acceptodds
Under review as a conference paper at ICLR 2027

LightDraft: Accelerating Masked Diffusion Image Editing with Look-Ahead Guidance

Abstract

Masked diffusion models support instruction-guided image editing through parallel token prediction. However, their reliance on bidirectional attention requires repeated full-sequence computation throughout sampling, while a fixed sampling schedule assigns the same refinement budget to edits with very different spatial extent and convergence behavior. To address this problem, we investigate whether intermediate predictions can reveal where computation is needed and how long refinement should continue. We observe that the spatial support of an edit often stabilizes before its visual appearance is fully refined. Based on this observation, we propose LightDraft, a training-free acceleration method that adapts both feature computation and sampling length to each edit. LightDraft decodes intermediate complete token predictions into look-ahead images, exposing the developing edit before sampling is complete. Differences from the reference image are used to progressively localize the edit region and guide a two-stage regional caching strategy, allowing most features outside the region to be reused. Meanwhile, changes between successive look-ahead images guide the sampling budget, reducing unnecessary refinement when predictions have largely stabilized. Experiments on ImgEdit and PIE-Bench show that LightDraft achieves an 8.14× speedup while maintaining instruction-following capability and editing quality.

Primary area:
generative models
Keywords:
Masked Diffusion, Image Editing, training free accelerate

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

LightDraft accelerates the pretrained Lumina-DiMOO image editor without further training. It decodes complete intermediate token predictions into look-ahead images, compares them with a reconstructed source to guide two-stage regional feature caching, and sets a bounded sampling budget from changes between these predictions. The implementation also applies region-guided residual correction to the decoded output. Experiments on 637 ImgEdit instructions and 700 PIE-Bench tasks report roughly eightfold speedups over 64-step inference with improved aggregate editing and background-preservation scores.

Critical strengths
  1. The method makes useful use of intermediate predictions that are already available before remasking. Comparing decoded predictions with the reconstructed reference, aggregating early candidate regions, and performing a full refresh before refining the support are sensible ways to manage uncertainty in early localization. The paper also explicitly recognizes that cached features can become stale under bidirectional attention and includes decoding and comparison costs in end-to-end latency.

  2. The experimental results look great. On ImgEdit, Table 1 reduces latency from 49.19 to 6.04 seconds while increasing overall EditScore from 4.520 to 4.786. On PIE-Bench, Table 2 reports 6.02 seconds, background PSNR of 31.28 versus 26.08, and CLIP-Edit of 21.49 versus 21.37.

Critical weaknesses
  1. The preservation improvement cannot be attributed to the proposed caching mechanism from the reported ablation. Section 3.3 adds the reference reconstruction residual outside the expanded edit support directly to the final RGB output. Table 3 groups two-stage caching, minimum-support expansion, and this residual correction together as R. R alone already obtains PSNR 31.70 and LPIPS 0.029, compared with 31.28 and 0.031 for the complete method. A correction-only control, caching without correction, and the same correction applied to a comparable baseline are needed to identify the source of the headline preservation result.

  2. The main uniform-reduction control uses 24 steps, whereas the adaptive rule is constrained to 12 to 18. The adaptive-only row improves latency from 18.36 to 10.74 seconds with almost unchanged CLIP-Edit, but this does not show that allocating steps according to prediction changes is better than simply choosing a fixed shorter schedule. Figure 2 explicitly contains schematic rather than measured distributions, and no distribution of allocated budgets or comparison against fixed 12-, 15-, and 18-step schedules is reported.

Questions for authors

How much of the background gain survives when residual correction is disabled, and how much can the same correction recover for uniform-step inference or an existing cache baseline using the same estimated support? This is the most consequential missing control.

At matched mean latency, how does adaptive allocation compare with a fixed budget or with the same budget distribution randomly reassigned across requests? Please report the allocation distribution and whether larger budgets correlate with actual quality improvements, rather than only with prediction change.

6 comments

Q3.1. The preservation improvement cannot be attributed to the proposed caching mechanism from the reported ablation. Section 3.3 adds the reference reconstruction residual outside the expanded edit support directly to the final RGB output. Table 3 groups two-stage caching, minimum-support expansion, and this residual correction together as R. R alone already obtains PSNR 31.70 and LPIPS 0.029, compared with 31.28 and 0.031 for the complete method. A correction-only control, caching without correction, and the same correction applied to a comparable baseline are needed to identify the source of the headline preservation result.

