cs.CVOct 29, 2026
SaveEditAgent: Rethinking Agentic Image Editing
Organizations: UC San Diego · Adobe Research
Abstract
Agentic image editors decompose complex requests into steps, but when a step fails, they mostly retry with the same narrow set of tools and a judge that scores each output by eye directly. We argue that such agents are bounded by two spaces: the action space of what the agent can do, and the observation space of what its judge can see. On our expert-annotated benchmark, a general-purpose editor yields no usable result in over 40% of the cases, and no judge is reliable across edit types. We present EditAgent, which expands both spaces. It routes each step to general editors and specialized methods, and attaches to each candidate a verification record. With these records, off-the-shelf VLM judges outperform finetuned reward models. Candidates are generated in parallel and judged in one listwise call: experts prefer this to sequential refinement in 78.5% of decided comparisons, at no more than half the judge calls. EditAgent improves two open-weight editors on three benchmarks and serves as a data engine for training stronger editors.
Figures & tables
Figure 2: Where the action space and the observation space fall short, by edit group. (a) Action space: share of the expert-labeled sets in which only general-editor candidates, only specialized candidates, both kinds, or no candidate are usable. (b) Observation space: Winner-Set Hit@1 on the expert-labeled sets with at least one usable candidate. Random is the expected Hit@1 of a uniform random pick.
Figure 3: Overview of EditAgent. Bottom: a request runs as dependent steps, and each accepted result becomes the source of the next step. Top: step 1, enlarged. (A) The planner orders the steps; the analyzer and router select the applicable methods for the current step. (B) The selected methods generate candidates in parallel from the same source image, each with a verification record (evidence). (C) An off-the-shelf VLM compares the candidates with their records and accepts one, or returns feedback for another attempt.
| Judge | Pointwise | Listwise (Ours) | |
| EditScore-7B | 65.9 | – | – |
| EditScore-32B | 68.0 | – | – |
| EditScore-72B | 70.3 | – | – |
| EditScore-7B, Ensemble@4 | 72.7 | – | – |
| EditScore-32B, Ensemble@4 | 73.3 | – | – |
| EditScore-72B, Ensemble@4 | 76.3 | – | – |
Table 1: Comparative judging is better. Overall accuracy (%) on EditReward-Bench [ 35 ] . Pointwise scores each candidate alone; listwise (ours) compares them together in one call.
Figure 4: Examples of the verification record (evidence). (a) A close-up of the edited region reveals the blurred checks left by the removal specialist. (b) Measured brightness, clipping and color shift accurately reveal the performance of the photometric operator.
| Listwise | Evidence | Style | Content | Scene | Photo. | Geom. | Enh. | All |
| 84.4 | 81.7 | 67.5 | 54.2 | 66.4 | 65.0 | 71.6 | ||
| ✓ | 87.1 | 87.1 | 72.0 | 47.2 | 78.6 | 36.8 | 73.6 | |
| ✓ | ✓ | 87.6 | 86.4 | 81.9 | 64.4 | 81.4 | 80.3 | 81.3 |
Table 2: Per-edit-group Winner-Set Hit@1 (%) of Qwen3.8-27B on our diagnostic benchmark.
Figure 5: Verification record (evidence) further lifts off-the-shelf VLM judge. Winner-Set Hit@1 on our diagnostic benchmark.
| Model | EdiVal-IF | EdiVal-CC | EdiVal-O |
| Closed-source | |||
| FLUX.1-Kontext-max † [ 23 ] | 69.49 / 46.89 / 31.83 | 93.93 / 90.90 / 88.40 | 80.79 / 65.29 / 53.04 |
| GPT-Image-1 † [ 41 ] | 73.12 / 54.89 / 38.35 | 81.00 / 77.78 / 75.50 | 76.96 / 65.34 / 53.81 |
| Nano Banana † [ 14 ] | 70.70 / 50.66 / 35.35 | 93.91 / 90.48 / 89.48 | 81.48 / 67.70 / 56.24 |
| FLUX.2-max [ 25 ] | 75.55 / 55.27 / 39.36 | 92.91 / 88.30 / 85.78 | 83.78 / 69.86 / 58.10 |
| Nano Banana 2 † [ 10 ] | 73.89 / 54.17 / 38.61 | 93.54 / 90.52 / 88.61 | 83.14 / 70.02 / 58.49 |
Table 3: Comparison of different editing models on EdiVal. Closed-source model numbers from EdiVal. Each cell reports turn 1 / turn 2 / turn 3 scores (0-100). † Generates at 1024 px, while all outputs are downsampled to 512 512 for evaluation.
