PaperlayerView exact revision on arXiv

Source-linked AI summary

EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing

Jiayi Song, Shijie Huang, Fangtai Wu, Yubo Huang, Zhenxiong Tan, Songhua Liu, Jiaming Liu, Ruihua Huang

arXiv:2608.18063v1cs.CV

TL;DR

High-resolution image editing remains limited by computational cost and fidelity challenges at larger resolutions. EditBridge uses a source-guided diffusion bridge with sparse attention and achieves computationally efficient editing with faithful detail preservation.

  • Problem

    Existing image editing models are generally limited to 1K resolution, while native high-resolution training and inference remain prohibitively costly.

  • Method

    EditBridge refines low-resolution edits into high-resolution outputs through source-conditioned data-to-data translation and prior-guided sparse attention.

  • Results

    EditBridge achieves high-resolution editing with computational efficiency and faithful detail preservation.

  • Takeaways & Limitations

    The framework supports more faithful and efficient high-resolution image editing by using original-source guidance during refinement.

  • Takeaways & Limitations

    The method requires a pre-defined indices prior, and iterative bridge sampling adds non-negligible overhead at high resolutions.

Abstract

from arXiv · show

High-resolution image editing is increasingly demanded in professional workflows, yet existing diffusion-based models remain constrained to resolutions below 1K due to quadratic attention complexity and prohibitive memory requirements. A prevalent workaround employs a two-stage pipeline: editing at low resolution followed by independent super-resolution. However, this approach suffers from two critical issues: information divergence, where hallucinated details contradict the original high-resolution (HR) source, and texture degradation, manifesting as over-smoothed or over-sharpened artifacts. We propose EditBridge, a diffusion bridge framework for efficient ultra high-resolution editing. Unlike conventional diffusion that regenerates from noise, we formulate refinement as structured data-to-data translation from the low-resolution (LR) edited result to its HR counterpart, explicitly conditioned on the original HR source to preserve authentic details. To efficiently incorporate HR source guidance, we introduce a prior-guided block-wise sparse attention mechanism that exploits semantic correspondence from first-stage editing to constrain cross-image interactions to spatially aligned regions, significantly reducing computational overhead. Extensive experiments demonstrate that EditBridge achieves high-fidelity editing with superior perceptual quality at resolutions up to 4K, delivering 3.6--8.4$\times$ speedup at 2K and enabling practical 4K editing in 61 seconds.

1 Introduction

Existing image editing models are generally limited to 1K resolution, while two-stage super-resolution pipelines can diverge from HR source details and degrade textures. EditBridge addresses these issues with HR-guided diffusion-bridge refinement and prior-guided block-wise sparse attention for efficient, faithful editing.

  • Motivation: Most existing image editing models remain constrained to resolutions no higher than 1K because attention has quadratic complexity and substantial memory requirements.The paper identifies computational complexity and memory demands as the primary causes of this resolution limitation.
  • Motivation: Super-resolution pipelines hallucinate high-frequency textures without HR source guidance, causing information divergence from the original and over-smoothed or over-sharpened artifacts.These failures undermine faithful editing by producing inconsistent details and degraded visual fidelity.
  • Method: EditBridge formulates high-resolution editing as diffusion-bridge refinement from an LR edited image to its HR counterpart, explicitly conditioning refinement on the original HR source.This structured data-to-data transition preserves edited semantic content while synthesizing fine-grained textures guided by authentic HR details.
  • Method: Prior-guided block-wise sparse attention uses semantic correspondence from LR editing to route HR-source information through spatially aligned regions instead of exhaustive global interactions.The mechanism reduces token and attention overhead while enabling semantically selective cross-image interaction.
  • Results: 2.2× speedup compared to conventional diffusion-based approaches is achieved while preserving faithful details in a computationally efficient DiT-based framework.The paper presents this as a high-resolution editing system combining computational efficiency with detail preservation.

2 Related Work

Related work spans instruction-based image editing, high-resolution visual generation, and specialized high-resolution editing pipelines. Instruction-based methods evolved from direct instruction-driven editing toward improved datasets and multimodal integration, while generation-oriented acceleration remains insufficient for editing’s strict source-target consistency requirements.

  • Instruction Based Image Editing: Instruction-based editing began with InstructPix2Pix and SuTI, while MagicBrush improved capabilities through high-quality, diverse training datasets and later methods incorporated MLLMs.The cited progression moves from instruction-driven editing to dataset and multimodal improvements.
  • High-Resolution Visual Generation: High-resolution visual generation has progressed from U-Net architectures to DiT backbones, alongside methods designed to accelerate inference at high resolutions.DiT’s attention mechanism is described as effective for modeling complex token-wise dependencies.
  • High-Resolution Visual Generation: Generation-oriented acceleration strategies are sub-optimal for image editing because editing requires fine-grained consistency between source and target images.The passage contrasts pure image generation with editing scenarios that must preserve source-target correspondence.
  • High-Resolution Visual Generation: MobilePicasso addresses high-resolution editing with hallucination-aware training and adaptive tiling, while ScaleEdit introduces a training-free transfer function.These methods are presented as specialized attempts to bridge the gap between high-resolution generation acceleration and editing.

