DuplexBridge: Coordinating Understanding and Generation for Unified Image Editing

1 Introduction

• A task-level score cannot distinguish an incorrect task expression from faithful execution of that incorrect condition or from faulty rendering of a correct condition.The clock examples contrast a wrong analog target of 11:50 with a correct 4:05 target whose protected digital display is changed.Source: §1 · p. 1
• DuplexBridge combines distinct local supervision references for both adapting roles with reciprocal checks that select their additional global-feedback losses.Understanding is checked against the task without the generated image; generation is checked against the condition it receives, while the global reward assesses the final edit.Source: §1 · p. 2
• The checkable bridge specifies source-bound targets, requested changes, and preservation constraints for the unified model’s native understanding-to-generation path.The bridge also defines the condition used for generation-local execution checks.Source: §1 · pp. 1–2
• 21.75–25.75 points: Full-system CCR gains over direct editors span three model configurations, with additional gains on public benchmarks.On a common supervised Wan-Image initialization, full reaches 77.08% CCR versus 70.92% without gates and 75.00% with original-task G-local supervision and both gates retained.Source: §1 · p. 2
• CE-Bench evaluates complete editing fulfillment by requiring all task-derived change and preservation checks to pass, independently of the predicted bridge.The benchmark contains 1,200 tasks.Source: §1 · p. 2

Understanding–generation coordination

• MGIE and SmartEdit establish understanding-guided editing and coupled adaptation, while SmartEdit’s bidirectionality concerns feature interaction rather than training feedback between roles.MGIE learns expressive instructions and jointly trains its multimodal language model and image editor (Fu et al., 2024); SmartEdit uses bidirectional interaction between image features and MLLM outputs (Huang et al., 2024).Source: §2 · p. 3
• DuplexBridge distinguishes task fidelity, received-condition execution, and task outcome as separate training signals during co-adaptation.The cited approaches include unified understanding and generation, successive understanding and generation phases, and decoupled role updates (Deng et al., 2025; Zheng et al., 2026; Jiao et al., 2026; Li et al., 2026).Source: §2 · p. 3

Structured prompt enhancement and feedback

• Unlike prior approaches described here, DuplexBridge combines U-local task-fidelity checks without rendered images and G-local execution checks against the received condition, with reciprocal trust selecting additional global losses.DDA-Thinker assesses plans and visual outcomes while keeping the editor frozen; DARS jointly trains planner and renderer using task checklists and post-render diagnostics (Yang et al., 2026).Source: §2 · p. 3

Reward modeling and editing evaluation

• CE-Bench emphasizes complete compliance with source-grounded changes and preservation, using original-task requirements independent of the predicted bridge and evaluating final images.The passage contrasts this emphasis with complementary editing and reasoning benchmarks and describes the distilled reward model’s shared outcome signal.Source: §2 · p. 3

3.1 Native roles and the forward path

• The unified model’s understanding role produces a public structured bridge, which its generation role receives with the source image in a fresh context.Generation does not receive the original instruction or the understanding role’s internal reasoning; no external understanding module constructs the bridge.Source: §3.1 · pp. 3–4

3.2 Making the editing requirements explicit

• The bridge organizes the goal, editable scope, operations, preservation requirements, and forbidden effects, with its goal summary entailed by the operative requirements.The same public condition defines the execution checks and is used by the compiler.Source: §3.2 · p. 4

3.3 Source binding, coverage, and validity

• The bridge binds changes to source entities, specifies terminal states, and separates editable content from protected content.Execution groups combine source grounding, applicable initial-state checks, and output requirements; repeated preservation obligations are deduplicated or partitioned, and public evaluation uses original-task requirements rather than the predicted bridge.Source: §3.3 · pp. 4–5
• In the clock example, the analog display’s requested terminal state is 4:05, while the digital display, clock appearance, and surrounding scene remain preservation obligations.The digital clock supplies the time used to calculate the analog target but is not an edit target.Source: §3.3 · p. 4

4 Coordinating Learning with Local and Global Feedback

• Prompt-fidelity, execution, and global-outcome checks use different information: the task and bridge, the bridge and generated image, and the task and generated image, respectively.Local rewards train each role’s responsibility; partner trust determines when the joint outcome also supervises that role.Source: §4 · p. 5

4.1 Role-local rewards

• Understanding’s local reference checks format, requirement-level semantic matching, and clarity or grounding without reading the generated image.The plan score uses maximum one-to-one matching on target identity, operation, terminal value, and invariant scope; unresolved judgments abstain.Source: §4.1 · p. 5
• Generation-local checks assess change groups and distinct invariants against the condition received, while quality scoring excludes instruction compliance.Invalid or unresolvable specifications and uncertain execution judgments are excluded; low scores on valid, confidently judged checks remain learning signals.Source: §4.1 · pp. 5–6