Thank you for this suggestion. We agree that grouping regional caching, support expansion, and RRC in Table 3 makes their individual contributions unclear. We therefore added separate component ablations and a 14-step Lumina-DiMOO baseline with RRC in Rebuttal Table SM-1 below.

RRC substantially improves background preservation. On PIE-Bench, adding RRC to LightDraft reduces background MSE from 0.005170 to 0.002791, a 46.0% reduction. RRC uses the expanded edit support estimated from look-ahead images to apply the reference reconstruction residual to the estimated background. This region-guided correction is itself part of LightDraft’s contribution. However, even without RRC, LightDraft achieves background preservation comparable to the original model using 24 or 18 steps (0.005170 vs. 0.005134 and 0.005116), and better than the 14-step baseline (0.005437). Support expansion provides a smaller additional benefit: with RRC disabled, adding support expansion reduces MSE from 0.005283 to 0.005170.

On ImgEdit, LightDraft without RRC also achieves an EditScore comparable to original 24- and 18-step inference. Using an average of 14.15 steps, it is 1.72× faster than the original model using 14 steps, while improving EditScore from 4.396 to 4.552.

Rebuttal Table SM-1: Component ablations and correction-only control. Mean steps, editing scores, and latency are measured on ImgEdit. The Speedup column is relative to original 64-step Lumina-DiMOO inference on ImgEdit (49.19 s/image). Background MSE is measured on PIE-Bench.

MethodMean stepsEditScore ↑Following ↑Consistency ↑Perceptual ↑Latency (s/image) ↓Speedup ↑Background MSE ↓
Original Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×0.00543709
Lumina-DiMOO, 14 steps + RRC144.7085.7927.2825.82410.524.68×—
Original Lumina-DiMOO, 18 steps184.4935.7966.9215.08813.523.64×0.00511588
Original Lumina-DiMOO, 24 steps244.5295.7906.9205.09018.062.72×0.00513444
LightDraft w/o RRC and support expansion14.154.5325.8006.7745.1135.948.28×0.00528280
LightDraft w/o support expansion14.154.8245.7847.4025.9285.948.28×0.00290128
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×0.00517001
LightDraft14.154.9575.9417.3515.8696.138.02×0.00279071

Q3.2. The main uniform-reduction control uses 24 steps, whereas the adaptive rule is constrained to 12 to 18. The adaptive-only row improves latency from 18.36 to 10.74 seconds with almost unchanged CLIP-Edit, but this does not show that allocating steps according to prediction changes is better than simply choosing a fixed shorter schedule. Figure 2 explicitly contains schematic rather than measured distributions, and no distribution of allocated budgets or comparison against fixed 12-, 15-, and 18-step schedules is reported.

Thank you for this suggestion. We agree that the 24-step control alone does not adequately address comparisons with shorter fixed schedules. We therefore add original Lumina-DiMOO inference with 12, 15, and 18 sampling steps in Rebuttal Table SM-2, together with 14-step inference for comparison with LightDraft’s measured average of 14.15 steps per image.

Even without RRC, LightDraft achieves an EditScore comparable to original 18- and 24-step inference. Compared with original 15-step inference, it improves EditScore from 4.402 to 4.552 while reducing latency from 11.27 to 6.13 s/image (1.84× speedup). Against the closer 14-step control, it also improves EditScore from 4.396 to 4.552 with 1.72× acceleration. Feature reuse reduces computation per step, allowing LightDraft to run faster even than original 12-step inference (6.13 vs. 9.02 s/image).

Rebuttal Table SM-2: Additional comparisons with fixed sampling schedules on ImgEdit. The Speedup column is relative to original 64-step Lumina-DiMOO inference (49.19 s/image).

MethodMean stepsEditScore↑Following↑Consistency↑Perceptual↑Latency (s/image)↓Speedup↑
Lumina-DiMOO, 64 steps (paper baseline)644.5205.6416.9685.34949.191.00×
Lumina-DiMOO, 24 steps244.5295.7906.9205.09018.062.72×
Lumina-DiMOO, 18 steps184.4935.7966.9215.08813.523.64×
Lumina-DiMOO, 15 steps154.4025.6906.7654.99311.274.36×
Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×
Lumina-DiMOO, 12 steps124.3425.6076.7614.9569.025.45×
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×
LightDraft14.154.9575.9417.3515.8696.138.02×

We also report the empirical distribution of allocated sampling budgets across all 637 ImgEdit samples in Rebuttal Table SM-3. The measured mean is 14.15 steps, with individual budgets ranging from 12 to 18 steps.

