- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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
- 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?
- Could you separately ablate low-to-high guidance and high-resolution Halo attention, and quantitatively evaluate the model without the 1K anchor?
- 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?