3 Preliminary: Diffusion Bridge

Diffusion bridges model stochastic paths between structured source and target domains, making them suitable for data-to-data translation. EditBridge applies this formulation to HR refinement by mapping an upsampled LR edit to the target HR output.

  • Diffusion Bridge: Diffusion bridges transport samples between two structured data domains through stochastic interpolation, unlike standard diffusion, which starts from an uninformative prior such as Gaussian noise.Given paired endpoints (x0,x1), the bridge defines a noisy interpolation path whose noise peaks at t=0.5 and vanishes at the boundaries.
  • Diffusion Bridge: The bridge velocity is approximated by a neural network vθ trained with a matching objective.The supplied formulation introduces the instantaneous velocity driving the transition and parameterizes it with vθ.
  • Problem Formulation: HR image editing maps an HR source xHR_s and instruction c to an edited HR output xHR_t.The problem formulation defines the desired transformation as HR-to-HR editing under an instruction.
  • Problem Formulation: To reduce direct HR optimization costs, the method uses a coarse-to-fine pipeline with a pretrained LR model G and a downsampled source xLR_s.The LR edit is generated first and then upsampled for HR refinement.
  • Problem Formulation: The HR refinement mapping H fits the bridge formulation by treating the upsampled LR edit as source endpoint x0 and the target HR output as endpoint x1.This reframes refinement as a structured transition between the upsampled LR edit and the desired HR result.

4 Method

EditBridge refines an upsampled low-resolution edit into its high-resolution counterpart through a conditional diffusion bridge guided by the original HR source. Prior-guided sparse attention uses first-stage semantic correspondences to reduce high-resolution cross-domain computation while preserving relevant details.

  • 4.1 Bridge-based HR Refinement: The bridge maps the upsampled LR edited result x̃HR_t to the desired HR edited image xHR_t while conditioning on the original HR source xHR_s.The model learns a conditional velocity field and integrates the probability path from t=0 to t=1.
  • 4.1 Bridge-based HR Refinement: Explicit conditioning on the uncorrupted HR source shortens the generative trajectory and supports faithful recovery of fine-grained details and high-frequency textures.This contrasts with generating high-resolution images from pure noise.
  • 4.2 Prior-Guided Sparse Attention: Standard attention has quadratic complexity O(N^2), creating unsustainable memory footprints and severe latency as token count grows with 2K or 4K resolution.EditBridge addresses this bottleneck with prior-guided sparse attention.
  • 4.2 Prior-Guided Sparse Attention: First-stage cross-domain attention affinity maps establish coarse source-target correspondences, whose maximum-response source anchors form a dense spatial guidance field.The mapping is upscaled by nearest-neighbor expansion to guide high-resolution refinement.
  • 4.2 Prior-Guided Sparse Attention: Cross-domain attention restricts each target query to a local k×k source window centered on its upscaled semantic anchor, while self-attention remains localized within each modality.The multi-modal sequence is [Text|Target Image|Source Image].
  • 4.2 Prior-Guided Sparse Attention: O(N·k^2) replaces O(N^2) for cross-domain attention, reducing inference memory while directing each target patch to its most relevant source counterpart for high-frequency synthesis.The efficiency condition is k^2 ≪ N.

5 Experiments

Experiments show that EditBridge consistently outperforms baseline methods across 1K–4K resolutions while providing substantial inference speedups. Ablations further support prior-guided sparse attention and single-step inference as effective choices for preserving fidelity and perceptual quality efficiently.

  • Evaluation Metrics: Region-aware metrics separately assess fidelity in unedited regions and generative fidelity and perceptual quality in edited regions.M-PSNR, M-SSIM, and M-MSE measure unedited-region fidelity, while M-LPIPS evaluates edited regions.
  • Qualitative Comparison: Qualitative comparisons show richer fine-grained details, improved visual clarity, and stronger structural and semantic consistency with the HR source.Conventional super-resolution methods are described as producing excessive sharpening artifacts and unnatural texture degradation.
  • Sparse Attention vs. Full Attention: Prior-guided sparse attention reduces noisy global interactions by routing queries to semantically aligned anchors, alleviating high-resolution memory and latency bottlenecks.Full attention may achieve slightly higher pixel-level metrics but can introduce source-induced artifacts in local editing.
  • Impact of Inference Steps: N = 1 inference achieves the optimal trade-off between computational efficiency and perceptual quality while preserving high-frequency source structures.The passage attributes possible multi-step degradation at extreme resolutions to accumulated quantization errors.

6 Conclusion

EditBridge is a prior-guided diffusion bridge framework for ultra-high-resolution image editing that treats high-resolution refinement as continuous data-to-data translation. It uses the original high-resolution source as conditional guidance to address information divergence and texture hallucination in conventional two-stage super-resolution pipelines.

