PaperlayerView exact revision on arXiv

Source-linked AI summary

RegionE: Adaptive Region-Aware Generation for Efficient Image Editing

Pengtao Chen, Xianfang Zeng, Maosen Zhao, Mingzhu Shen, Peng Ye, Bangyin Xiang, Zhibo Wang, Wei Cheng, Gang Yu, Tao Chen

arXiv:2510.25590v1cs.CVcs.AI

TL;DR

Instruction-based image editing applies uniform computation despite differences between edited and unchanged regions, while its denoising-only redundancy remains underexplored. RegionE uses adaptive region-aware generation and caching, achieving reported speedups across three models while preserving quality.

  • Problem

    IIE applies uniform computation to edited and unchanged regions despite their different generation behavior, and redundancy in denoising-only IIE remains largely unexplored.

    Source: §1 · p. 2

  • Method

    RegionE partitions edited and unedited regions, uses one-step prediction for unchanged areas and local iterative generation for edited areas, and applies caches to reduce spatial and temporal redundancy.

    Source: §1 · pp. 2–3

  • Results

    2.57× on Step1X-Edit, 2.41× on FLUX.1 Kontext, and 2.06× on Qwen-Image-Edit, with GPT-4o evaluations confirming no quality degradation.

    Source: §1 · p. 3

  • Takeaways & Limitations

    RegionE reduces redundancy in IIE while maintaining minimal bias and negligible quality loss in the evaluated models.

    Source: §6 · p. 10

Abstract

from arXiv · show

Recently, instruction-based image editing (IIE) has received widespread attention. In practice, IIE often modifies only specific regions of an image, while the remaining areas largely remain unchanged. Although these two types of regions differ significantly in generation difficulty and computational redundancy, existing IIE models do not account for this distinction, instead applying a uniform generation process across the entire image. This motivates us to propose RegionE, an adaptive, region-aware generation framework that accelerates IIE tasks without additional training. Specifically, the RegionE framework consists of three main components: 1) Adaptive Region Partition. We observed that the trajectory of unedited regions is straight, allowing for multi-step denoised predictions to be inferred in a single step. Therefore, in the early denoising stages, we partition the image into edited and unedited regions based on the difference between the final estimated result and the reference image. 2) Region-Aware Generation. After distinguishing the regions, we replace multi-step denoising with one-step prediction for unedited areas. For edited regions, the trajectory is curved, requiring local iterative denoising. To improve the efficiency and quality of local iterative generation, we propose the Region-Instruction KV Cache, which reduces computational cost while incorporating global information. 3) Adaptive Velocity Decay Cache. Observing that adjacent timesteps in edited regions exhibit strong velocity similarity, we further propose an adaptive velocity decay cache to accelerate the local denoising process. We applied RegionE to state-of-the-art IIE base models, including Step1X-Edit, FLUX.1 Kontext, and Qwen-Image-Edit. RegionE achieved acceleration factors of 2.57, 2.41, and 2.06. Evaluations by GPT-4o confirmed that semantic and perceptual fidelity were well preserved.

1 INTRODUCTION

Instruction-based image editing faces high inference latency and applies uniform computation despite spatial and temporal redundancy. RegionE uses region-specific prediction and caching to accelerate editing, with reported speedups and preserved quality.

1 INTRODUCTION

  • Efficient inference research has largely targeted image generation, while redundancy and acceleration feasibility in denoising-only IIE remain largely unexplored.The paper situates its motivation amid diffusion-based editing models that follow textual instructions.Source: §1 · pp. 1–2
  • IIE models spend comparable computation on edited and unchanged regions, while attention keys and values and adjacent-timestep velocities exhibit temporal similarity.These spatial and temporal patterns motivate reducing computation in unchanged regions and reusing information across timesteps.Source: §1 · p. 2
  • Unedited regions follow nearly linear trajectories, so early-stage velocity reliably estimates later denoised images; edited regions follow curved trajectories and are harder to predict.Velocity remains consistent across consecutive timesteps, even though its direction and magnitude patterns differ across the denoising process.Source: §1 · pp. 2–3
  • RegionE is a training-free, region-aware framework that reduces spatial redundancy through adaptive predictions and local generation, and temporal redundancy through velocity-decay caching.It combines Adaptive Region Partition, Region-Instruction KV Cache, and Adaptive Velocity Decay Cache.Source: §1 · pp. 2–3
  • 2.57× on Step1X-Edit, 2.41× on FLUX.1 Kontext, and 2.06× on Qwen-Image-Edit, with PSNR and SSIM maintained and GPT-4o evaluations confirming no quality degradation.The reported PSNR values are 30.520, 32.133, and 31.115; SSIM values are 0.939, 0.917, and 0.937, respectively.Source: §1 · p. 3

2 RELATED WORK

Prior work spans diffusion acceleration and image-editing paradigms. RegionE is positioned as addressing spatial and temporal redundancy in instruction-based editing, using trajectory properties specific to that task.

