Finding effective prompts and seeds for one-step image editors requires balancing edit fidelity, source preservation, and computational cost. To address this challenge, we explore a large collection of candidate edits, where each candidate is defined by a unique prompt-seed pair, and generate corresponding images on demand. We present SearchChord, a training-free framework that leverages a large pool of such candidates, combining text-based preselection, image verification, and latent-field recomposition. Our ablation studies across pool sizes up to 200 candidates show that a pool of 100 offers a cost-effective operating point. Compared with feeding all 100 candidates into recomposition, TextSim preselects a subset of 10 candidates. Rather than uniform averaging, we incorporate image-verifier scores to weight each candidate during recomposition. Experiments on PIE-Bench and the proposed CropWeed-EditBench reveal a trade-off: candidate selection improves target-text alignment, while recomposition better preserves unmodified source content. The oracle verifier achieves the best performance in both preservation and target-text alignment. Practical verification partially recovers semantic alignment, but recomposition still compromises edit fidelity, including in agricultural image edits. These results motivate text-first preselection and stronger verification to balance preservation, fidelity, and cost. Code and data will be made publicly available.
Figures & tables
Figure 2: SearchChord pipeline. The arrows indicate the words changed between the source and target prompts, while bracketed placeholders mark where those words appear within the shared sentence.
Field estimator
t
PSNR ↑
CLIP-Edited ↑
Chord (default)
.90
23.52
21.85
Chord (matched center)
.80
24.37
21.66
BiChord ( α=.25 )
.80
23.71
21.67
BiChord-RA
.80
23.53
21.72
Table 1: ChordEdit-based estimator comparison on PIE-Bench. Bold marks the best value in each metric column.
Scoring method
Stage
Image?
Benchmark GT/mask?
Deployable?
Relative scoring cost
Main role
Joint-5 balanced
diagnostic
Yes
Yes
No
Very high
primary headroom estimate
TextSim
pre-generation
No
No
Yes
Very low
reduce field/image budget
EditReward
post-generation
Yes
No
Yes
High
edit quality + instruction following
EditScore
post-generation
Yes
No
Yes
High
visual quality + instruction following
Table 2: Candidate scoring methods and information access. “Image?” indicates whether the edited candidate must already be generated. “Benchmark GT/mask?” indicates whether the method uses mask information for evaluation. “Deployable?” indicates whether the method can be used in a practical deployment setting.
Method / selector
N
T
Fields
Unedited region
CLIP Semantics
Time (s)
PSNR ↑
MSE ×103↓
LPIPS ×103↓
Whole ↑
Edited ↑
Editing/scoring ↓
Est. total ↓
ChordEdit baseline
1
1
1
22.44
8.28
121.06
25.87
22.58
0.49
0.49
Joint-5 oracle: Best-of- N †
100
1
100
23.39
6.95
110.58
27.93
25.22
146.55
175.63
Joint-5 oracle: Mean Top- T †
100
10
100
24.72
5.08
91.02
24.52
21.13
146.98
176.06
Joint-5 oracle: Weighted Top- T †
100
10
100
24.73
5.07
91.00
24.52
21.17
146.74
175.83
Joint-5 oracle: Mean Top- T †
100
100
100
24.53
5.21
91.40
24.13
20.83
146.74
175.83
Table 3: PIE-Bench quality–compute comparison with common proximal refinement and β=1 . All rows below the ChordEdit baseline are SearchChord variants. Estimated total includes prompt-pool construction. Black bold marks overall column-best values; blue bold and teal bold mark the best within the full-verifier (EditReward/EditScore) and TextSim/cascade blocks, respectively.
Figure 3: Effects of candidate budget and field recomposition. (a) PSNR and CLIP-Edited for balanced Joint-5 Best-of- N as the candidate pool grows. (b) Pool composition averaged over 700 edits. Gray counts trials using the original target prompt, teal counts the first trial of each distinct prompt variant, and rose counts further trials of those variants with new seeds. (c) Per-image LPIPS and CLIP-Edited scores for methods evaluated on the same images. (d) Mean scores for candidate selection and field recomposition using N=100 , with ChordEdit shown as a reference.
Unedited region
CLIP Semantics
Edit group
Method
PSNR ↑
MSE ×103↓
LPIPS ×103↓
Whole ↑
Edited ↑
Illumination/weather
ChordEdit baseline
25.87
3.07
95.87
25.39
20.39
TextSim: Weighted Top- T
26.53
2.65
93.23
25.27
19.50
TextSim → EditReward: Weighted
26.47
2.68
93.07
25.32
19.58
Soil appearance/type
ChordEdit baseline
25.24
3.62
103.71
26.52
22.13
TextSim: Weighted Top- T
25.85
3.16
97.41
26.44
20.59
Table 4: Separate illumination, soil, and relative leaf-amount evaluations for 3,387 CropWeed-EditBench plant instances. Bold marks the best values within each group.
Hainan University, School of Computer Science and Technology, Haikou, China · Shanghai University, School of Mechatronic Engineering and Automation, Shanghai, China