Figure 6: The history check extends what the judge observes to earlier turns. Letting in more daylight in turn 2 clears the frosted glass from turn 1. The history check detects this and repairs.
| Model | ComplexBench-Edit | Complex-Edit | ||||||||
| Obj. | Obj.-At. | Obj.-At.-G. | Two-chain | Three-chain | Avg. | IF | IP | PQ | O | |
| Qwen-Image-Edit-2511 | 56.98 | 60.81 | 67.97 | 57.59 | 28.67 | 54.40 | 8.721 | 6.259 | 7.132 | 7.371 |
| + EditAgent | 68.92 | 70.74 | 73.17 | 60.21 | 31.11 | 60.83 | 8.736 | 6.391 | 7.209 | 7.445 |
| FLUX.2-klein-9B | 67.52 | 70.56 | 73.87 | 61.67 | 33.61 | 61.45 | 8.763 | 7.304 | 6.964 | 7.677 |
| + EditAgent | 70.05 | 74.69 | 76.80 | 64.48 | 36.53 | 64.51 | 8.783 | 7.131 | 7.224 | 7.713 |
Table 4: Editing performance on ComplexBench-Edit (scores 0-100) and Complex-Edit (scores 0-10).
| Model | Qwen-Image-Edit-2511 [ 63 ] | FLUX.2-klein-9B [ 27 ] | ||||||
| Turn 1 | Turn 2 | Turn 3 | Inference Cost | Turn 1 | Turn 2 | Turn 3 | Inference Cost | |
| Standalone Editor | 75.97 | 58.58 | 47.64 | 1.1 | 81.37 | 66.30 | 53.20 | 0.5 |
| MIRA † [ 75 ] | 75.79 | 57.59 | 45.59 | 11.3 | 82.57 | 67.08 | 53.76 | 5.6 |
| IMAGAgent [ 49 ] | 76.83 | 60.15 | 49.78 | 10.9 | 82.63 | 68.00 | 55.21 | 4.6 |
| EditThinker [ 28 ] | 76.58 | 60.47 | 49.23 | 9.4 | 82.62 | 67.20 | 55.97 | 4.6 |
| GenArtist † [ 61 ] | 79.51 | 62.94 | 50.54 | 9.5 | 82.94 | 67.48 | 54.74 | 5.2 |
Table 5: Comparison with agentic editors on EdiVal. We report EdiVal-O at turns 1/2/3 and inference cost (H100 GPU hour). All agents use the same base editors and Qwen3.8-27B [ 56 ] as the VLM. Best bold , second underlined . † Re-implemented from the paper.
Figure 7: Multi-turn editing comparison between FLUX.2-klein-9B and EditAgent.
| Baseline | Training Data Source | EdiVal-O Turn 3 | Complex-Edit O | ComplexBench-Edit Avg |
| Qwen-Image-Edit-2511 | — | 49.66 | 7.371 | 54.40 |
| FLUX.2-klein-9B | 48.51 | 7.334 | 55.84 | |
| EditAgent | 51.61 | 7.435 | 57.57 | |
| FLUX.2-klein-base-9B | — | 54.53 | 7.517 | 57.84 |
| FLUX.2-klein-9B | 52.83 | 7.351 | 57.47 | |
| EditAgent | 56.53 | 7.615 | 59.98 |
Table 6: Training data from EditAgent versus data generated directly by FLUX.2-klein-9B. Each baseline model is fully finetuned for one epoch; “—” denotes no finetuning. To bridge the train-eval gap of prompting, we rewrite the input prompts for EdiVal.
| Vanilla & Agent | Action Space | Observation Space | EdiVal-O (T3) | |
| Parallel Comparison | Verification Record | |||
| 47.64 | ||||
| 49.43 | ||||
| 52.37 | ||||
| 53.18 | ||||
| 47.53 | ||||
Table 7: Component ablation on EdiVal with Qwen-Image-Edit-2511. Rows with replace the action space with identical copies of the general editor. The last row shows the full EditAgent.