Rebuttal Table SM-3: Measured sampling-budget distribution on ImgEdit.

Allocated stepsNumber of samplesPercentage
1211718.37%
1315724.65%
1412219.15%
1510917.11%
16609.42%
17335.18%
18396.12%
Total637100.00%

Regarding Figure 2, the illustrated distributional trends are consistent with our measurements. We will replace the schematic plots with plots based on measured data.

Q3.3. How much of the background gain survives when residual correction is disabled, and how much can the same correction recover for uniform-step inference or an existing cache baseline using the same estimated support? This is the most consequential missing control.

Thank you for this suggestion. Rebuttal Table SM-4 reports LightDraft with and without RRC, together with original Lumina-DiMOO inference and a 14-step baseline with RRC.

On PIE-Bench, without RRC, LightDraft preserves the background comparably to original 18- and 24-step inference, with background MSE of 0.005170 versus 0.005116 and 0.005134, respectively. It also achieves lower MSE than original 14-step inference (0.005437). Enabling RRC further reduces LightDraft’s background MSE by 46.0%, from 0.005170 to 0.002791, demonstrating its substantial contribution to background preservation.

For uniform-step inference, applying RRC to the original 14-step model improves ImgEdit EditScore from 4.396 to 4.708 and consistency from 6.798 to 7.282. With RRC enabled in both methods, LightDraft achieves an EditScore of 4.957 while remaining 1.72× faster (6.13 vs. 10.52 s/image). These results demonstrate that RRC also benefits uniform-step inference on ImgEdit; the corresponding PIE-Bench background recovery remains unquantified.

Rebuttal Table SM-4: RRC ablations and uniform-step comparisons. Background MSE is measured on PIE-Bench; all other metrics are measured on ImgEdit. The Speedup column is relative to original 64-step Lumina-DiMOO inference on ImgEdit (49.19 s/image).

MethodMean stepsEditScore↑Following↑Consistency↑Perceptual↑Latency (s/image)↓Speedup↑Background MSE↓
Original Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×0.00543709
Lumina-DiMOO, 14 steps + RRC144.7085.7927.2825.82410.524.68×—
Original Lumina-DiMOO, 18 steps184.4935.7966.9215.08813.523.64×0.00511588
Original Lumina-DiMOO, 24 steps244.5295.7906.9205.09018.062.72×0.00513444
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×0.00517001
LightDraft14.154.9575.9417.3515.8696.138.02×0.00279071
Reviewer_zV9wOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LightDraft, a training-free acceleration method for Masked Diffusion-based image editing. Its core idea is to utilize look-ahead images generated from intermediate predictions to simultaneously guide spatial computation allocation and temporal step allocation. Specifically, LightDraft locates editing regions by comparing differences with a reference image and employs a two-stage region caching approach to reuse features from non-edited regions. It then adaptively adjusts the sampling budget based on the degree of change between adjacent predicted images. In experiments on the ImgEdit and PIE-Bench datasets using the Lumina-DiMOO model, the method achieves 8.14× and 7.99× acceleration, respectively, while delivering good results in terms of editing quality and background preservation metrics.

Critical strengths
  1. The LightDraft method is practical and requires no additional training. It combines edit region localization, feature caching, and adaptive sampling, while optimizing both the computational cost per step and the total number of sampling steps. The method has good engineering applicability.

  2. The experimental evaluation in this paper is thorough. The authors not only compare the method with various acceleration techniques on two image editing benchmarks but also provide component-wise ablation experiments to validate the method’s acceleration performance and background preservation capabilities.

Critical weaknesses
  1. The paper does not sufficiently substantiate the effectiveness of its core mechanisms. The localization of editing regions relies on fixed thresholds and preset update times, and estimation errors in editing regions may affect the subsequent cache accuracy. The region-processing component proposed in the paper simultaneously incorporates region caching, minimum support region expansion, and residual correction. The existing ablation studies have not adequately distinguished the independent contributions of each component to background preservation gains.

  2. The experimental evaluation is relatively thorough, but the scope of the domain is rather limited. The method was primarily validated on a single model, Lumina-DiMOO, and its applicability to scenarios such as global style changes and large-scale edits remains unclear.

Questions for authors

As shown in Table 2, the background PSNR improved from 26.08 dB to 31.28 dB, and the SSIM improved from 0.804 to 0.934. Since the method includes a mechanism for output correction by reconstructing residuals using a reference image, could the authors provide results for this correction when enabled in isolation, and explain whether it would compromise editing tasks that require changing the global background?

