ConserveEdit: Source-Anchored Transactions for Drift-Bounded Iterative Image Editing
Organizations: University of California, Los Angeles · Aimakj · College of Computer Science and Technology, National University of Defense Technology · Key Laboratory of Advanced Microprocessor Chips and Systems, National University of Defense Technology
Abstract
Instruction-driven image editors produce convincing local edits in one pass, but inside an iterative refinement loop every accepted output becomes the next input, so unintended changes outside the requested region accumulate. We present ConserveEdit, an editor-agnostic controller that treats each refinement step as a transaction against a contract compiled once from the source image. The contract separates what may change (the authorized region of each target, with a bounded halo for shadows and reflections) from what must survive (protected content with explicit tolerances). Candidates from a frozen editor are composed through a source-anchored mask and committed only if target progress, drift on protected content, and image quality pass as separate tests; otherwise the step is rolled back and repaired. For any editor, this keeps every pixel outside the authorized region bit-exact and bounds drift independently of the number of iterations, whereas no fixed scalar trade-off reproduces the separate tests. On all 1053 MagicBrush test edits with a frozen Qwen-Image-Edit, masking is necessary but not sufficient. A direct call already alters the rest of the image (12.1 dB PSNR outside the mask); composed through the source-anchored mask, it stays bit-exact there and its DINO similarity to the ground truth rises from 0.67 to 0.85, matching ConserveEdit, so this gain comes from the mask. Next to automatic masks, however, masked direct editing changes a protected face or text on 48% of turns, and masked scalar selection avoids this mostly by abstaining (23% constrained success). Because its drift test names the violated entity, ConserveEdit shrinks the support instead (2% violations, 53% constrained success); in a simulator with known ground truth it raises constrained success from 55.0% to 61.8%. Automatic grounding remains the main bottleneck. Code and logs will be released.
Figures & tables
| Controller | CLIP-D | Abstain | L1 | CLIP-I | DINO | PSNR | Calls |
|---|---|---|---|---|---|---|---|
| Editor changes the whole image | |||||||
| Direct (1 call) | 0.058 | 0% | 0.196 | 0.820 | 0.666 | 12.1 | 1.0 |
| Naive loop | 0.058 | 0% | 0.262 | 0.734 | 0.485 | 9.9 | 3.4 |
| Best-of- scalar | 0.057 | 27% | 0.149 | 0.849 | 0.719 | 25.2 | 6.9 |
| Gate, no mask † | 0.000 | 100% | (0.036) | (0.930) | (0.868) | exact | 8.0 |
| Committed through the human mask | |||||||
| Controller | Abstain | CLIP-D | DINO | L1 |
|---|---|---|---|---|
| Direct + mask | 33% | 0.037 | 0.816 | 2.80 |
| FLUX.1 Fill | 33% | 0.025 | 0.769 | 3.19 |
| Scalar + mask | 40% | 0.046 | 0.811 | 2.77 |
