Instruction-based image editing models exhibit pronounced sample-specific performance variation: different models can produce substantially different results for the same source image and editing instruction, and the best-performing model often varies across requests. These models also differ substantially in inference latency, making fixed-model deployment inefficient. We therefore study image editing model routing, which dynamically selects an editor from a candidate pool given an input image and editing instruction. We introduce EditJev, a System One routing model that directly predicts sample-level preference scores over candidate image editing models in a single forward pass. A lightweight preference head on a vision-language backbone produces a fixed-dimensional score vector without autoregressive model-name generation or response parsing. To train EditJev, we construct pairwise preferences from offline candidate outputs and optimize a pairwise ranking objective to learn relative model suitability. At inference time, EditJev combines the predicted preference scores with model-level speed scores to balance editing quality and generation latency. Experiments on ImgEdit and GEdit-En show that EditJev improves aggregate quality over the strongest fixed editor, from 4.540 to 4.592 and from 8.194 to 8.255, respectively, while reducing estimated end-to-end latency by 60.8% and 31.7%. These results demonstrate the effectiveness of direct, structured model routing for efficient image editing.
Figures & tables
Figure 1: Left: Different image editing models exhibit complementary strengths: the optimal model varies across image-instruction pairs, and no single model consistently performs best for all editing requests. Middle: This observation motivates model routing, where EditJev achieves a better quality–latency trade-off by dynamically selecting suitable models instead of relying on a fixed model. Right: Given an input image and instruction, EditJev uses a multimodal router to understand the request, estimate model suitability, and select the model that provides better editing quality with lower inference latency.
Figure 2: Overview of EditJev. The upper part shows offline candidate generation, automatic evaluation, and router training. The lower part shows quality–latency-aware inference, which combines preference scores with model-level speed scores.
Method
Add
Adjust
Extract
Replace
Remove
Background
Style
Compose
Action
Overall ↑
Latency(s) ↓
Fixed Models
FLUX.2-Klein-4B
4.470
4.660
1.690
4.480
4.300
4.470
4.850
3.510
4.800
4.073
–
FLUX.2-Klein-9B
4.580
4.730
1.950
4.680
4.710
4.680
4.910
3.590
4.880
4.252
–
FLUX.1-Kontext-dev
4.050
3.930
1.950
4.440
3.420
3.940
4.400
2.460
4.240
3.670
–
OmniGen2
3.730
3.640
1.640
4.040
3.110
3.680
4.580
2.670
4.420
3.461
–
Qwen-Image-Edit
4.660
4.650
4.120
4.750
4.630
4.410
4.770
3.770
4.750
4.540
–
Table 1: Quantitative comparison on ImgEdit. Methods are grouped into fixed models, MLLM-based routing methods, and System One routing methods. The best and second-best results for each editing category, overall quality, and latency are marked in bold and underline, respectively. Latency denotes the additional routing overhead before executing the selected editor.
Figure 3: Qualitative routing results on GEdit-En and ImgEdit. A red border marks the model selected by EditJev for each input.
Method
G_SC
G_PQ
G_O ↑
Fixed Models
FLUX.2-Klein-4B
7.731
8.397
7.667
FLUX.2-Klein-9B
8.081
8.416
7.965
FLUX.1 Kontext
6.404
7.579
6.241
OmniGen2
6.391
7.762
6.297
Qwen-Image-Edit
8.441
8.326
8.194
Table 2: Quantitative evaluation results on GEdit-En.
Method
ImgEdit (s) ↓
GEdit-En (s) ↓
Qwen-Image-Edit
56.90
56.90
GPT-5.5 Routing
51.15
53.55
EditJev
22.33
38.88
Table 3: Latency comparison among Qwen-Image-Edit, GPT-5.5 Routing, and EditJev. Latency includes both the routing inference time and the generation time of the selected editor.
Training Objective
Overall ↑
MSE Regression
4.5545
MCQ Generation
4.5491
Pairwise Ranking (Ours)
4.6002
Table 4: Ablation on the router training objective on the ImgEdit benchmark.
Variant
Overall ↑
w/o Image Input
4.5717
w/ Image Input (Ours)
4.6002
Table 5: Ablation on the visual input to the router on ImgEdit.