6 comments

Q2.1. The paper does not sufficiently substantiate the effectiveness of its core mechanisms. The localization of editing regions relies on fixed thresholds and preset update times, and estimation errors in editing regions may affect the subsequent cache accuracy. The region-processing component proposed in the paper simultaneously incorporates region caching, minimum support region expansion, and residual correction. The existing ablation studies have not adequately distinguished the independent contributions of each component to background preservation gains.

Thank you for this suggestion. In the paper, we distinguish two stages of masked diffusion image editing: Region Localization and Detail Refinement. As illustrated in Figure 1 in the paper, the regions where look-ahead images differ from the reconstructed reference tend to stabilize toward the end of localization, while subsequent generation mainly refines their appearance. This observation motivates estimating the edit region at a preset transition step, then concentrating computation within it while reusing most features elsewhere.

To measure how well our estimated edit region covers the edits in the final output, we analyzed original 24-step Lumina-DiMOO inference on all 700 PIE-Bench tasks. Following the procedure described in the paper, we computed the estimated edit region from intermediate look-ahead images and identified the final edited region from the completed output. These regions were measured for comparison without changing the original generation process.

Rebuttal Table ZV-1: Coverage of final changes by the estimated edit region. Final changes are detected by comparing the final output with the reconstructed reference using the same difference measure and threshold as localization. Each ratio is computed per image and then averaged.

Coverage metricMean
Area of final changes outside the estimated edit region / Total area of final changes3.58%
Area of final changes outside the estimated edit region / Total image area0.78%

These results show that the estimated edit region covers the final edited region well. Manual inspection further showed that nearly all uncovered areas fell into two categories, rather than missed edits requiring further refinement:

  1. Barely perceptible texture variations unrelated to the editing instruction.
  2. Regions where actual edits were already complete and local predictions had stabilized early in Region Localization.

We will include representative examples in the final paper.

Rebuttal Table ZV-2 reports the component ablations. On PIE-Bench, RRC reduces LightDraft’s background MSE by 46.0%, from 0.005170 to 0.002791. RRC uses the expanded edit support estimated from look-ahead images to apply the reference reconstruction residual to the estimated background. This region-guided correction is itself part of LightDraft’s contribution.

Even without RRC, LightDraft maintains background preservation comparable to original 24- and 18-step inference, and better than original 14-step inference. With RRC disabled, support expansion further reduces background MSE from 0.005283 to 0.005170.

On ImgEdit, LightDraft without RRC also achieves an EditScore comparable to original 24- and 18-step inference. Using 14.15 steps on average, it is 1.72× faster than original 14-step inference (6.13 vs. 10.52 s/image), while improving EditScore from 4.396 to 4.552.

Rebuttal Table ZV-2: Component ablations. EditScore and latency are measured on ImgEdit; background MSE is measured on PIE-Bench.

MethodEditScore ↑Latency (s/image) ↓Background MSE ↓
Original Lumina-DiMOO, 24 steps4.52918.060.00513444
Original Lumina-DiMOO, 18 steps4.49313.520.00511588
Original Lumina-DiMOO, 14 steps4.39610.520.00543709
LightDraft w/o RRC and support expansion4.5325.940.00528280
LightDraft w/o support expansion4.8245.940.00290128
LightDraft w/o RRC4.5526.130.00517001
LightDraft4.9576.130.00279071

Q2.2. The experimental evaluation is relatively thorough, but the scope of the domain is rather limited. The method was primarily validated on a single model, Lumina-DiMOO, and its applicability to scenarios such as global style changes and large-scale edits remains unclear.

Thank you for this suggestion. To test transfer beyond Lumina-DiMOO, we applied LightDraft to EditMGT [1], with RRC disabled in all transfer experiments.

As shown in Rebuttal Table ZV-3, LightDraft achieves comparable EditScore with substantially fewer sampling steps. Compared with original 64-step inference, it uses only 43.92 steps on average to achieve an EditScore of 5.331 versus 5.315, reducing the step count by 31.4%. This also reduces latency from 8.04 to 4.55 s/image, yielding 1.77× acceleration.

Rebuttal Table ZV-3: Transfer to EditMGT. RRC is disabled in all LightDraft configurations. Mean steps report the actual average sampling budget. Speedups are calculated using unrounded latency measurements.

