acceptodds
Withdrawn from ICLR 2027

Efficient and Faithful High-Resolution Image Editing with Sparse Cross-Resolution Denoising

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Instruction-based image editors remain predominantly 1K systems, yet practical editing increasingly demands faithful 2K/4K outputs. Direct high-resolution adaptation preserves a unified editing trajectory but incurs dense target–source attention over rapidly growing token sequences; low-resolution editing followed by external refinement is cheaper but can separate detail restoration from the editing trajectory. We introduce sparse cross-resolution denoising, a native high-resolution editing paradigm that retains a frozen 1K stream as a semantic and structural anchor and trains a target-resolution stream to reconstruct local detail within the same Transformer. An asymmetric sparse topology preserves full 1K attention, sends spatially aligned noisy-target guidance from the low-resolution trajectory to high-resolution queries, and restricts high-resolution target–reference interaction to local Halo neighborhoods. Resolution-specific diffusion-state mapping and spatial YaRN adapt the editor to larger grids. A training-free progressive path further replaces an initial portion of target-resolution denoising with a 1K prefix and completes refinement in the same model. On VINS-4KEval, we achieve the highest reported metrics at both 2K and 4K; in a matched 4K forward benchmark, the sparse topology is \(4.2\times\) faster overall and \(25.3\times\) faster in attention than dense attention. Progressive inference further reduces end-to-end latency by \(30.5%\) at 4K while retaining competitive quality.

Primary area:
generative models
Keywords:
diffusion, efficient, image editing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UK92OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a dual-resolution approach to efficient high-resolution image editing. A frozen 1K stream provides the editing trajectory, while a trainable high-resolution stream recovers local details through sparse cross-resolution and Halo attention. The authors also introduce progressive inference, which starts denoising at 1K before switching to the target resolution. Experiments cover 2K and 4K editing, with comparisons against edit–super-resolution pipelines and high-resolution adaptation.

Critical strengths
  1. The problem is practically relevant. High-resolution editing needs to preserve small details while following the instruction, and dense attention becomes expensive at these resolutions.
  2. The division between a low-resolution editing anchor and local high-resolution refinement is reasonable. The paper also considers sparse-kernel implementation, rather than only presenting an attention mask.
  3. The matched forward-pass benchmark provides useful evidence that the sparse topology reduces computation relative to dense attention. Progressive inference offers an additional way to trade quality for latency.
Critical weaknesses
  1. My main concern is that the reported quality metrics do not establish whether the requested edit was completed. Appendix C computes edit-region LPIPS between the output and the original source, with lower values treated as better. The other main metrics also measure source preservation. Simply returning the source image would score ideally on these metrics, even if the instruction was ignored. Preservation is important, but the main comparison needs a separate measure of editing success.
  2. The ablations do not fully isolate the central design choices. Table 2 changes the low-resolution support and high-resolution Halo together, so their individual effects remain unclear. Removing the low-resolution anchor is mainly evaluated qualitatively. I would like to see quantitative evidence for how much each pathway contributes, particularly to instruction following.
  3. The progressive inference procedure is described inconsistently. Section 3.4 transfers the clean estimate through RGB resizing and VAE re-encoding, with latent interpolation described as an ablation. Appendix A.4 and Algorithm 4 instead specify latent bilinear interpolation as the default. These are different inference procedures, and it is unclear which produced the reported results.
Questions for authors
  1. Could you report instruction-following accuracy or a human evaluation alongside source-preservation metrics? How would an unchanged-source baseline score under the current protocol?
  2. Could you separately ablate low-to-high guidance and high-resolution Halo attention, and quantitatively evaluate the model without the 1K anchor?
  3. Which transition operator was used for the main progressive results: RGB/VAE lifting or latent interpolation? Could you reconcile the main text and appendix and report the quality and runtime of both?
Reviewer_5qQ2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a method to extend instruction-based image editing to high-resolution settings, especially in 2K and 4K. In order to address the high computational cost of directly applying such techniques to HR images, the proposed method uses sparse cross-resolution denoising that splits the process into LR and HR branches. The LR branch uses the information generated by the pretrained LR editor, while the HR branch uses that information as an anchor to 'transfer' the editing to HR images.