Router Backbone
Overall ↑
Qwen3-VL-2B
4.5898
Qwen3-VL-4B (Ours)
4.6002
Qwen3-VL-32B
4.5898
Table 6: Ablation on the router backbone scale on ImgEdit.
Figure 4: Quality–latency trade-off on ImgEdit under different speed weights.
Method
Agreement (%) ↑
Spearman ρ↑
GPT-5.5 Routing
34.50
—
EditJev (Ours)
81.75
0.7478
Table 7: Human preference alignment on 20 cases from ImgEdit and GEdit-En.
Aligning generative models with human preferences remains a central challenge. While RLHF is widely adopted, it relies on large-scale preference data. This limitation is particularly evident in evaluating instruction-guided image edits, even though humans can often learn such preferences from only a few examples. We present RewardHarness, a training-free, self-evolving agentic reward framework that reframes reward modeling as context evolution rather than weight optimization. Instead of learning from large-scale annotations, RewardHarness aligns with human preferences by iteratively evolving a library of tools and skills from as few as 100 preference demonstrations. Given a source image, an edited image, and an editing instruction, an orchestrator agent selects the most relevant subset of tools and skills from the maintained library, and sub-agents invoke them to construct a reasoning chain that produces a reward score. By comparing predicted scores with ground-truth preferences and analyzing successes and failures in the reasoning process, the orchestrator automatically refines its library of tools and skills without additional human annotation. Using only 0.05% of the EditReward preference data, RewardHarness achieves 47.4% average accuracy on EditReward-Bench K=2/3/4 and GenAI-Bench, surpassing GPT-5 by 5.3 points. When used as a reward signal in FlowGRPO, RL-tuned models achieve 3.52 on ImgEdit-Bench. Code and model are available at the project page: https://rewardharness.com/
Yuxuan Zhang, Cong Wei, Penghui Du +11
University of British Columbia · Vector Institute · Kolors Team, Kuaishou Technology +5
Image Chain-of-Thought (Image-CoT) is a test-time scaling paradigm that improves image generation by extending inference time. Most Image-CoT methods focus on text-to-image (T2I) generation. Unlike T2I generation, image editing is goal-directed: the solution space is constrained by the source image and instruction. This mismatch causes three challenges when applying Image-CoT to editing: inefficient resource allocation with fixed sampling budgets, unreliable early-stage verification using general MLLM scores, and redundant edited results from large-scale sampling. To address this, we propose ADaptive Edit-CoT (ADE-CoT), an on-demand test-time scaling framework to enhance editing efficiency and performance. It incorporates three key strategies: (1) a difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty; (2) edit-specific verification in early pruning that uses region localization and caption consistency to select promising candidates; and (3) depth-first opportunistic stopping, guided by an instance-specific verifier, that terminates when intent-aligned results are found. Extensive experiments on three SOTA editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show that ADE-CoT achieves superior performance-efficiency trade-offs. With comparable sampling budgets, ADE-CoT obtains better performance with more than 2x speedup over Best-of-N.
Xiangyan Qu, Zhenlong Yuan, Jing Tang +9
Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences · AMAP, Alibaba Group +1
Common image editing tasks typically adopt powerful generative diffusion models as the leading paradigm for real-world content editing. Meanwhile, although reinforcement learning (RL) methods such as Diffusion-DPO and Flow-GRPO have further improved generation quality, efficiently applying Reinforcement Learning from Human Feedback (RLHF) to diffusion-based editing remains largely unexplored, due to a lack of scalable human-preference datasets and frameworks tailored to diverse editing needs. To fill this gap, we propose HP-Edit, a post-training framework for Human Preference-aligned Editing, and introduce RealPref-50K, a real-world dataset across eight common tasks and balancing common object editing. Specifically, HP-Edit leverages a small amount of human-preference scoring data and a pretrained visual large language model (VLM) to develop HP-Scorer--an automatic, human preference-aligned evaluator. We then use HP-Scorer both to efficiently build a scalable preference dataset and to serve as the reward function for post-training the editing model. We also introduce RealPref-Bench, a benchmark for evaluating real-world editing performance. Extensive experiments demonstrate that our approach significantly enhances models such as Qwen-Image-Edit-2509, aligning their outputs more closely with human preference.
Fan Li, Chonghuinan Wang, Lina Lei +9
Huawei Noah’s Ark Lab · Nankai University · Harbin Institute of Technology