MethodMean stepsEditScore ↑Following ↑Consistency ↑Perceptual ↑Latency (s/image) ↓
Original EditMGT18.004.4536.5596.1264.2222.56
EditMGT + LightDraft17.954.6866.7536.3234.4902.13
Original EditMGT32.004.8456.8236.5454.7694.23
EditMGT + LightDraft32.355.0577.1026.6894.8603.47
Original EditMGT40.005.1227.0666.7464.9445.18
EditMGT + LightDraft40.545.2267.2046.8015.0744.36
Original EditMGT44.005.1247.1056.7144.9285.66
EditMGT + LightDraft43.925.3317.3176.9805.1544.55
Original EditMGT64.005.3157.2056.8335.2838.04

We acknowledge that global style changes and large-scale edits remain challenging for LightDraft: modifying most of the image leaves fewer regions available for feature reuse, limiting the acceleration benefit. Although such edits reduce the gains from Two-Stage Regional Caching, Adaptive Sampling Budget can still allocate sampling steps according to the magnitude of prediction changes during generation. LightDraft can therefore achieve efficiency gains even on global editing tasks. Please see our response to Q2.3(b) for a detailed analysis and supporting experiments.

[1] Wei Chow et al. EditMGT: Unleashing Potentials of Masked Generative Transformers in Image Editing. CVPR, 2026.

Q2.3(a). As shown in Table 2, the background PSNR improved from 26.08 dB to 31.28 dB, and the SSIM improved from 0.804 to 0.934. Since the method includes a mechanism for output correction by reconstructing residuals using a reference image, could the authors provide results for this correction when enabled in isolation?

Thank you for this suggestion. We applied RRC to original 14-step Lumina-DiMOO inference and report this correction-only control alongside LightDraft and the original 64-step baseline in Rebuttal Table ZV-4.

With RRC enabled in both methods, LightDraft achieves higher scores on all four ImgEdit metrics than original 14-step inference with RRC. At similar sampling budgets (14.15 vs. 14 steps), EditScore increases from 4.708 to 4.957, while latency decreases from 10.52 to 6.13 s/image, providing 1.72× acceleration. This speedup comes from lower computation per step: original inference recomputes features across the full sequence, whereas LightDraft reuses most features outside the estimated edit region. Thus, applying RRC alone does not recover LightDraft’s overall quality–efficiency advantage.

Even without RRC, LightDraft retains this feature reuse, improving EditScore over original 14-step inference (4.552 vs. 4.396) while remaining 1.72× faster. It also achieves an EditScore comparable to the original 64-step baseline (4.552 vs. 4.520) with 8.02× acceleration. RRC further improves background preservation on PIE-Bench, reducing LightDraft’s background MSE from 0.005170 to 0.002791, a 46.0% reduction.

Rebuttal Table ZV-4: Correction-only control and LightDraft comparisons. Mean steps, editing scores, and latency are measured on ImgEdit; background MSE is measured on PIE-Bench. The Speedup column is relative to original 64-step Lumina-DiMOO inference (49.19 s/image).

MethodMean stepsEditScore ↑Following ↑Consistency ↑Perceptual ↑Latency (s/image) ↓Speedup ↑Background MSE ↓
Original Lumina-DiMOO, 64 steps644.5205.6416.9685.34949.191.00×—
Original Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×0.00543709
Lumina-DiMOO, 14 steps + RRC144.7085.7927.2825.82410.524.68×—
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×0.00517001
LightDraft14.154.9575.9417.3515.8696.138.02×0.00279071
Reviewer_HeAxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed, LightDraft, a training-free method for accelerating masked diffusion image editing. It uses intermediate look-ahead images to identify edit regions, cache background features, and adapt the number of sampling steps. It also uses a residual correction to improve background preservation. Experiments shows almost 8x speedup with maintained or improved the quality.

Critical strengths
  • The proposed method is well motivated. Reducing the computational cost on unchanged reason is make sence.
  • The proposed method requires no additioanl masks.
  • Reported resuts shows the strong efficiency with comparable performance in two different benchmarks.
Critical weaknesses
  • One of my concen is the unmatched denoise steps setting in the comparison. The baselines uses the fixed 24steps during the experiments, while the LightDraft use a reduced sampling budget (12-18). The efficiency comparison on this unmached denoise steps is kind of unfair.
  • While Section 3.3 explains the role of RRC, Table 3 groups regional caching, support expansion, and RRC into a single component. This makes it difficult to determine how much of the improvement in preservation comes from caching versus the final residual correction. Separate ablations of these components, along with a baseline using the same residual correction, would help clarify their individual contributions.
Questions for authors

There is no further question of this paper.