4.2 Global task feedback

• A frozen pairwise reward model compares the generated edit with baseline edits under the original instruction and rubric, without seeing the bridge or private reasoning.Its direct-scoring model distills a thinking teacher’s judgment distribution and remains frozen during U/G training.Source: §4.2 · p. 6

4.3 Bidirectional gates

• Reciprocal trust gates U-global by execution trust and G-global by prompt trust, while valid local supervision remains independent of those gates.The gates require usable judgments; the routing examples show that a role can retain local supervision when global feedback is excluded.Source: Figure 3 · p. 6, §4.3 · p. 7

4.4 Alternating updates with four loss components

• Training alternates frozen-role updates and computes each loss component on its eligible samples, allowing a sample excluded from global feedback to retain local supervision.Components with fewer than two eligible samples or zero reward variance are skipped, including their KL penalty.Source: §4.4 · p. 7

4.5 Reasoning-distilled global judgment

• The direct student learns the frozen teacher’s marginal judgment distribution and supplies global feedback after distillation, while the reward model stays frozen during U/G training.The passage describes independent preference labels, candidate-order swaps, and difficult-pair evaluation as reliability assessments.Source: §4.5 · pp. 7–8

5.1 Models and baselines

• BAGEL, Qwen-2511, and Wan-Image-7B use their own understanding and generation capabilities; direct baselines edit the original instruction, while full systems add native PE, structured SFT, and joint RL.Wan-Image is used for mechanism ablations, with variants trained from the common B0 supervised initialization.Source: §5.1 · p. 8

5.2 CE-Bench and complete task fulfillment

• CE-Bench evaluates 1,200 tasks for joint satisfaction of target binding, terminal-state changes, and preservation constraints, independently of the model’s predicted bridge.CCR requires every task-derived change and preservation group to pass; a failed output remains in the denominator.Source: §5.2 · p. 8

5.3 Taxonomy and task allocation

• CE-Bench assigns each task one mutually exclusive primary category and a secondary capability category, with precedence E>D>C>B>A.Categories distinguish reference-image use, coupled obligations, inferred targets or states, exact text or structured content, and direct source grounding.Source: §5.3 · p. 9, Table 1 · p. 9

5.4 Public benchmark protocols

• Tasks use a source image and original instruction, with additional task-provided references only for category E; public comparisons use each benchmark’s native metric scale.The public protocols cover KRIS-Bench, RISEBench-360, ImgEdit, and GEdit-Bench English Full-set GO.Source: §5.3 · p. 9, §5.4 · p. 9

6.1 End-to-end editing performance

• On CE-Bench, DuplexBridge improves CCR by 25.75, 21.75, and 23.25 points on BAGEL, Qwen-2511, and Wan-Image, respectively, alongside preservation (table 2).Table 2 reports percentage scores; full systems include native PE, structured SFT, and gated RL.Source: §6.1 · p. 10, Table 2 · p. 10
• Wan-Image reaches 80.13/29.17 KRIS/RISE, gaining 19.29/16.11 points over its direct baseline; Qwen-2511 reaches 82.16/31.11 versus DARS’s 80.72/27.50.The public benchmark table also includes ImgEdit and GEdit-Bench metrics.Source: §6.1 · p. 10, Table 3 · p. 10
• CE-Bench examples illustrate image-grounded answers, spatial placement, geometric reasoning, and multi-reference composition while preserving source content.The figure’s final row combines entity placement with preservation of the base scene.Source: §6.1 · p. 11, Figure 4 · p. 11

6.2 Supervised initialization controls

• From B0 to full DuplexBridge, CE CCR increases by 17.91, 16.92, and 17.66 percentage points for BAGEL, Qwen-2511, and Wan-Image-7B.B0 is the supervised checkpoint with native PE and structured SFT before RL; all three configurations improve four public metrics and each RISE category over B0.Source: §6.2 · p. 11, Table 5 · p. 12

6.3 Information-matched communication and verification

• Grouping raises CCR from 56.67 to 59.00 for key–value serialization and from 57.42 to 59.42 for JSON; the smaller syntax gap suggests organization matters more than JSON syntax.Removing an information-redundant goal or shuffling fields changes CCR little.Source: §6.3 · p. 11, Table 6 · p. 12
• Compiling execution checks improves source fidelity and reduces duplication; the construction comparison also reports editing CCR after retraining.Fidelity, task recall, and duplication diagnostics cover all construction attempts, while human agreement uses an adjudicated content-matched subset.Source: §6.3 · p. 12, Table 7 · p. 12