| ConserveEdit | 41% | 0.046 | 0.812 | 2.74 |
| Mask | Controller | Viol. | CS | Abstain | CLIP-D | DINO |
|---|---|---|---|---|---|---|
| human | Direct + mask | 1% | 80% | 0% | 0.059 | 0.850 |
| human | Scalar + mask ∗ | 1% | 82% | 11% | 0.079 | 0.847 |
| human | ConserveEdit ∗ | 0% | 83% | 10% | 0.078 | 0.847 |
| auto | Direct + mask | 48% | 27% | 18% | 0.037 | 0.808 |
| auto | Scalar + mask ∗ | 4% | 23% | 68% | 0.028 | 0.861 |
| auto | ConserveEdit ∗ | 2% | 53% | 32% | 0.038 | 0.830 |
| Parameter | SimEdit | Real (MagicBrush) |
|---|---|---|
| Budget / width | 4 / 4 (sweeps: ) | 2 / 4 |
| Clause threshold / progress floor | 0.7 / 0.02 | 0.5 / 0.02 |
| Target predicate | redness on true target | local CLIP-B/32 directional |
| Core support | noisy grounding | human mask, or auto (GD + SAM) |
| Envelope threshold | – | 0.8 (dev split) |
| Halo | 1 px (+repairs, 3 px) | 8 px (+repairs, 3 px) |
| Controller | CLIP-D | Abstain | L1 | CLIP-I | DINO | PSNR | Calls |
|---|---|---|---|---|---|---|---|
| EditThinker [ 30 ] | 0.062 | 0% | 0.184 | 0.833 | 0.708 | 12.5 | 3.5 |
| + same | 0.069 | 0% | 0.042 | 0.900 | 0.871 | exact | 3.5 |
| RePlan [ 44 ] | 0.044 | 0% | 0.080 | 0.901 | 0.837 | 23.2 | 1.0 |
| + same | 0.044 | 0% | 0.035 | 0.931 | 0.882 | exact | 1.0 |
| EditRefiner [ 58 ] | 0.061 | 0% | 0.182 | 0.828 | 0.697 | 13.0 | 2.4 |
| + same | 0.070 | 0% | 0.042 | 0.904 | 0.869 | exact | 2.4 |
| Controller | CLIP-D | Abstain | CLIP-I | DINO | L1 | PSNR |
|---|---|---|---|---|---|---|
| Direct (1 call) | 0.066 | 0% | 0.817 | 0.660 | 0.198 | 12.1 |
| Naive loop | 0.066 | 0% | 0.734 | 0.489 | 0.262 | 9.9 |
| Best-of- scalar | 0.065 | 17% | 0.839 | 0.703 | 0.165 | 20.5 |
| Direct + mask (1 call) | 0.078 | 0% | 0.893 | 0.850 | 0.048 | exact |
| Scalar + mask | 0.094 | 0% | 0.889 | 0.850 | 0.049 | exact |
| Gate, no mask | 0.000 | 100% | 0.930 | 0.870 | 0.036 | exact |
| Res. | Controller | CLIP-D | Abstain | DINO | L1 | PSNR | PSNR |
|---|---|---|---|---|---|---|---|
| Direct (1 call) | 0.061 | 0% | 0.678 | 0.193 | 12.2 | – | |
| Naive loop | 0.059 | 0% | 0.477 | 0.258 | 9.9 | – | |
| Direct + mask (1 call) | 0.070 | 0% | 0.865 | 0.042 | exact | – | |
| Scalar + mask | 0.081 | 15% | 0.865 | 0.041 | exact | – | |
| ConserveEdit | 0.079 | 18% | 0.867 | 0.041 | exact | – | |
| Direct (1 call) | 0.073 | 0% | 0.788 | 0.094 | 22.3 | 22.3 |
| Stratum | Controller | CLIP-D | Abstain | DINO | L1 | |
|---|---|---|---|---|---|---|
| Mask area: small | 351 | Direct (1 call) | 0.066 | 0% | 0.696 | 0.186 |
| Direct + mask (1 call) | 0.065 | 0% | 0.931 | 0.014 | ||
| Scalar + mask | 0.082 | 15% | 0.933 | 0.014 | ||
| ConserveEdit | 0.080 | 16% | 0.935 | 0.014 | ||
| Mask area: medium | 351 | Direct (1 call) | 0.059 | 0% | 0.684 | 0.187 |
| Direct + mask (1 call) | 0.073 | 0% | 0.874 | 0.033 |
| Predicate | Ctrl. | Abstain | add | repl. | attr. | pose | remove | CLIP-D | DINO |
|---|---|---|---|---|---|---|---|---|---|