2 comments

Q1.1. One of my concen is the unmatched denoise steps setting in the comparison. The baselines uses the fixed 24steps during the experiments, while the LightDraft use a reduced sampling budget (12-18). The efficiency comparison on this unmached denoise steps is kind of unfair.

Thank you for this suggestion. In the paper (Section 4.1), we use original 64-step Lumina-DiMOO inference as the baseline for all reported speedups. The paper also includes 24-step inference as a direct step-reduction control. We now add 14- and 18-step comparisons in Rebuttal Table HA-1.

LightDraft without RRC achieves an EditScore comparable to the original Lumina-DiMOO model using 24 or 18 sampling steps, while being 2.20× faster than 18-step inference. LightDraft uses 14.15 steps per image on average. Compared with the original model using 14 steps, feature reuse reduces latency from 10.52 to 6.13 s/image (1.72× speedup), while EditScore increases from 4.396 to 4.552.

Rebuttal Table HA-1: Additional comparisons on ImgEdit. The Speedup column is relative to original 64-step Lumina-DiMOO inference.

MethodMean stepsEditScore↑Following↑Consistency↑Perceptual↑Latency (s)↓Speedup↑
Lumina-DiMOO, 64 steps (paper baseline)644.5205.6416.9685.34949.191.00×
Lumina-DiMOO, 24 steps244.5295.7906.9205.09018.062.72×
Lumina-DiMOO, 18 steps184.4935.7966.9215.08813.523.64×
Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×
LightDraft14.154.9575.9417.3515.8696.138.02×

Q1.2. While Section 3.3 explains the role of RRC, Table 3 groups regional caching, support expansion, and RRC into a single component. This makes it difficult to determine how much of the improvement in preservation comes from caching versus the final residual correction. Separate ablations of these components, along with a baseline using the same residual correction, would help clarify their individual contributions.

Thank you for this suggestion. We agree that Table 3 in the paper groups regional caching, support expansion, and RRC into a single component. We therefore conducted additional component ablations and applied RRC to the original Lumina-DiMOO model with 14 sampling steps (Rebuttal Table HA-2).

On ImgEdit, even without RRC, LightDraft achieves an EditScore comparable to the original Lumina-DiMOO model using 24 or 18 sampling steps. LightDraft uses 14.15 steps per image on average. Compared with the original model using 14 steps, feature reuse reduces latency from 10.52 to 6.13 s/image (1.72× speedup), while EditScore increases from 4.396 to 4.552.

On PIE-Bench, RRC substantially improves background preservation, reducing LightDraft’s background MSE by 46.0%, from 0.005170 to 0.002791. However, even without RRC, LightDraft preserves the background comparably to the original model using 24 or 18 steps (MSE of 0.005170 vs. 0.005134 and 0.005116, respectively), while outperforming the 14-step baseline (0.005170 vs. 0.005437). Support expansion also contributes to preservation, as removing it from LightDraft w/o RRC raises MSE from 0.005170 to 0.005283.

Rebuttal Table HA-2: Component ablations on ImgEdit and PIE-Bench. Mean steps, editing scores, and latency are reported on ImgEdit. The Speedup column is relative to original 64-step Lumina-DiMOO inference on ImgEdit (49.19 s/image). Background MSE is evaluated on PIE-Bench.

MethodMean stepsEditScore ↑Following ↑Consistency ↑Perceptual ↑Latency (s/image) ↓Speedup ↑Background MSE ↓
Original Lumina-DiMOO, 24 steps244.5295.7906.9205.09018.062.72×0.00513444
Original Lumina-DiMOO, 18 steps184.4935.7966.9215.08813.523.64×0.00511588
Original Lumina-DiMOO, 14 steps144.3965.6996.7984.99210.524.68×0.00543709
Lumina-DiMOO, 14 steps + RRC144.7085.7927.2825.82410.524.68×—
LightDraft w/o RRC and support expansion14.154.5325.8006.7745.1135.948.28×0.00528280
LightDraft w/o support expansion14.154.8245.7847.4025.9285.948.28×0.00290128
LightDraft w/o RRC14.154.5525.7806.7955.1326.138.02×0.00517001
LightDraft14.154.9575.9417.3515.8696.138.02×0.00279071

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.