Equal Dose, Different Coverage:Tracing Rate and Duration in Image Editing
Organizations: University of Science and Technology of China
Abstract
Benchmarks for scientific image editing test whether a model draws the outcome of a described process, leaving open which representations carry its errors and whether correcting them yields the right quantity. We study this question with an airbrush of fixed deposition profile, under which exchanging rate and duration preserves the dose and should preserve the endpoint, yet changes the edited coverage. Matched-norm activation patches across text positions and denoising stages show that, in distilled FLUX with Chinese prompts, representations at unchanged text positions (context) carry this discrepancy: early context transfers 3.51 of the 3.98 percentage-point (pp) signed route difference, against 0.09 pp for equal-norm updates at the two changed digits. Early image latents transmit native (unpatched) route differences across two FLUX configurations and OmniGen, while text-position effects depend on configuration. The trace predicts an intervention: share early context between routes while each retains its own digits. On independent English layouts, this reduces the distilled-FLUX absolute route gap from 1.13 to 0.60 pp, outperforms early digit sharing, and improves agreement more than in base FLUX under directly paired tests. Calibrated evaluation, which prescribes an area for each dose, separates this gain from accuracy: sharing the full prompt encoding produces identical route images with target errors of 8-16 pp, and stating the target percentage directly retains a compressed area response. Localization predicts a consistency gain on new layouts, and calibration measures the quantity error that agreement leaves unresolved.
Figures & tables
| Task | Language | Configurations | Role | Paired units |
|---|---|---|---|---|
| Natural airbrush | Chinese | FLUX distilled | Representation and stage tracing | 12 |
| Natural airbrush | English | FLUX distilled/base, OmniGen | Representation and stage tracing | 8 |
| Natural airbrush | Chinese | FLUX distilled | Independent sharing validation | 12 |
| Natural airbrush | English | FLUX distilled/base | Independent sharing validation | 12 |
| Calibrated airbrush | Chinese | FLUX distilled | Gap and target MAE after sharing | 12 |
| Calibrated airbrush | English | FLUX distilled/base | Gap and target MAE after sharing | 8 |
| FLUX distilled | FLUX base | |||
| Quantity / comparison | Nine units | Eight units | Nine units | Eight units |
| Native absolute gap | 1.642 0.328 | 1.822 0.470 | 0.993 0.420 | 1.180 0.708 |
| Signed transfer and stage interaction | ||||
| Two digits, all steps | +0.170 0.054 | +0.038 0.048 | −0.024 0.061 | +0.020 0.026 |
| Context, matched norm, all steps | +1.401 0.283 | +2.356 0.493 | +0.084 0.153 | +0.004 0.108 |
| Context, shuffled rows, all steps | +0.281 0.506 | −0.236 0.286 | +1.182 1.181 | +0.338 0.392 |
| Chinese FLUX | English FLUX | English FLUX base | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Conditioning | MAE | MAE | MAE | ||||||
| Native | 1.79 0.76 | 0.48 0.10 | 16.03 0.91 | 0.16 0.06 | 0.16 0.02 | 13.35 0.59 | 0.23 0.11 | 0.82 0.19 | 8.33 1.45 |
| Share digits early | 1.61 0.72 | 0.50 0.10 | 16.01 0.89 | 0.12 0.03 | 0.19 0.03 | 13.35 0.59 | 0.26 0.11 | 0.80 0.17 | 8.34 1.44 |
| Share context late | 1.79 0.76 | 0.49 0.10 | 16.03 0.91 | 0.16 0.05 | 0.16 0.02 | 13.35 0.59 | 0.22 0.10 | 0.84 0.19 | 8.33 1.45 |
| Share context early | 1.02 0.50 | 0.32 0.05 | 16.02 0.94 | 0.09 0.03 | 0.17 0.07 | 13.35 0.59 | 0.10 0.04 | 0.96 0.25 | 8.36 1.45 |
| Share full encoding | 0.00 0.00 | 0.00 0.00 | 15.96 0.88 | 0.00 0.00 | 0.00 0.00 | 13.37 0.60 | 0.00 0.00 | 0.00 0.00 | 8.49 1.47 |
| Chinese FLUX | English FLUX | English FLUX base | English OmniGen | |||||
|---|---|---|---|---|---|---|---|---|
| Instruction contrast | Spot | Tank | Spot | Tank | Spot | Tank | Spot | Tank |
| Prescribed gain, 9 to 27 units | +25.00 0.00 | +50.00 0.00 | +25.00 0.00 | +50.00 0.00 | +25.00 0.00 | +50.00 0.00 | +25.00 0.00 | +50.00 0.00 |
| Dose from process alone | −0.27 0.25 | −0.21 0.06 | −0.09 0.10 | −0.08 0.93 | −0.08 0.18 | −0.49 0.33 | +0.02 0.03 | −0.05 0.08 |
| Dose with correct total supplied | +1.16 0.42 | +0.21 0.06 | +0.25 0.10 | −0.56 0.50 | +0.49 0.36 | +0.30 0.51 | −0.27 0.23 | +0.55 0.34 |
| Target percentage alone | +1.11 0.31 | −0.14 0.23 | +4.95 2.05 | +0.64 0.43 | +2.25 1.40 | +0.14 0.08 | −1.04 0.48 | −1.31 0.98 |
| Dose at fixed size cue | +1.26 0.94 | −0.04 0.16 | −0.63 0.29 | −1.16 0.47 | −0.18 0.10 | −0.01 0.76 | +0.01 0.14 | +0.42 0.49 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Early vs. late | Intact vs. shuffled | |||
|---|---|---|---|---|
| 9 units | 8 units | 9 units | 8 units | |
| FLUX distilled | 1.481 | 1.682 | 1.268 | 1.423 |
| FLUX base | 0.815 | 0.955 | 0.660 | 0.843 |
| OmniGen | 0.979 | 1.031 | 0.579 | 0.623 |
| Model / dose | Native | Early C | Early D | Late C |
|---|---|---|---|---|
| Distilled / 9 | 1.311 | 0.651 | 1.241 | 1.290 |
| Distilled / 8 | 0.954 | 0.543 | 0.886 | 0.962 |
| Base / 9 | 0.964 | 1.054 | 1.041 | 0.926 |
| Base / 8 | 1.084 | 0.936 | 1.027 | 1.080 |
| 95% limits | 95% limits | |||
|---|---|---|---|---|
| 0.25 | 0.348 | 0.221 to 0.468 | 0.358 | 0.213 to 0.495 |
| 0.50 | 0.492 | 0.311 to 0.658 | 0.443 | 0.237 to 0.648 |
| 1.00 | 1.184 | 0.724 to 1.592 | 0.908 | 0.474 to 1.335 |
| Control | Gain | 95% CI | Adj. |
|---|---|---|---|
| Native | 0.618 | 0.047 to 1.263 | 0.328 |
| Digits | 0.631 | 0.063 to 1.283 | 0.328 |
| Spatial shuffle | 0.381 | to 0.880 | 0.328 |
Image Editing
Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.
See all 336 papers