Source-linked AI summary
EditJev: System One Routing for Image Editing Models
Zhile Guan, Xu Jiang, Yule Duan, Gongbo Zhang, Shuyu Wang, Xiao Chen, Jian Zhang
TL;DR
Image-editing models differ in suitability and latency across requests, while selection must happen before candidate outputs are available. EditJev directly predicts preferences over candidate editors and combines them with model speed scores. It improves aggregate quality over the strongest fixed editor on ImgEdit and GEdit-En while reducing estimated latency.
Problem
Candidate editing quality is unknown before generation, so a router must select using only the source image and instruction despite request-specific quality and latency trade-offs.
Source: §1 · p. 2, §3.1 · p. 3Method
EditJev predicts fixed-dimensional candidate preference scores in one forward pass, trains from offline pairwise preferences, and combines scores with model speed at selection time.
Source: §1 · p. 2, §3.2 · p. 4, §3.3 · p. 5Results
60.8% and 31.7% estimated end-to-end latency reductions accompany aggregate-quality gains over the strongest fixed editor on ImgEdit and GEdit-En, respectively.
Source: §1 · p. 2Takeaways & Limitations
EditJev demonstrates sample-level image-editing routing that selects suitable editors for image–instruction pairs while balancing quality and latency.
Source: §5 · p. 9The authors identify more reliable, scalable routing supervision and support for dynamically evolving editor pools as future work.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Image-editing models vary in quality and latency across image–instruction pairs, making fixed-model deployment or manual selection a poor fit. EditJev formulates sample-level routing and predicts preferences over candidate editors to balance quality and latency.
1 INTRODUCTION
- Model strengths and computational costs vary by request, while a fixed default can waste computation and manual per-request selection is impractical.The model best suited to a global transformation may differ from one suited to precise local manipulation.Source: §1 · p. 2
- The paper formulates sample-level image-editing routing: choose a candidate editor from the input image and instruction while accounting for quality and inference cost.The candidate’s eventual output quality is unknown at selection time.Source: §1 · p. 2
- EditJev uses a vision-language backbone and lightweight preference head to predict a fixed-dimensional score vector in one forward pass, requiring 94 ms for routing.The framework directly predicts preferences rather than generating model names autoregressively.Source: §1 · p. 2
- Offline candidate outputs are evaluated by a multimodal judge, converted into pairwise preferences, and used to train a ranking objective; inference combines preferences with model speed scores.The resulting selection balances editing quality and generation latency.Source: §1 · p. 2
- EditJev improves aggregate quality over the strongest fixed editor while reducing estimated end-to-end latency by 60.8% on ImgEdit and 31.7% on GEdit-En, respectively.The latency estimates include routing overhead.Source: §1 · p. 2
2 RELATED WORK
Prior image-editing research has expanded task datasets, editing frameworks, and model capabilities, while scaling can increase latency and deployment cost. Routing is established in language and text-to-image generation but remains underexplored for instruction-based image editing; EditJev applies direct structured preference prediction to this setting.
2.1 INSTRUCTION-BASED IMAGE EDITING
- Prior instruction-based editing work covers diverse operations and has scaled model capacity, data, and compute, while such scaling increases inference latency and deployment cost.The operations include object insertion and removal, attribute modification, and style transformation.Source: §2.1 · p. 3
2.2 MODEL ROUTING
- Routing has been studied for language models and extended to text-to-image generation, but remains underexplored for instruction-based image editing; EditJev uses compact direct preferences for fixed-candidate selection.Structured predictions can be consumed downstream without additional parsing.Source: §2.2 · p. 3
3 METHOD
The method formulates image editing as routing among a fixed candidate set, predicting preferences from each request and combining them with model speed. It trains the router on reliable pairwise comparisons derived from offline candidate outputs.