Figure S1 : Diagnostic sets in which only the specialized candidate is usable (✓). The general editor turns the bear around, puts the snow on the TV screen, and changes the saturation.
| Group | Edit Type | Definition | General | Hit@1 |
| Content | Add Text | add readable text/writing/signage somewhere it fits naturally | 19 | 19 |
| Remove Text | remove existing text/logo/signage (only if some is visible) | 15 | 15 | |
| Replace Text | replace visible text with different text (only if text is visible) | 14 | 15 | |
| Change Text Style | change color/font/size/style of visible text | 9 | 12 | |
| Add Object | add a new object/animal/person that fits the scene | 18 | 18 | |
| Remove Object | remove an existing object/person/animal | 15 | 17 |
Table S1 : The 30 edit types of the diagnostic benchmark and the definitions given to the instruction writer. Each type has 20 sets. General counts the sets with at least one usable general-editor candidate from either run, and Hit@1 counts the sets used for Winner-Set Hit@1.
| Role | Model | Input | Output | Prompt |
| Planner | VLM | Request of the turn, the version graph and the carried requirements | Goal and ordered steps, each with its operation, target, regions to keep, parameters and new requirements | |
| Analyzer | VLM | One step, source image of the step, carried requirements, feedback on earlier attempts, reports of earlier steps, parameter documentation of the applicable methods | Explicit instruction stating what changes and what stays, operation, target, regions to keep, parameters, acceptance criteria, candidate count | |
| Router | Code | Analyzed step | Methods, each contributing at least one candidate ( Tab. S4 ) | None |
| Grounder | VLM and SAM 2.1 | Image and a referring expression | One box per referred object in coordinates normalized to , refined into a mask by SAM 2.1 and reviewed by the VLM | and |
| Editing Methods | Tab. S5 | Source image of the step, instruction, target, parameters | Candidate image and its execution artifacts, such as a mask | None |
| Judge | VLM | Request of the step with its scope, acceptance criteria and regions to keep, source image, and all candidates with their verification records in a shuffled order | A note, a pass/fail verdict and a score from 1 to 10 for every candidate, then the best candidate and a ranking |
Table S2 : Roles in EditAgent. In the benchmark experiments and in the runs that produced the diagnostic candidates, every VLM role uses Qwen3.8-27B with the settings of Tab. S3 . The data engine uses Qwen3.8-27B as the judge ( Sec. H ).
| Setting | Value |
| Candidates per Attempt | at most 3 |
| Attempts per Step | 3 |
| Re-Plans per Turn | 2 |
| Repairs per Turn | 1 |
| Repair Confidence | at least 0.4 |
| Seed of Candidate in Attempt |
Table S3 : Default settings of EditAgent, used in all benchmark experiments and in the data engine (EdiVal images at px).
| Step | Operation | Methods |
| Remove a target | remove | General Editor Removal Specialist |
| Recolor a target | recolor | General Editor Masked Recolor |
| Restyle a target | restyle | General Editor Masked Local Editor |
| Replace or add an object | replace, add | General Editor |
| Move a target | move | General Editor Move Specialist |
| Resize a subject | scale_subject | General Editor Subject Scaling |
Table S4 : Methods raced for each kind of step, with the operation that the planner and analyzer assign to it.
