acceptodds
Withdrawn from ICLR 2027

Unifying Task Reformulation Pipelines into an Image Editing Harness

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent work has shown that external pipelines can substantially improve image editors by reorganizing how editing requests are processed. However, existing pipelines are typically designed around specific weaknesses or failure modes. We collect and devise several pipelines, observe their behavior on five benchmarks, and identify three characteristic behaviors. Their benefits are request-dependent and complementary; unnecessary interventions can degrade otherwise successful edits; and naively combining multiple pipeline guidances can introduce conflicts rather than accumulate their benefits. Motivated by these observations, we propose a unified image-editing harness that selectively configures task reformulation for each request. Specifically, a two-stage router selects an appropriate reformulation, on-demand skills provide pipeline-specific guidance, and shared execution controls manage visual context, tool interventions, and recovery. We evaluate the resulting system with FLUX.2-klein-9B as the primary editor across five image-editing benchmarks, against a broad set of strong baselines, together with module ablations and cost analysis. Without updating the primary editor's parameters, our harness substantially improves its editing performance and approaches the quality of Nano Banana Pro at a lower inference cost.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Instruction-Based Image Editing, Task Reformulation, Multimodal Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TUMsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a routing method to use different image editing frameworks. The proposed method achieves better performance than a strong proprietary baseline, Nano Banana Pro, with a small open-source model.

Critical strengths
  1. The analysis in Section 3 reveals how different methods can cooperate to improve editing quality.
  2. The experiment demonstrated improvement on multiple benchmarks.
Critical weaknesses
  1. The paper is incomplete with multiple errors:
  • In Appendix A, experiment details are not provided, but there is a placeholder message saying "[Pending: Confirm the Banana transfer model and all API identifiers, all baseline editor/planner roles, the router and captioner versions, and the Qwen3.8 planner checkpoint.]"
  • On Figure 2, Step 2 says "Change the blue trash can to green", which is unrelated to the image.
  • In the last row of Figure 3, there is no highlighted red outline in the original image, despite the prompt saying so.
  • The first stage was called query profiling in Figure 2 and Section 4.2, but is called captioning later.
  1. The reported performance doesn't match the official numbers:
  • On PICABench, the performance of Nano Banana Pro is previously reported as 66.16 by the PICABench paper, but the performance is 57.89 in this paper.
  • On VIBE, the performance of Nano Banana Pro is reported as 65.15 by the official VIBE-Benchmark repository, but the performance is 76.51 in this paper. Also, the performance of FLUX.2-klein is 44.94, way higher than the official performance of the stronger model FLUX.2-dev (30.95). This may arise due to the mismatch of scoring models. The paper uses Qwen3.8-27B, but the official benchmarks used various different scoring models.
  1. VIBE is an image editing benchmark with 1,034 samples, but Table 4 reported only 900 samples. Similarly, ImgEdit-Basic has 734 samples and ImgEdit-UGE has 47 samples according to the original paper, but Table 4 reported 737 and 45, respectively.
  2. Since the router component is the primary contribution of the paper, the experiment scale is currently too limited, as the only configuration is with 3 certain methods to route to. The effectiveness of the router should be demonstrated with more/fewer number of methods, or an alternative set of 3 methods. For example, the existing literature has a large number of localized editing methods, so it would help if the paper can further demonstrate that routing works using different methods of localized editing.
  3. Overall, using a harness on FLUX.2-klein to achieve better performance than Nano Banana Pro is a bold claim. It seems that the current routing scheme (rewrite, local, or spatial) is unable to close fundamental gaps between the two models, such as in text editing (see the "Text" column of Table 3 and Table 4 in https://arxiv.org/pdf/2608.14546). It would be great if this can be acknowledged as a limitation.
Questions for authors
  1. Could you provide an analysis about why the reported performances significantly differ from the official numbers?
  2. Could you clarify why the size of the benchmarks differ from official numbers?
  3. Could you clarify what is "a Banana editor" in Section 5.5? Its performance of 3.150 in Table 6 (Banana Direct) does not match the performance of 3.568 in Table 3 (Nano Banana Pro). Is it a different model?
Reviewer_NbuaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a framework/harness that uses a frozen image editor (FLUX.2-klein-9B in their experiments) and routes image editing requests to one particular branch (instruction rewriting, crop-and-paste local editing, segment-and-rearrange spatial editing) depending on the type of the request. The idea is well motivated and the usefulness of different branches is shown in Section 3, in particular in Table 1.