Critical strengths

None. To be clear, I'm not claiming that this is necessarily a poorly written paper. The reason for this assessment is that I am not familiar with this field enough to appreciate the usefulness of the proposed architecture.

Critical weaknesses
  • Lack of details: Overall, I found this paper to be too abstract to be read as a paper in representation learning. There are many details missing that are necessary to understand the paper. For example, section 3.3 describes the key idea behind the system, but the modeling details are only given as high-level descriptions. There are no diagrams, equations, or pseudocodes to describe the overall system. Yes, the appendix does contain three pseudocodes, but only one (Alg. 1) seems to be directly related to the key idea, and even that is missing details: (1) 'shift' operator, mu_H, mu_L are undefined. (2) The statement 'high-resolution segment of X' is ambiguous. (3) 'sparse_cross_resolution_transformer()' function is undefined, which is critical because that seems to be the core of this algorithm. Is it Algorithm 3? Also, many symbols and functionalities are being used without proper definition. Not including the pseudocode argument above: (1) a Transformer is suddenly appearing in line 184 without prior mentioning, and (2) also abruptly followed by a vθ(...)v_\theta(...) function. (3) ϕ()\phi() is undefined in line 235. (4) The role of μ\mu in line 223 is unclear. (5) Subscripts and superscripts of the variables appearing in the equation of line 310 are unexplained. (6) RRGBR_{RGB} is undefined in line 313.

  • More importantly, I feel that there is a lack of rigorous contribution at the level that is expected from successful ICLR submissions. The proposed solution is more of an engineering practice than of a novel algorithmic contribution, or even an empirical discovery. Existing concepts such as diffusion, RoPE, YaRN, and Halo attention are assembled to form a pipeline of a system. Again, this doesn't mean this is a bad approach. I'm just saying that there are other good conferences that value such approaches than ICLR does. Or maybe there are such contributions that I don't see, but even so, the text does not do a good job at explaining them properly.

Questions for authors
  • Could the authors clarify the issues raised in the first weakness?
  • In defense of the second weakness written above, could the authors provide justification for the novelty of the proposed approach?
Reviewer_9TKoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To address the issue of surging computational costs in instruction-driven image editing at 2K/4K high resolutions, this paper proposes a sparse cross-resolution denoising framework. The method utilizes a frozen 1K-resolution branch as an anchor for global semantics and structure, preserving global information through asymmetric sparse attention. It further incorporates local high-resolution "Halo" attention and spatially aligned low-to-high-resolution guidance to achieve efficient reconstruction of local details. Additionally, the authors introduce a progressive inference strategy to further reduce end-to-end latency for 4K editing. Experimental results demonstrate that this approach enhances computational efficiency while maintaining competitive editing quality.

Critical strengths
  • Efficient Sparse Computational Topology: By employing "halo" local attention and cross-resolution guidance, the reliance on global dense attention for high-resolution (2K/4K) inputs is reduced, effectively lowering computational overhead.

  • Authors compare the impact of different low-resolution guidance radii (RR) on quality and computational time, and also evaluate both direct and progressive inference. This helps in understanding how the local support range and the timing of the transition influence the efficiency-versus-quality trade-off of the method.

Critical weaknesses
  • The core experiments utilize FLUX.2 Klein Base 4B. Although this setup is of a certain scale, relying on a single editing backbone makes it difficult to fully verify whether the proposed attention topology is applicable to other architectures or to models with different editing capabilities.
Questions for authors
  • It is suggested that the authors include additional generalization experiments using different backbone architectures or independent datasets. This would help clarify whether the method’s superior performance gains stem from the fundamental innovation of the sparse cross-resolution design, or if they are contingent upon a specific combination of backbone architecture and training data.

  • For edits involving large-scale geometric transformations or substantial object restructuring, could the 1K anchor constrain the high-resolution branch's ability to accommodate significant structural changes?

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

Related papers

Discussion (0)

Sign in to comment.