.147.4
GrowDiT––1.0×63.842.2

FID-50k on ImageNet-256 (seed 0).

Scope of baseline comparisons (Weakness two). To clarify the scope of the baseline comparisons, we added a variant, depth growth with the need-score allocation (blocks activated per noise level; Sec. 5.3, Table 5). On DiT-S its FID is 64.1, against 69.2 for depth growth without noise dependence (table below; Table 5), so the need-score allocation also improves depth growth. With this baseline, the scope is clear: in every tested configuration (Sec. 5.2 and 5.3), noise-dependent growth with the need score, in width or depth, outperforms the baselines without growth (uniform, dense, train-then-prune) and growth without noise dependence (random growth, depth growth), and among noise-dependent growth methods the need score finds a better allocation than the hand-set one. Growing diffusion models remains less explored; we hope the added ablations, which show where the gain comes from, make the paper a useful reference for this direction.

DiT-SFIDgain
random growth65.6–
depth growth without noise dependence69.2-5.6%
depth growth with need-score allocation64.12.3%

FID-50k on DiT-S (ImageNet-256, 1× budget), averaged over seeds. Gain: FID reduction relative to random growth.

Figure 6 (Weakness one). We replaced Figure 6 with a measurement of how M, the number of samples used to estimate the need score, affects the downstream training gains (Appendix C, Table 8). Training with M=32 or M=512 instead of the default M=128 gives an FID within the spread of the default over seeds, so the FID after training is insensitive to M.

Thank you for the thorough revision. The per-seed results make the modest FID gains over random growth more convincing, with improvements in all nine paired ImageNet comparisons. The allocation-matched ablation addresses my main methodological question and suggests that the need score is most useful for allocating capacity across noise bins and layers. The corrected FLOP accounting and the additional pruning-budget splits also address my earlier concerns.

One point would still benefit from clarification. Section 4 describes the wall-clock comparison as single-seed, while Table 6 says the gains are averaged over seeds as in Table 2. Please clarify which seeds were used for Table 6 and whether the runtimes estimated from per-step timings in Appendix G were checked against elapsed times from complete training runs.

We thank the Reviewer for the careful reading of the revision and have updated the paper accordingly.

Seeds of Table 6. Table 6 uses the same three seeds as Table 2 for GrowDiT, random growth and uniform. The single-seed statement in Section 4 was left from the previous version, and we have corrected it.

Check against complete training runs. Following the Reviewer's suggestion, we trained GrowDiT, random growth and uniform again on DiT-S and DiT-B (seed 0), each as one uninterrupted run on one A100 with the model compiled, and recorded the time of the training loop. The measured time is 3.4% to 5.1% longer than our previous estimate (first table).

previous estimateestimatedmeasureddifference
DiT-S, GrowDiT5.705.94+4.3%
DiT-S, random growth5.695.98+5.1%
DiT-S, uniform5.395.61+4.1%
DiT-B, GrowDiT12.3112.72+3.4%
DiT-B, random growth12.3012.78+3.9%
DiT-B, uniform11.7112.27+4.7%

Training time at the 1x budget (hours on one A100, compiled; seed 0), estimated as in the previous revision. Step times are averaged over different A100 nodes (Appendix G). Difference: measured relative to estimated.

The previous estimate missed two things. (1) The other operations of a training step (data loading, the EMA update and the rest of the training loop), which take 5.9 ms per step on DiT-S and DiT-B and 26 ms on DiT-L (DiT-L trains with four micro-batches per step). (2) The step time, which the previous estimate interpolated coarsely and therefore inaccurately; we now use a finer interpolation, which is more accurate. With both corrections, the estimate is within 0.6% of the measured time in all six runs (second table; Appendix G, Table 15).

corrected estimateestimatedmeasureddifference
DiT-S, GrowDiT5.955.94-0.1%
DiT-S, random growth5.945.98+0.6%
DiT-S, uniform5.595.61+0.3%
DiT-B, GrowDiT12.7612.72-0.3%
DiT-B, random growth12.7612.78+0.2%
DiT-B, uniform12.2012.27+0.5%

Same runs, with the corrected estimate.

Updated results. We recomputed Table 6 and Table 14 with the corrected estimate and added the standard deviation over seeds to Table 6 (third table). The results differ little from those of the previous revision, and the calibration does not change any conclusion of the paper.

FID gain over uniform (%)DiT-SDiT-BDiT-L
budgets matched byFLOPsclockFLOPsclockFLOPsclock
GrowDiT2.1 ± 1.1-0.3 ± 1.06.0 ± 1.04.0 ± 1.19.6 ± 0.87.1 ± 0.7
random growth-0.3 ± 0.9-2.6 ± 0.93.2 ± 1.61.1 ± 1.46.0 ± 1.23.4 ± 1.2
GrowDiT, wall-clock gain / FLOP gain≈ 0%66%74%

ImageNet-256, mean ± standard deviation over three seeds; each seed is compared with uniform of the same seed (Table 6). FLOPs: every method trains for the 1x FLOP budget. Clock: every method trains for the wall-clock time that uniform needs for this budget.

We thank the Reviewer again for their constructive comment.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful evaluations, constructive feedback, and valuable suggestions, which have helped us further improve the quality and clarity of our manuscript.

Our work introduces GrowDiT, a training strategy that progressively grows diffusion transformers by allocating new capacity across noise levels, guided by a gradient-based need score. Following the reviewers' recommendations, we have carefully addressed all comments by clarifying the methodological contributions, strengthening the empirical evidence through additional targeted experiments and analyses, and improving the transparency and reproducibility of our results. These refinements provide further support for our main findings while preserving the core methodology, contributions and overall scope of the paper.

All changes are highlighted in blue in the revised manuscript, and detailed explanations are provided in our individual responses. We are grateful to the reviewers for their valuable contributions to improving our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.