| Method | Backend | Established by Code or Construction | Left to the Judge |
| General Editor | The backbone (FLUX.2-klein-9B or Qwen-Image-Edit-2511) edits the whole frame. | None | Requested change, preservation, realism. |
| Masked Local Editor | Target mask from grounding, SAM 2.1 [ 46 ] and VLM review. The general editor edits the frame, and only its pixels inside the dilated and feathered mask are kept. | Unchanged outside the feathered mask (code). | Requested look in the mask, seam realism. |
| Masked Recolor | As above, with a tighter mask. | Unchanged outside the feathered mask (code). | Requested color, coverage, seam. |
| Removal Specialist | OSOR [ 80 ] with the reviewed target mask, whose output is used directly. | Canvas size unchanged (code). | Target and attached parts removed, background reconstructed, neighbors and supports kept. |
| Move Specialist | A diffusion transformer trained for object movement, conditioned on the source mask and a destination box designed by the VLM. Pixels outside the source and destination regions are restored from the source. | None | Object gone from its origin, same object at the destination, size consistent with depth, blending. |
| Subject Scaling | Reviewed mask, ViTMatte [ 69 ] alpha matte, uniform rescaling anchored at the feet, inpainting of the old footprint, contact shadow and color harmonization. A second candidate refines the composite with the general editor. | Every pixel outside the old and new subject and its shadow equals the source (code). No candidate whose scaled subject would cross a frame edge (construction). | Size change in the requested direction, recognizable subject, no halo. |
Table S5 : Methods of the capability pool, each with its backend, the claims established by code or by construction, and what is left to the judge. A candidate that breaks its guarantee is removed before judging, and the record of each remaining one carries both columns ( Sec. F ).
| Measurement | Used For | Definition | Shown to the Judge |
| Visible Change | Every step | Share of pixels whose CIE76 color difference to the source exceeds 8, after a Gaussian blur with px of both images (no blur for detail edits). | One value per candidate, e.g . “visible change 4.2% of the frame” for a small local edit. |
| Change Map | Every step | Per-pixel to the source, shown as brightness on one scale shared by all candidates. | An image per candidate, black where a pixel is unchanged and brighter where it differs more. |
| Close-Up | Localized steps | The region that the request names, localized once in the source. The crop pads its box by 30% of its size and covers at least 30% of the frame on each axis. | A BEFORE | AFTER image per candidate, at the same place and scale for all. |
| Photometric Statistics | Tone, color and lighting | Brightness is the mean CIELAB and contrast the standard deviation of . Saturation is the mean chroma . Clipped highlights and crushed shadows are the shares of pixels whose brightest channel is at least 250 or at most 5, and color shift is the change of mean and . | One line for the source and one per candidate, e.g . “brightness 0.62, contrast 0.24, saturation 18.3, clipped 2.1%, color shift ”. |
| Sharpness and Noise | Sharpening, denoising, blur | Variance of the Laplacian and the noise level of the luminance, with both images downsampled to a common size, and for a localized step also the sharpness inside the target relative to the source. | Ratios to the source, e.g . “sharpness 1.84, noise 0.71” for a sharpened candidate. |
| Structure Kept | Global edits that keep the layout | Mean luminance SSIM between the source and the candidate over the frame. | One value per candidate, where 1 means identical structure and lower values mean more change. |
Table S6 : Measurements of the verification record, chosen by the operation of the step. Code computes them from the pixels in the same way for the source and every candidate. The change map and the close-up are images, and every other measurement is one line of text per image.
| Agent | How It Is Run |
| GenArtist † | Official prompts. The unreleased editing loop is re-implemented, and its tools become backbone edits, masked for local tools. |
| MIRA † | Trained policy model not released. The planning loop of the paper runs zero-shot, with Qwen3.8-27B serving as its policy model. |
| IMAGAgent | Released code and prompts, with closed models replaced by Qwen3.8-27B and the backbone, and inpainting by masked edits. |
| EditThinker-8B | Released code and 8B thinker. |
| VisionDirector-TF † | Trained planner not released. A training-free loop uses the released verifier prompt and runs local goals as masked edits. |
| RISE-Agent | Released code, with Qwen3.8-27B as planner and verifier and without web search. |
Table S7 : How each agentic baseline of Tab. 5 of the main paper is run. Every agent edits with the same backbone and uses Qwen3.8-27B wherever it calls a multimodal LLM. † Re-implemented.
