Instruction-based image editors still primarily operate at 1K, whereas practical editing increasingly requires faithful 2K/4K outputs. Dense high-resolution adaptation preserves the editing trajectory but incurs expensive target-source attention; editing at 1K followed by external refinement is cheaper but can separate detail restoration from the editing trajectory. We introduce FLASH, a native high-resolution editor based on sparse cross-resolution denoising. A frozen 1K branch provides a semantic and structural anchor, while a target-resolution branch reconstructs local detail within the same Transformer. An asymmetric topology retains full 1K attention, guides high-resolution queries with spatially aligned low-resolution noisy-target states, and confines high-resolution target-reference interaction to local Halo neighborhoods. Resolution-specific diffusion-state mapping and spatial YaRN adapt the pretrained editor to larger grids. An optional, training-free progressive path replaces early high-resolution denoising with a 1K prefix and completes refinement in the same editor. On VINS-4KEval, the direct model achieves the strongest structural fidelity among the compared methods at both 2K and 4K. A separate 4K forward benchmark with R = 1 records 4.2x faster total forward execution and 25.3x faster attention than dense target-reference attention. Progressive inference reduces end-to-end 4K latency by 30.5%, with a modest quality cost.
Figures & tables
Figure 1: Three routes to high-resolution editing: dense native 4K editing, 1K editing with super-resolution, and our FLASH framework.
Figure 2: Per-forward runtime across resolutions. Each bar shows attention and other forward computation. Dense attention uses target–reference sequences; the separate sparse diagnostic uses R=1 and h=n . At 4K, it is 4.2× faster overall and 25.3× faster in attention. These timings do not represent the main R=3 configuration.
Figure 3: Overview of FLASH. A frozen 1K branch anchors global editing, local high-resolution attention reconstructs detail, and progressive inference replaces early high-resolution steps with 1K denoising.
Method
HaarPSI ↑
mPSNR ↑
mSSIM ↑
mMSE ↓
mLPIPS ↓
Time ↓
2K Resolution
FaithDiff ( Chen et al., 2025a )
0.3840
21.7671
0.6900
0.01700
0.3101
29.43
DiT4SR ( Duan et al., 2025 )
0.3761
21.3047
0.6966
0.01757
0.3130
52.91
PiSA-SR ( Sun et al., 2025 )
0.3651
20.7503
0.6647
0.01843
0.3177
16.63
TSD-SR ( Dong et al., 2025 )
0.3600
20.1663
0.6468
0.02035
0.3338
16.89
VINS ( Chen et al., 2026b )
0.4164
22.2907
0.7065
0.01693
0.2486
100.82
Table 1: Main 2K/4K comparison. Bold and underline mark the best and second-best values per resolution; time is not ranked. HaarPSI uses the full image. mPSNR/mSSIM/mMSE use the keep region; mLPIPS uses the edit region, or the full image for global edits. Shaded rows do not use external super-resolution. Time is end-to-end latency in seconds, excluding model loading.
Figure 4: 2K outputs for “Add a white dove flying above the bus.” Direct and progressive predictions are compared with edit–super-resolution cascades and VINS; crops show the edit and nearby text detail.
Support
h
HaarPSI ↑
mPSNR ↑
mSSIM ↑
mMSE ↓
mLPIPS ↓
Time ↓
R=1
2
0.4234
22.4356
0.7774
0.01591
0.2966
94.42
R=2
4
0.4213
22.4387
0.7732
0.01650
0.2977
99.64
R=3
6
0.4235
22.5668
0.7780
0.01616
0.2976
108.95
Table 2: 2K synchronized Halo ablation. Increasing R expands low- resolution support and the matched high-resolution Halo of width h=Rn ; time is not ranked.
Transition
HaarPSI ↑
mPSNR ↑
mSSIM ↑
mMSE ↓
mLPIPS ↓
Time ↓
Direct
0.4235
22.5668
0.7780
0.01616
0.2976
108.95
K=8
0.4173
22.4866
0.7768
0.01551
0.3132
102.32
K=12
0.4137
22.4258
0.7733
0.01554
0.3165
88.19
K=16
0.4097
22.4041
0.7693
0.01563
0.3157
74.84
Table 3: Progressive transition ablation at R=3 . K is the initial 1K-prefix length; Direct denotes joint denoising. Time is end-to-end latency in seconds, excluding model loading.
Figure 5: 2K Halo ablation under a shared instruction. R=3 is compared with no Halo (R=0,h=0) , highlighting cross-window continuity and texture.
Figure 6: 2K anchor ablation. The anchor-free configuration removes the 1K branch; the full model retains instruction-driven scene layout.
Figure 7: 2K lifting ablation at R=3 . RGB/VAE and latent configurations share the same progressive denoising and differ only in the lifting operator.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Item
Setting
Base model
FLUX.2 Klein Base 4B
Low-resolution editor
1K supervised fine-tuned checkpoint, merged and frozen
Training data
Full VINS-120K split; 131,010 examples
High-resolution branch
Rank-32 LoRA on attention and FFN projections
Optimizer
AdamW; learning rate 2×10−5 ; weight decay 0.01
Warm-up
100 optimizer steps
Appendix
Table 4: Settings used in this work.
Query \ Key
Text
LN
LR
HN
HR
Text
✓
✓
✓
–
–
LN
✓
✓
✓
–
–
LR
✓
✓
✓
–
–
HN
✓
local
–
local
local
HR
✓
–
–
local
local
Appendix
Table 5: Query–key adjacency of the evaluated configuration. Local denotes the Chebyshev- R parent-cell neighborhood ( KL=(2R+1)2 keys) or the Halo support ( KH=(n+2h)2 keys); ✓ denotes the native global pattern.
Figure 8: Additional 2K comparisons on VINS-4KEval under the shared GroundingDINO+SAM2 masks. Each group shows the source, instruction, and method outputs. Local structure, small text, and thin boundaries reveal differences between 1K-upsampled and native high-resolution editing.
Figure 9: Additional 4K comparisons on VINS-4KEval under the same shared-mask protocol, contrasting native target-resolution detail with 1K editing followed by super-resolution.