| CLIP-B/32 (paper) | S+M | 12% | 7 | 18 | 7 | 7 | 55 | 0.083 | 0.849 |
| ConserveEdit | 13% | 8 | 19 | 7 | 8 | 55 | 0.081 | 0.850 | |
| detector, rule as fixed | S+M | 76% | 57 | 100 | 100 | 100 | 100 | 0.026 | 0.858 |
| ConserveEdit | 76% | 57 | 100 | 100 | 100 | 100 | 0.026 | 0.859 | |
| detector + seam, rule as fixed | S+M | 76% | 57 | 100 | 100 | 100 | 100 | 0.026 | 0.858 |
| ConserveEdit | 100% | 100 | 100 | 100 | 100 | 100 | 0.000 | 0.868 |
| Process | Effect | Parameter |
|---|---|---|
| Edit | blend target toward red with strength | fail w.p. 0.15 |
| Spill | edit region = target dilated by px | |
| Shadow | darken true shadow by | – |
| Re-synthesis noise | i.i.d. Gaussian, | |
| Color cast | global offset per channel | |
| Geometric drift | whole-frame shift by – px |
| Controller | CSR (%) | Target | BG PSNR (dB) | Ent. drift ( ) | DriftAUC ( ) | Calls |
|---|---|---|---|---|---|---|
| Naive loop | 5.8 12.7 | 0.873 | 29.5 | 165.83 111.97 | 187.00 | 2.2 |
| Best-of- scalar | 14.5 19.4 | 0.858 | 34.9 | 28.54 23.91 | 45.21 | 5.2 |
| Scalar + mask | 55.0 38.2 | 0.752 | 56.9 | 2.11 5.98 | 2.15 | 6.5 |
| Gate, current ref. | 50.7 31.7 | 0.591 | 48.1 | 4.26 1.65 | 6.09 | 11.9 |
| Gate, no mask | 46.0 31.1 | 0.552 | 48.5 | 3.99 1.51 | 5.86 | 12.2 |
| ConserveEdit w/ current ref. | 61.3 38.6 | 0.753 | 57.1 | 0.55 1.33 | 0.61 | 6.8 |
| Controller vs. ConserveEdit | CSR (pts) | ||
|---|---|---|---|
| Naive loop | +56.0 | 1.39 | |
| Best-of- scalar | +47.3 | 1.15 | |
| Scalar + mask | +6.8 | 0.33 | |
| Gate, current ref. | +11.2 | 0.002 | 0.23 |
| Gate, no mask | +15.8 | 0.33 | |
| ConserveEdit w/ current ref. | +0.5 | 0.249 | 0.12 |
| Controller | IoU | IoU | IoU | IoU | all |
|---|---|---|---|---|---|
| Naive loop | 2.8 | 6.4 | 5.9 | 5.3 | 5.8 |
| Best-of- scalar | 5.6 | 16.2 | 14.3 | 14.0 | 14.5 |
| Scalar + mask | 0.0 | 41.9 | 70.0 | 71.9 | 55.0 |
| Gate, current ref. | 33.3 | 49.6 | 55.3 | 43.9 | 50.7 |
| Gate, no mask | 33.3 | 45.7 | 49.1 | 40.4 | 46.0 |
| Editor-call budget | 4 | 8 | 16 | 32 |
|---|---|---|---|---|
| Naive loop | 7.3 | 7.3 | 7.3 | 7.3 |
| Best-of- scalar | 18.7 | 19.3 | 20.0 | 20.0 |
| Scalar + mask | 30.3 | 49.7 | 50.0 | 50.0 |
| ConserveEdit (full) | 31.0 | 53.0 | 55.7 | 55.7 |
| Best-of- scalar | Scalar + mask | |||||
|---|---|---|---|---|---|---|
| CSR | Ent. ( ) | CSR | Ent. ( ) | |||
| 0.25 | 20.0 | 0.86 | 51.33 | 61.3 | 0.77 | 1.54 |
| 1 | 22.7 | 0.86 | 23.54 | 61.3 | 0.77 | 1.50 |
| 4 | 27.3 | 0.85 | 19.30 | 62.0 | 0.76 | 1.14 |
| 16 | 44.3 | 0.82 | 9.78 | 62.7 | 0.76 | 0.52 |
| 64 | 32.0 | 0.40 | 1.38 | 63.3 | 0.76 | 0.29 |
| 0 | 0.015 | 0.03 | 0.06 | |
|---|---|---|---|---|
| Best-of- scalar | 57.0 | 43.7 | 21.0 | 8.3 |
| Scalar + mask | 57.3 | 57.3 | 57.3 | 56.7 |
| Gate, no mask | 89.7 | 73.0 | 51.7 | 28.0 |
| ConserveEdit (full) | 60.0 | 60.0 | 60.3 | 60.3 |
Image Editing
Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.
See all 336 papers