  • EditBridge introduces a prior-guided diffusion bridge framework tailored for ultra-high-resolution image editing.
  • The framework formulates high-resolution refinement as a continuous, data-to-data translation process.
  • Original high-resolution source images provide conditional guidance to address information divergence and texture hallucination in conventional two-stage super-resolution pipelines.

Appendices … A User study

A blind user study with 30 participants and 50 image pairs finds that EditBridge receives the highest human preference across Detail Preservation, Realism, and Aesthetics. It significantly surpasses PiSA-SR and HiFlow in Detail Preservation and Realism, supporting its ability to avoid texture hallucination and information divergence.

  • A User study: 30 participants evaluated 50 randomly sampled image pairs in a blind Four-Alternative Forced Choice study.The comparison included EditBridge, direct inference, PiSA-SR, and HiFlow, using full images and 100% cropped patches.
  • A User study: EditBridge consistently achieves the highest preference rates across Detail Preservation, Realism, and Aesthetics.Figure 7 reports that EditBridge consistently outperforms all baselines in human preference.
  • A User study: EditBridge significantly outperforms PiSA-SR and HiFlow in Detail Preservation and Realism.The study attributes this result to prior-guided sparse attention preventing texture hallucination typical of super-resolution pipelines and information divergence common in pure generative models.

B Training Configuration · C Pseudo Code · D Prompts

The supplementary sections specify implementation settings, training and inference procedures for the resolution bridge with sparse attention, and the structured prompting protocol used in experiments. Together, they describe the model inputs, latent transitions, attention-mask updates, and concise JSON-formatted editing prompts.

  • B Training Configuration: Training uses LoRA configuration with rank 128, alpha 128, Gaussian initialization, and all-linear target modules.Additional settings include learning rate 1, bias correction, safeguard warmup, and weight decay 0.01.
  • C Pseudo Code: The training procedure learns a velocity field bridging the ground-truth and conditional distributions.The ground-truth latent zgt is the flow source z0, while the low-resolution target latent zlr tgt is the flow target z1.
  • C Pseudo Code: Training samples a timestep and constructs a fused latent through the Bridge forward process to predict the corresponding vector field.The high-resolution source zhr is concatenated to preserve high-frequency details.
  • C Pseudo Code: Inference progressively denoises from the initial high-resolution source latent z1 toward the target state z0 by solving a stochastic differential equation.Each denoising step downsamples the current latent to re-estimate the sparse attention mask relative to the fixed high-resolution reference, using a Brownian Bridge update.
  • C Pseudo Code: The inference procedure takes the initial latent, high-resolution reference, low-resolution source, text embedding, timesteps, noise scale, and rescale flag, and outputs denoised latent z0.The training procedure instead takes high-resolution source, low-resolution target, ground truth, and text prompt inputs to optimize loss L.
  • D Prompts: The experiment prompts are ordered by their appearance in the main paper and instruct Gemini to produce one concise English editing sentence from one image.The required output is a single JSON object containing an edit prompt, with no explanation or markdown outside JSON.
  • D Prompts: The prompt examples cover scene transformations, object and color changes, style transfers, object removal or insertion, and compositional integration.Examples include changing umbrellas to red, adding a canoe, removing a card, transforming character aesthetics, and integrating a bedside monitor with matched lighting and perspective.

E Limitation and Future Work

EditBridge remains constrained by its dependence on pre-defined indices priors, motivating automated prior estimation. Future work also targets localized post-hoc repair and extending the bridge mechanism to video and 3D generation tasks.

  • Limitations: The method relies on pre-defined indices priors that must be extracted or synthesized before translation, limiting fully automated pipelines.Final output quality is bounded by the precision of these initial guidance signals.
  • Automated Prior Estimation: A learnable prior estimation module could infer structural or semantic indices directly from source images using self-supervised representation learning.This would eliminate manual or external prior acquisition and broaden applicability to “in-the-wild” images.
  • Post-hoc Refinement and Localized Repair: The Diffusion Bridge could support localized resampling within masked regions to repair mode collapse or textural artifacts after image generation.Re-initializing the bridge within a selected area would enable post-hoc refinement and interactive content repair.
  • Generalizing the Bridge Mechanism: Future research will investigate video-to-video editing and 3D asset stylization while preserving temporal and spatial consistency across frames or viewpoints.Extending the bridge trajectory to these domains is identified as a challenging direction for controllable generative tasks.

F More Cases

Additional 1K and 2K visual cases show that EditBridge preserves source fidelity, structural integrity, and subject identity while producing natural, artifact-free edits. The results also demonstrate seamless background integration and a balance between following complex instructions and retaining essential character.

  • Additional visual cases: Additional samples at 1K and 2K resolutions show fidelity to source images and perceptually natural edited regions.The visual results are presented in Figures 8 and 9.
  • Additional visual cases: The Diffusion Bridge preserves original structural integrity and subject identity while integrating synthesized content seamlessly into existing backgrounds.The cases avoid common generative artifacts, including blurriness and texture repetition.
  • Additional visual cases: The extensive cases balance adherence to complex textual instructions with faithful retention of subjects’ essential character.
Loading 2608.18063v1…