acceptodds
Withdrawn from ICLR 2027

Conditional Geometric Feedback Control for Training-Free Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-guided image editing requires realizing target changes while preserving non-target content in the source image. Existing work uses latent-space geometry to characterize generative representations and editing directions, but how to translate this geometry into preservation feedback adapted to current editing conditions remains less explored. We propose Conditional Geometric Feedback Control (CGFC), a training-free and inversion-free image editing method. We construct local Fisher geometry from the source–target condition-label posterior to characterize the sensitivity of condition discrimination to latent corrections. Under Gaussian linear interpolation and population-optimal conditional velocities, its nonzero sensitive axis aligns with the same-state conditional velocity difference, motivating the model-predicted difference as a practical proxy for the control axis. Using this shared axis, one CGFC update controls both the direction of the current prediction residual at fixed norm within dynamic spatial support and the accumulated source displacement orthogonal to the axis. On PIE-Bench with SD3, CGFC improves content preservation and round-trip consistency. These results support using condition-dependent latent geometry to design feedback control for stepwise image editing.

Primary area:
generative models
Keywords:
image editing, flow matching, information geometry

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JsBgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper constructs a local Fisher geometry from the posteriors under the source and target conditions and relates its sensitive directions to the conditional velocity difference. Based on this, the method adjusts the fidelity residual and applies orthogonal, bounded feedback to the accumulated source-image displacement, in order to control content drift in inversion-free editing.

Critical strengths
  1. Linking the conditional velocity difference to the posterior discriminative geometry provides an interpretable basis for choosing the feedback direction.
  2. The control mechanism has a structural design: Adjusting the residual along the discriminative direction and applying bounded displacement feedback in the orthogonal direction separates local conditional change from accumulated source-image drift.
Critical weaknesses
  1. The geometric formulation adds limited incremental over direct directional control. The Fisher geometry amounts to projecting and adjusting along the conditional velocity difference. The paper does not sufficiently explain whether the geometric derivation cannot be obtained by using this direction directly.
  2. The control components are largely a combination of existing pieces: Projection, norm restoration, and bounded displacement feedback are each common operations.
  3. Local discriminative properties do not directly equal to semantic preservation. Orthogonality to the conditional discriminative axis corresponds only to first-order posterior invariance under idealized conditions and does not imply preservation of identity, structure, or background.
Questions for authors

Please refer to Weaknesses.

Reviewer_SS72OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CGFC, a training-free and inversion-free control method for image editing. Its core idea is to extract a condition-sensitive axis from the local Fisher geometry of the source–target conditional-label posterior and use the conditional velocity difference at the same state as a computable proxy, then jointly adjust the current residual direction and the accumulated source displacement. The theoretical presentation is relatively complete, and the results show improvements in source content preservation and round-trip consistency. However, the main results are accompanied by a decrease in target CLIP, and the explanation that “weaker editing makes the result more similar to the source image” has not been ruled out. The discrepancy between the theoretical axis and the proxy from the actual model also lacks direct validation.

Critical strengths
  1. The connections among the method's motivation, mathematical objects, and control operations are relatively clear.
  2. The method requires no additional training or image inversion and introduces no additional denoiser calls. Its deployment cost is relatively low compared with methods requiring training or per-sample optimization.
  3. The experiments do show relatively consistent improvements in source content preservation.
Critical weaknesses
  1. The most critical issue is that the trade-off between editing capability and fidelity has not been fairly controlled. Compared with Prism, target CLIP for the full image and the edited region decreases from 24.569/21.183 to 24.335/21.019, respectively. The increase in similarity to the source image may partly result from more conservative, smaller-magnitude edits rather than better fidelity control.
  2. The paper lacks comparisons at matched editing strength and a convincing Pareto analysis. Although Appendix E.1 / Figure 6 provides CLIP–PSNR curves for a small number of CFG settings, the authors themselves acknowledge that this small-scale sweep “cannot establish Pareto superiority.” To demonstrate that CGFC offers a superior fidelity–editing trade-off, source-image fidelity metrics should be compared over the full benchmark at matched target CLIP, matched region-level editing success rates, or matched perceptual editing magnitudes.
  3. The paper does not adequately present failure cases or behavior when the geometric proxy fails. In particular, it should analyze whether the control axis degenerates and whether necessary edits are suppressed when the conditional velocity difference is very small, source and target semantics are highly similar, masks are inaccurate, or edits involve large structural changes.
Questions for authors
  1. The alignment between the Fisher sensitivity axis and the conditional velocity difference relies on strong assumptions, including Gaussian linear interpolation and population-optimal conditional velocity. Prediction errors in the actual SD3 model, multimodal data distributions, and high-dimensional nonlinearities may disrupt this correspondence. However, the paper does not measure the angle between the theoretical and proxy axes, their stability, or their discrepancy in the actual model.
  2. The ablation results indicate that accumulated displacement feedback accounts for a large share of the gains, whereas the additional gains from the most geometrically distinctive residual-direction adjustment are relatively limited. The paper needs stronger evidence that the Fisher geometry component itself is the primary source of novelty, rather than generic source displacement regularization.
  3. The distinction from the closest methods is insufficient. Proposition 1 is essentially a reformulation of the CFG implicit-classifier identity, yet the paper does not discuss this connection. FlowDC likewise protects source structure using accumulated displacement and orthogonal decomposition, but it is mentioned only in related work and is not included in experimental comparisons.
Reviewer_PMezOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CGFC, a training-free and inversion-free image editing method that aims to improve source content preservation during text-guided editing. The method derives a condition-sensitive direction from the Fisher information geometry of the source-target condition posterior and approximates it using the difference between conditional velocity predictions. Based on this direction, CGFC introduces two complementary feedback mechanisms: norm-preserving residual direction adjustment to reduce interference with intended edits, and bounded displacement feedback to suppress accumulated content drift. Experiments on PIE-Bench demonstrate improved content preservation and round-trip reconstruction compared to existing methods, although these improvements come with a slight reduction in target-text alignment.

Critical strengths

S1. The proposed method provides a theoretically motivated framework for jointly controlling instantaneous preservation errors and accumulated content drift using conditional Fisher geometry, without requiring additional training or denoiser evaluations.

Critical weaknesses

W1. Comparisons with recent models such as Nano Banana and GPT-image are missing. Although differences in computational requirements would make a direct comparison imperfect, at least a qualitative comparison is needed to establish the significance of this work.

W2. The quantitative evaluation focuses on content preservation, which does not directly establish editing quality. Assessing the quality of the edited images themselves, using an aesthetic quality or human preference metric, is essential.

W3. Although the proposed method is motivated by conditional Fisher geometry, the theoretical analysis establishes only local properties of individual feedback operations. The connection between reduced Fisher sensitivity and global semantic content preservation remains insufficiently justified.

W4. Related to W2, the proposed method has the lowest CLIP Similarity in Table 2. This raises concerns about whether it performs the requested edits effectively or largely reproduces the source image. Can the authors explain or analyze why the proposed method achieves the lowest CLIP Similarity?

W5. The experiments and analyses in this paper are limited to SD3. Evaluation on stronger models, such as FLUX and Qwen-image 2.1, is essential but missing.

W6. The qualitative results do not clearly demonstrate the advantages of the proposed method. The main text includes too few qualitative examples (Figure 5), and it is difficult to discern improvements over Prism or SAM-Flow. This raises concerns about the effectiveness of the proposed method.

Questions for authors

Please see the weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.