2 RELATED WORK

  • Diffusion acceleration methods target parameter, temporal, or spatial redundancy through techniques including pruning, quantization, distillation, feature reuse, and selective token updates.The passage contrasts these approaches with RegionE’s focus on trajectory characteristics unique to IIE and both spatial and temporal redundancy.Source: §2 · p. 3
  • Image editing progressed from ControlNet and inversion-based pipelines to general-purpose IIE models that use textual instructions without masks or task-specific designs.Earlier modular approaches often relied on language models, spatial cues, or synthetic instruction–image pairs and were described as task-specific.Source: §2 · p. 3

3 PRELIMINARY

The preliminaries describe rectified-flow inference and the Instruction-DiT inputs used in IIE. They frame the model’s denoising iterations as the process RegionE seeks to optimize.

3 PRELIMINARY

  • Figure 2 compares traditional DiT with IIE DiT and depicts denoising, adjacent-timestep velocity similarities, and cross-step key similarities.The paper identifies the denoising iterations as the inference process RegionE targets for optimization.Source: Figure 2 · p. 4, §3 · p. 4
  • Rectified flow linearly connects a noise sample X1 and target image X0 across timestep t, and inference reverses this path by iteratively solving for X0.The paper describes discrete Euler sampling as an approximation to this iterative inference process.Source: §3 · p. 4
  • Instruction-DiT estimates the velocity field from text, noise, and instruction tokens; its output corresponds only to the noise-token component.Text and instruction tokens serve as static control signals, while noise tokens evolve across timesteps.Source: §3 · p. 4

4 METHODOLOGY

RegionE divides denoising into stabilization, region-aware generation, and smoothing. It uses the distinct trajectories of edited and unedited regions, plus velocity similarity across timesteps, to reduce computation.

4 METHODOLOGY

  • The method has three stages: Stabilization, Region-Aware Generation, and Smooth, with the middle stage addressing spatial and temporal redundancy.Region-Aware Generation combines adaptive region partition, region-aware generation, and adaptive velocity decay caching.Source: §4 · p. 5

Stabilization and Smooth Stages

  • The method leaves early unstable denoising unchanged, periodically recomputes the full image to update RIKVCache, then smooths stitched regions with unaccelerated denoising.The authors report that two smoothing steps are sufficient to eliminate boundary gaps effectively.Source: §4 · pp. 5–6

Adaptive Region Partition

  • After stabilization, unedited regions follow straighter trajectories and can be estimated at earlier timesteps in one step, while edited regions are less accurately predicted this way.The final unedited-region estimate is nearly identical to the true region; the same estimate is inaccurate for edited regions.Source: §4 · pp. 5–6

Region-Aware Generation

  • Unedited regions are denoised with one-step predictions, while edited regions are generated locally; RIKVCache retains global context for attention.Removing instruction and unedited-region inputs can bias edited-region velocity estimates, motivating the cache.Source: §4 · p. 6

Adaptive Velocity Decay Cache

  • AVDCache exploits similar adjacent-step velocity directions and gradually decaying magnitudes, using a solver-aware and timestep-aware decay factor.Its criterion measures cumulative error across the cache interval, and AVDCache is described as an improved residual-cache method.Source: §4 · p. 6

5 EXPERIMENT

RegionE is evaluated on three IIE models for editing quality and efficiency, with ablations examining its cache and stage components. It reports cross-model speedups while preserving quality, and the ablations identify quality and efficiency trade-offs.

5.1 EXPERIMENTAL SETTINGS

  • The evaluation used 606 image-prompt pairs for Step1X-Edit and Qwen-Image-Edit, and 1,026 for FLUX.1 Kontext, with 28 sampling steps per model.Latency was measured on one NVIDIA H800 GPU, one image at a time.Source: §5.1 · p. 7

5.2 EXPERIMENTAL RESULTS ANALYSIS

  • Against the pretrained models, RegionE achieved PSNR values of 30.520 dB, 32.133 dB, and 31.115 dB across the three models, with high SSIM and lower LPIPS scores.The authors report consistent quality preservation across architectures.Source: §5.2 · pp. 7–8
  • 2.572×, 2.409×, and 2.059× speedups were achieved on Step1X-Edit, FLUX.1 Kontext, and Qwen-Image-Edit, respectively.Latency fell from 27.945s to 10.865s, 14.682s to 6.096s, and 32.125s to 15.604s, respectively.Source: Table 1 · p. 8, §5.2 · p. 8
  • GPT-4o assigned Qwen-Image-Edit a G-SC score of 8.242 alongside a 2.059× speedup, while RegionE's G-PQ scores exceeded competing acceleration methods by 0.1 to 0.3 points.The evaluation covers semantic consistency, perceptual quality, and overall quality.Source: §5.2 · p. 8
  • Figure 4 shows Step1X-Edit outputs where RegionE is reported to preserve details and contours better than baselines at higher speedups, with ARP masks closely matching human perception.The last column displays ARP predictions of unedited regions.Source: §5.2 · p. 8, Figure 4 · p. 9