6.4 Joint adaptation and supervision references

• Joint ungated training reaches 70.92 CCR, while adding selected global feedback to dual-role local-only training raises CCR from 68.25 to 77.08.All RL variants start from B0; the full score is 4.36 on ImgEdit and 7.74 on GEdit EN, compared with 4.37 and 7.71 for the DARS-style transplant.Source: §6.4 · pp. 12–13, Table 8 · p. 13
• With both gates retained, G-local supervision on the received condition reaches 77.08 CCR, versus 75.00 when it checks the original task instead.U-local and global feedback retain the original task reference.Source: §6.4 · p. 13, Table 9 · p. 13

6.5 Selecting reliable global feedback

• 77.08 CCR with both gates exceeds 74.17 or 73.33 with either gate alone and 70.92 with neither; semantic gating beats group-matched random selection by 4.83 points.Soft semantic weighting reaches 76.33 CCR, while annotation-assisted oracle gates reach 78.67 CCR using additional adjudicated annotations.Source: Table 10 · p. 13, §6.5 · pp. 13–14

6.6 Checkpoint cross-pairing

• The positive interaction in checkpoint cross-pairing indicates complementary adaptation of the two roles, and KRIS and RISE also improve under the learned pair.The comparison pairs shared B0 states with adapted checkpoints from the same full training run.Source: §6.6 · p. 14, Table 13 · p. 14

6.7 Requirement load and joint fulfillment

• Full has higher CCR point estimates at every observed requirement load, with an overall paired gain of 6.17 points [4.50, 7.83].Across tasks, change all-pass rises from 80.42% to 83.50%, preservation all-pass from 76.75% to 81.25%, and mean failed groups fall from 0.659 to 0.508.Source: Table 14 · p. 15, §6.7 · p. 15, Figure 5 · p. 15

6.8 Reward quality, cost, and downstream effects

• Soft KD raises difficult-pair accuracy from 66.5% to 74.8%, approaching explicit thinking’s 75.8% at 1.0× rather than 6.0× online latency.M5 raises CCR by 2.83 points with gates; gating adds 6.16 points under M5, versus 5.17 under base direct reward.Source: Table 15 · p. 16, §6.8 · pp. 16–17, Table 16 · p. 16
• M5 improves MMRB2 editing pair accuracy from 66.2 to 70.6 and EditReward-Bench K = 2, 3, 4 ranking scores from 57.2/35.9/10.8 to 61.0/43.5/15.8.M1 achieves the highest scores on these benchmarks, while M5 provides direct scoring at lower online cost.Source: §6.8 · p. 17, Table 17 · p. 17

6.9 Human evaluation of final edits

• 75.77% human CCR for full versus 69.07% for ungated on 194 commonly resolved tasks, a 6.70-point gain [1.03, 12.37].The automatic gain is 6.19 points, and automatic minus human gain is −0.52 points [−3.61, 2.58].Source: §6.9 · p. 17, Table 19 · p. 17, Table 20 · p. 18
• Instruction preference favors full over both ungated and direct, while preservation preference favors full over direct.The visual-quality intervals in both comparisons and the preservation interval against ungated span zero.Source: §6.9 · p. 18, Table 21 · p. 18, Figure 6 · p. 18

6.10 Text-to-image adaptation

• All three DuplexBridge text-to-image configurations improve on all five reported metrics after retraining both roles.U specifies requirements from text, and G synthesizes from that condition.Source: §6.10 · p. 18, Table 22 · p. 19

CE-Bench

• CE-Bench examples include sign-text replacement and translation, localized color changes, a traffic-light state change, and object appearance edits.Paired images show edited regions alongside the surrounding scene; auxiliary references support flower replacement and dog insertion.Source: Figure 9 · p. 20

RISEBench

• On RISEBench, examples include a connected maze route, physical state changes, and an equivalence-consistent cylinder replacement.The maze route passes through the white corridor between red and green endpoints; the rope is untied and egg whites become opaque.Source: Figure 7 · p. 19, §7 · p. 21

KRIS-Bench

• KRIS-Bench examples pair requested changes with physical consequences, reference fidelity, and preservation.Examples include a balance responding to added load, oil separating above water, cups equalizing in size, and garment replacement preserving white trousers.Source: Figure 8 · p. 19, §7 · p. 21

8 Limitations

• DuplexBridge has not yet been evaluated on more complex video generation and editing tasks.The authors identify extension to video as a future research direction.Source: §8 · p. 21

• DuplexBridge connects a unified model’s own understanding and generation through a checkable editing condition.Source: §9 · p. 21
• Local supervision distinguishes task fidelity from execution of the condition generation receives.Source: §9 · p. 21
• Reciprocal trust gates select additional global feedback while retaining valid local learning.Source: §9 · p. 21
• Across three configurations, complete compliance improves over direct editors, and controlled Wan-Image comparisons and human evaluation support the findings.Source: §9 · p. 21
• The findings support separating local responsibilities when jointly adapting understanding and generation.Source: §9 · p. 21