| Model | EdiVal-O | CBE | CE-O |
| Qwen-Image-Edit-2511 | 75.97 / 58.58 / 47.64 | 54.40 | 7.371 |
| + EditAgent (Qwen3.8-27B) | 80.64 / 67.02 / 55.46 | 60.83 | 7.445 |
| + EditAgent (Gemma-4-31B) | 80.34 / 66.28 / 54.17 | 60.53 | 7.457 |
| FLUX.2-klein-9B | 81.37 / 66.30 / 53.20 | 61.45 | 7.677 |
| + EditAgent (Qwen3.8-27B) | 84.46 / 71.10 / 59.45 | 64.51 | 7.713 |
| + EditAgent (Gemma-4-31B) | 85.17 / 71.64 / 60.25 | 63.44 | 7.715 |
Table S8 : EditAgent with Qwen3.8-27B or Gemma-4-31B as its internal judge, under unchanged external scoring. The columns are EdiVal-O at turns 1 / 2 / 3, the ComplexBench-Edit average (CBE) and the Complex-Edit overall score (CE-O).
| Benchmark | Backbone | Masked Local Editor | Masked Recolor | Removal Specialist | Photometric Operators | Style Filter | Zoom and Other | General Editor | Accepted Steps |
| EdiVal | Qwen-Image-Edit-2511 | 9.2 | 6.7 | 8.1 | 0.0 | 0.0 | 0.6 | 75.4 | 1,253 |
| FLUX.2-klein-9B | 7.8 | 7.7 | 3.3 | 0.0 | 0.0 | 0.0 | 81.2 | 1,628 | |
| ComplexBench-Edit | Qwen-Image-Edit-2511 | 10.1 | 5.2 | 16.0 | 0.0 | 0.1 | 0.0 | 68.6 | 1,891 |
| FLUX.2-klein-9B | 8.3 | 5.3 | 7.2 | 0.0 | 0.0 | 0.0 | 79.2 | 2,246 | |
| Complex-Edit | Qwen-Image-Edit-2511 | 6.0 | 4.6 | 3.7 | 5.4 | 4.4 | 1.2 | 74.8 | 28,712 |
| FLUX.2-klein-9B | 4.8 | 4.7 | 1.8 | 4.5 | 3.6 | 1.1 | 79.5 | 33,224 |
Table S9 : Accepted steps by method for each backbone and benchmark (% of the accepted steps on the final edit chain of each request). Masked Local Editor and Masked Recolor run the backbone inside a mask. Removal Specialist, Photometric Operators (whole-frame and local) and Style Filter do not use the backbone. Zoom and Other covers Zoom-In Crop, Canvas Extension, Flip and Exact Rotation, Rotation Tool and the two style-transfer methods.
| Group | Turns | Accepted (%) | First (%) | Attempts | Candi- dates |
| Content | 46,675 | 94.1 | 79.6 | 1.58 | 4.06 |
| Style & Effects | 22,864 | 94.6 | 82.0 | 1.58 | 4.03 |
| Scene | 28,595 | 92.2 | 70.4 | 2.23 | 5.15 |
| Photometric | 22,841 | 93.8 | 65.5 | 2.13 | 3.35 |
| Geometry | 31,390 | 83.3 | 69.6 | 2.56 | 5.29 |
| Enhancement | 11,441 | 89.9 | 63.8 | 2.59 | 4.63 |
Table S10 : Data engine statistics by edit group over all 163,806 executed turns. Accepted is the share of turns accepted by the judge, and First the share whose steps are all accepted at the first attempt, without re-planning. Groups follow Tab. S1 . Attempts and candidates are means per turn.
Figure S3 : Multi-turn editing comparison with FLUX.2-klein-9B.
Figure S4 : Multi-turn editing comparison with GPT-Image-2.
Figure S6 : Single-turn editing comparison with Qwen-Image-Edit-2511.
Figure S7 : Single-turn editing comparison with FLUX.2-klein-9B.
| Term in the Paper | Name in the Prompts |
| Candidate | TRIAL n |
| Source Image of a Step | BEFORE |
| Change Map | CHANGE-MAP |
| Close-Up | CLOSE-UP , THEN | NOW close-up |
| General Editor | edit_model |
| Masked Methods | masked_composite |
Table S11 : Terms of the paper and their names in the prompts.
More in this topic
Image Editing
Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.
See all 318 papers