3.1 PROBLEM FORMULATION
- Routing selects one model from a fixed candidate list for each image–instruction request, targeting high editing quality and low latency despite unknown pre-execution quality.The policy must choose using only the source image and instruction, without observing candidate outputs.Source: §3.1 · p. 3
3.2 SYSTEM ONE ROUTING
- EditJev predicts a fixed-dimensional preference score for each candidate and combines those scores with normalized speed scores to select a model for execution.The score vector can be consumed directly without autoregressive decoding or response parsing; only the selected model is executed.Source: §3.2 · p. 4
3.3 PREFERENCE-BASED ROUTER TRAINING
- Training evaluates every candidate output offline with an MLLM on editing success, preservation, naturalness, and artifacts, then aggregates semantic-consistency and perceptual-quality scores by their geometric mean.Each component score is the minimum of its paired evaluator dimensions.Source: §3.3 · p. 5
- The router learns from preferred model pairs whose aggregate-quality gap meets threshold δ, using a pairwise ranking objective that avoids reducing supervision to one best-model label.The threshold excludes near-tied pairs whose ordering may be unreliable; pairwise comparisons can also avoid matching evaluator-score scales.Source: §3.3 · p. 5
4 EXPERIMENTS
EditJev improves quality while reducing end-to-end latency against fixed-editor and general-purpose routing baselines. Ablations and user judgments examine its ranking objective, visual input, backbone scale, speed weighting, and alignment with human preferences.
4.1 EXPERIMENTAL SETUP
- The ImgEdit comparison covers fixed models, MLLM-based routing, and System One routing; reported latency in Table 1 is routing overhead before editor execution.The benchmark setup uses GPT-5.1 as the primary evaluator and follows each benchmark’s official protocol.Source: §4.1 · pp. 5–6, Table 1 · p. 6
4.2 MAIN RESULTS
- EditJev achieves 2.55× and 1.46× inference speedups on ImgEdit and GEdit-En, respectively, while improving aggregate scores on both benchmarks.It also achieves a better quality–latency trade-off than fixed-model deployment and general-purpose MLLM routing.Source: Table 1 · p. 6, Table 2 · p. 6, §4.2 · pp. 7–8
- For the floral-background edit, EditJev selects FLUX.2 Klein 9B for garment preservation; for garment extraction, it selects Qwen-Image-Edit, the only successful candidate.For car removal, it selects the fastest successful candidate, FLUX.2 Klein 9B, illustrating request-specific choices that account for quality and generation efficiency.Source: Figure 3 · p. 7, §4.2 · p. 8
4.3 ABLATION STUDY
- On ImgEdit, pairwise ranking scores 4.6002, versus 4.5545 for MSE regression and 4.5491 for MCQ generation.The authors attribute the advantage to learning relative ordering from multiple pairwise comparisons rather than absolute evaluator scores or a single best-model label.Source: §4.3 · p. 8, Table 4 · p. 8
- Removing the source image lowers ImgEdit’s overall score from 4.6002 to 4.5717, supporting the use of both image and instruction for routing.The authors note that candidate suitability can vary with the source image even when the instruction is unchanged.Source: §4.3 · p. 8, Table 5 · p. 8
- Across Qwen3-VL backbone scales, routing performance is comparable, and increasing backbone size does not consistently improve it.The authors interpret this as stable performance across different multimodal backbone scales.Source: §4.3 · p. 9, Table 6 · p. 9
- On ImgEdit, increasing β from 0.4 to 1.0 lowers estimated end-to-end latency from 26.79 to 17.73 seconds (33.8%) as the score changes from 4.5966 to 4.5667.With the trained router fixed, β = 0.7 gives a score of 4.5916 at 22.33 seconds, enabling inference-time trade-off adjustment without retraining.Source: §4.3 · p. 9, Figure 4 · p. 9
4.4 USER STUDY
- In 400 human judgments, EditJev’s selected model matched participants’ choice 327 times (81.75%), compared with 138 times (34.50%) for GPT-5.5 routing.Mean per-case Spearman correlation between EditJev’s speed-adjusted scores and candidate votes was ρ = 0.7478.Source: §4.4 · p. 9
5 CONCLUSION
EditJev formulates sample-level image-editing model routing to select suitable editors for different image–instruction pairs. The paper identifies more reliable, scalable supervision and support for evolving model pools as future directions.
5 CONCLUSION
- Table 7 reports human preference alignment on 20 cases from ImgEdit and GEdit-En.The supplied caption gives the table’s scope but no outcome values.Source: Table 7 · p. 9
- EditJev learns preferences over candidate editors and selects suitable models for individual image–instruction pairs.The paper presents this as sample-level image-editing model routing.Source: §5 · p. 9
- Future work targets more reliable, scalable routing supervision and routers that support larger, evolving editor pools without extensive retraining.The paper proposes stronger multimodal judges for more comprehensive and fine-grained editing assessments.Source: §5 · p. 9