5.3 ABLATION STUDY

  • Removing RIKVCache raised speedup to 2.734× but reduced PSNR from 30.520 to 22.868 and G-O from 6.948 to 5.191; removing AVDCache limited acceleration to 1.733×.The authors attribute the RIKVCache quality loss to removing instruction information and unedited-region context from local attention.Source: §5.3 · p. 9, Table 2 · p. 9
  • Removing STS reduced PSNR from 30.520 to 21.441 and G-O from 6.948 to 6.325; removing SMS or the forced RAGS step also lowered reported quality metrics.The authors state that STS addresses instability in speed estimation, SMS bridges edited and unedited regions, and the forced step mitigates KV-similarity decay.Source: §5.3 · p. 9, Table 2 · p. 9

6 CONCLUSION

RegionE combines early spatial-region prediction with region-wise editing and a cache for temporal redundancy. It reports 2.57×, 2.41×, and 2.06× end-to-end speedups across three IIE models while maintaining minimal bias and comparable GPT-4o quality scores.

6 CONCLUSION

  • RegionE combines ARP and RIKVCache for region-wise editing with AVDCache to reduce spatial and temporal redundancy.The conclusion reports end-to-end speedups of 2.57×, 2.41×, and 2.06× on Step1X-Edit, FLUX.1 Kontext, and Qwen-Image-Edit, respectively, with PSNR of 30.52–32.13 and comparable GPT-4o results.Source: §6 · p. 10

Supplementary Material

The supplementary material is organized into four sections covering RegionE pseudocode, cache analysis, and per-task visual and quantitative benchmark results.

A PSEUDOCODE OF REGIONE

The pseudocode takes a diffusion transformer, image and text inputs, noise, stage lengths, thresholds, and forced steps, then outputs the edited target image. It sequences stabilization, region-aware generation, and smoothing, with cache operations and local denoising within the region-aware stage.

  • The algorithm accepts a diffusion transformer, sampling step, instruction image, text tokens, and random noise, along with stage lengths, thresholds, and a forced-step list.Source: A PSEUDOCODE OF REGIONE · p. 14
  • During stabilization, it iteratively updates the noisy image and initializes the Region-Instruction KV Cache.The pseudocode then estimates an early result and partitions edited and unedited regions using a threshold.Source: A PSEUDOCODE OF REGIONE · p. 14
  • In region-aware generation, unedited regions receive one-step estimates while edited regions undergo iterative denoising with cache operations.The algorithm also checks the adaptive velocity decay cache during edited-region denoising.Source: A PSEUDOCODE OF REGIONE · p. 14
  • The smooth stage resumes iterative updates before the algorithm outputs the target image after editing.Source: A PSEUDOCODE OF REGIONE · p. 14

B ANALYSIS OF ADAPTIVE VELOCITY DECAY CACHE

The analysis relates residual caching to velocity caching, characterizing the residual cache as a decayed form of the velocity cache. It introduces a timestep correction coefficient to account for timestep-dependent velocity decay.

B ANALYSIS OF ADAPTIVE VELOCITY DECAY CACHE

  • For N residual-caching steps, the analysis derives an expression for the cached result, with a coefficient before v_ti that is less than one.The derivation is illustrated as a residual-cache pipeline in Figure 5.Source: B ANALYSIS OF ADAPTIVE VELOCITY DECAY… · p. 15, Figure 5 · p. 15
  • Residual caching is equivalent to velocity caching, with the residual cache acting as a decayed form of the velocity cache.The solver determines the decay coefficient; an external timestep correction coefficient γ_ti accounts for timestep-dependent decay, and setting γ_ti to 1 recovers Equation 13.Source: B ANALYSIS OF ADAPTIVE VELOCITY DECAY… · p. 15

C PER-TASK VISUALIZATION RESULTS IN THE BENCHMARK

The visual appendix presents RegionE and baseline editing examples across Step1X-Edit-v1p1, FLUX.1 Kontext, and Qwen-Image-Edit. It covers 11 tasks for Step1X-Edit and Qwen-Image-Edit, and five for FLUX.1 Kontext.

C PER-TASK VISUALIZATION RESULTS IN THE BENCHMARK

D PER-TASK QUANTITATIVE RESULTS IN THE BENCHMARK

The benchmark tables compare RegionE with baseline methods in terms of quality and efficiency across 11 GEdit-Bench tasks and five KontextBench tasks. The supplied passages identify the tables and their scope but do not provide cell values or comparative outcomes.

D PER-TASK QUANTITATIVE RESULTS IN THE BENCHMARK

Related papers

Loading 2510.25590v1…