acceptodds
Withdrawn from ICLR 2027

Elastic Bridge Model for Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image editing has two competing goals: it should preserve unedited regions with high fidelity while generating realistic and coherent content in edited regions. Existing data-to-data bridge paradigms struggle to satisfy both. Flow matching often under-edits due to overly deterministic trajectories, whereas Brownian bridges introduce stochasticity via global noise, which harms regions that should remain intact. We interpret this stochasticity-determinism tension through a spatial signal-to-noise ratio (SNR) lens. High SNR yields more deterministic trajectories and better fidelity in unedited regions, while lower SNR increases stochasticity and enables substantial edits. As a result, edited and unedited regions require different SNR levels. To resolve this, we propose Elastic Bridge, which replaces the global noise scale with a spatially adaptive elasticity map to balance SNR across regions. To make this design practical, we introduce Source-Grounded Elasticity, a simple approximation that mitigates the resulting train-inference mismatch. Under fair comparison, Elastic Bridge achieves the best accuracy-preservation trade-off among all compared paradigms and scales effectively to state-of-the-art large-scale models.

Primary area:
generative models
Keywords:
image editing, diffusion bridges, flow matching, stochastic generative models

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_vFS1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Elastic Bridge for instruction-based image editing. The authors argue that flow matching under-edits, while the Brownian bridge corrupts unedited regions because it applies the same noise level to every pixel. Based on an SNR analysis, they replace this global noise scale with a spatial map that injects noise mainly where edits occur. Since the map requires the target image, they approximate it with the model’s own prediction at the start of sampling. Experiments show a better accuracy–preservation trade-off than the compared paradigms.

Critical strengths
  1. The SNR analysis explains why a single global noise scale fails, and the fix (a spatial noise map) is simple and easy to plug into existing bridge models.
  2. Under a well-controlled comparison, Elastic Bridge clearly improves preservation over the Brownian bridge at similar accuracy (Table 1, Figure 6).
Critical weaknesses
  1. The justification for Source-Grounded Elasticity is weak. vθ(x0,0,c)v_θ(x_0, 0, c) only estimates the average displacement E[δ∣x0,c]E[δ | x_0, c], the same conditional mean that the paper blames for flow matching's collapse.
  2. How closely K^\hat{K} matches the oracle KK is never quantified.
  3. The ablation in Table 2 does not isolate the effect of SGE, because it changes the training-time map (oracle KK vs. K^\hat{K}) and the inference-time estimator (per-step K^t\hat{K}_t vs. fixed K^\hat{K} from t=0t = 0) at the same time. Moreover, the oracle variant (36.75%) performs even worse than Flow Matching (45%), which makes it a weak reference point.
Questions for authors
  1. For edits with several plausible outcomes, is K^\hat{K} in the edited region noticeably smaller than for unambiguous edits?
  2. Could you quantitatively compare K^\hat{K} with the oracle KK on the test set, e.g., by their correlation or the IoU of high-noise regions?
  3. In Table 2, how do the two missing combinations perform? That is, oracle training with a fixed K^\hat{K} from t=0t = 0, and source-grounded training with per-step K^t\hat{K}_t.
Reviewer_6pN6OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper replaces the global noise level in a Brownian bridge with spatially varying noise levels and predicts a fixed noise map from the source image to accommodate different regional editing needs. The experiments compare different transport formulations and evaluate the method on public benchmarks and through adaptation to pretrained editing models.

Critical strengths

Spatial noise allocation has some practical value. Under the same training conditions, the method improves the trade-off between editing accuracy and background preservation compared with the evaluated Brownian bridge implementation. Some gains on public benchmarks also support the practical utility of this design.

Critical weaknesses
  1. Equation (3) uses the square of the mean drift without accounting for the stochastic component of the bridge drift. This does not justify the conclusion that regions with small displacement become random walks that damage the background. Even when the starting and ending points coincide, an ideal Brownian bridge retains a restoring drift, and its variance vanishes at the endpoint. The paper’s explanation for the failure of the learned baseline still lacks a connection to drift estimation errors.
  2. The implemented SGE uses the absolute value of the conditional mean displacement, whereas the theoretical derivation uses the absolute value of the actual displacement. These quantities are not interchangeable: displacements associated with different valid edits may cancel, leading to nearly zero predicted noise even when the region requires changes and diversity. Consequently, the conclusion about “equalizing the signal-to-noise ratio” does not directly carry over to the implemented method.
Questions for authors

When the mean displacement is zero but valid targets require substantial changes, how does SGE avoid suppressing randomness in that region? What additional assumptions are needed for the conclusion about equalizing the signal-to-noise ratio to hold in this case?

Reviewer_4mU6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the Elastic Bridge, which replaces the global diffusion scale of a standard Brownian Bridge with a spatial elasticity map which is derived from the source-target displacement, to achieve a more balanced signal-to-noise ratio (SNR) across different regions. Furthermore, the proposed source-grounded elasticity utilizes vθ(x0,0,c)v_\theta(x_0, 0, c) to estimate the elasticity map during both training and inference, which avoids train-inference mismatch while enabling more precise localization of the editing-related regions. Experiments and visualizations demonstrate the effectiveness of the proposed model.

Critical strengths
  1. The motivation behind the method is clear. The approach of using spatially adaptive stochasticity to improve the balance between editing and preservation is both concise and somewhat novel.
  2. The experiment results are generally strong, and the effectiveness of the method is validated across multiple benchmarks as well as through its adaptation to Qwen-Image-Edit model.
Critical weaknesses
  1. The core argument of diversity lacks direct validation. The paper attributes the primary advantage of the stochastic bridge to its ability to restore one-to-many editing diversity, but the evaluations merely focus on accuracy, quality, and preservation without directly measuring the diversity among the model outputs. The current experimental results provide support for improvements in editing-preservation trade-off, but they are insufficient to substantiate the preservation-diversity mechanism emphasized in the paper. This viewpoint is now just an intuition of the modeling I think.
  2. The learning mechanism for source-grounded elasticity seems to be unclear. During training, K^\hat{K} is derived from the network's prediction at t=0t=0. However, this branch is stop-gradient. Since the flow-matching training uses t∼U(0,1)t \sim U(0,1), there is no direct supervision at the precise point of t=0t=0. Although the authors derive the approximation v(x0,0,c)≈x1−x0v(x_0, 0, c) \approx x_1 - x_0 based on an idealized vector field, it's still not so clear why this estimation remains accurate and stable after the actual training process.
Questions for authors
  1. Is it possible to directly evaluate whether Elastic Bridge enhances the output diversity in editing-related regions? Comparing it against a conditional DiT might be a nice try.
  2. Under the stop-gradient setting, how is vθ(x0,0,c)v_\theta(x_0, 0, c) able to learn an effective displacement estimate?

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

Related papers

Discussion (0)

Sign in to comment.