In Section 5, the paper quantitatively shows that the proposed harness beats direct FLUX editing, and other methods that use only one particular type of task reformulation (e.g., EditThinker, RePlan, and GenArtist) and that it is mostly on-par with a commercial image editor (Nano Banana Pro) but for a much cheaper price (~1/15 of the per-image cost). The paper is further evaluated qualitatively in Figure 3 and in the supplementary showing its usefulness in various image editing results.

Critical strengths
  1. The paper presents a somehow simple and straight-forward, but still well motivated and engineered framework/harness for image editing. Routing a request to the branch that has the highest chance for success appears almost trivial, but I can imagine that if the number of branches grows it makes more and more sense to choose the method with the best chance to give good results.

  2. The paper is well-written, well motivated, all steps are well-justified.

Critical weaknesses
  1. Even though the proposed framework seems to be useful, I would still consider the contribution to be minor. It remains somehow unclear why one could not just run the image on all branches and simply pick the best result, the total cost would still be cheaper than using Nano Banana Pro alone.
Questions for authors
  1. It was not completely clear to me from reading the paper if the competing baseline methods were using the exact same image editor, e.g., was EditThinker and RePlan also evaluated using FLUX.2-klein?

  2. What are the average number of editor calls and agent steps per request? How does the harness compare with Direct best-of-N under the same GPT-5.6 Luna selector at matched editor calls?

Reviewer_UEWGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an inference-time harness that wraps an existing image editor and, for each request, picks how to reformulate the task. A pilot study on 480 requests compares three pipelines (instruction rewriting, crop-edit-paste, and segmentation-based spatial editing) with direct editing. It reports that their benefits differ by request, that unneeded steps can hurt, and that merging all guidance into one prompt works worse. Based on this, a two-stage router picks one pipeline or direct editing, an MLLM planner (GPT-5.6 Luna) follows the matching skill prompt to call the editor and tools, and the system ends with VLM verification and final candidate selection. With FLUX.2-klein-9B as the editor and Qwen3.8-27B as the judge, the paper reports gains over direct editing on all five benchmarks (e.g., VIBE from 44.9% to 74.3%), better scores than EditThinker, RePlan and GenArtist, and scores above Nano Banana Pro on four of five benchmarks at lower cost.

Critical strengths
  1. The gains with a fixed editor are large and consistent. Without changing FLUX.2-klein-9B, the system improves all five benchmarks, for example VIBE from 44.9% to 74.3% and ImgEdit-UGE from 3.05 to 3.70. It also helps two other editors.
  2. The design is simple and practical. It needs no training and works with different editors and planners. The context ablation gives a useful finding: dropping old intermediate images from the planner's context helps (3.433 vs. 3.697 on ImgEdit-UGE).
Critical weaknesses
  1. The evaluation is not convincing. All scores come from a single VLM judge (Qwen3.8-27B), and there is no human evaluation. The harness itself uses VLM verification and selection at inference time, so part of the gain may come from matching VLM preferences rather than producing better edits. The main table also compares only against a single direct call, while the harness makes many editor calls and picks the final image from several candidates. Removing that final selection drops PICABench from 59.8% to 54.3% and VIBE from 74.3% to 64.2%. A best-of-N baseline with the same number of editor calls and the same selector is therefore needed.
  2. It is unclear that an agent is needed. The router picks one of four preset pipelines once per request, and each skill describes a fixed procedure, so the system looks more like a routed LLM workflow than an agent. The paper does not compare with the router plus fixed scripted pipelines, with a single planner that has all skills and chooses by itself, or with always using Local, which has the best rescue rate and the lowest harm rate in Table 2. "Aggregated Context" removes both routing and the tool loop, so it does not answer this question. Replacing the two-stage router with a single-stage one barely changes the results and is even better on VIBE (75.35% vs. 74.32%), which contradicts the claim that routing contributes the most.
Questions for authors
  1. Can you add a human study, e.g., blind pairwise comparisons against Direct and Nano Banana Pro on a few hundred requests, and report how often the VLM judge agrees with humans?
  2. How does the harness compare with best-of-N direct editing at the same number of editor calls, using the same selector?
  3. How do the router plus fixed scripted pipelines (no planner loop) and a single planner with all skills (no router) perform?
  4. In Section 3, how often would a third plain direct sample "recover" the Bo2 failures? Without this, the 87% recovery may partly be resampling noise.
  5. Which models are used as the verifier and the final selector, and do they overlap with the evaluation judge?
  6. How many editor calls and planner steps does a request use on average? How often does the planner deviate from the skill's default procedure or hit the 10-step limit?
  7. Which editors do EditThinker, RePlan and GenArtist use in Table 3?
  8. Appendices A–C, Table 9 and Algorithm 1 are cited but missing from the PDF. Are they in the supplementary material?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.