acceptodds
Withdrawn from ICLR 2027

VAGS: Velocity Adaptive Guidance Scale for Image Editing and Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Classifier-free guidance (CFG) is the primary control over how strongly text semantics move a flow-based sampler, yet standard practice holds its scale fixed across the entire ODE trajectory. This is a fundamental mismatch: early steps are noise-dominated and carry weak semantic signal, while late steps commit image structure and demand stronger directional commitment; more critically, the value of any guidance strength depends on whether the guided velocity is consistent with the model's current dynamics or working against them. We propose Velocity-Adaptive Guidance Scale (VAGS), a training-free replacement that multiplies the nominal scale by a bounded factor combining a temporal signal-level term with the cosine similarity between task-relevant velocity fields. For inversion-free editing, VAGS measures the alignment between source- and target-guided velocities, so edit strength at each step reflects local compatibility between preservation and transformation. For generation, VAGS-Gen uses the alignment between unconditional and conditional velocities as the analogous signal. Neither variant requires fine-tuning, auxiliary networks, or extra forward passes, and fixed CFG is recovered as a special case. On PIE-Bench and DIV2K for editing, and COCO17, CUB-200, and Flickr30K for generation, VAGS consistently improves structural fidelity and generation quality over fixed CFG and recent training-free guidance variants. The codebase is available in the supplementary materials.

Primary area:
generative models
Keywords:
Generative AI, Flow Matching, Image Editing

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

This paper proposes VAGS, a training-free adaptive classifier-free guidance method for flow-based image editing and generation. VAGS adjusts the guidance scale at each sampling step using a combination of denoising progress and velocity alignment. For editing, it measures source-target velocity similarity, while for generation, it uses conditional-unconditional velocity similarity. Experiments on five benchmarks demonstrate improved image preservation and generation quality with negligible computational overhead.

Critical strengths
  1. Simple and efficient design. VAGS is training-free, requires no additional network evaluations, and can be incorporated into existing flow-based samplers with negligible computational overhead.
  2. Encouraging experimental results. The method achieves substantial improvements in background preservation and consistent FID reductions across multiple generation benchmarks. The scheduler comparisons and sensitivity analyses provide additional evidence of its effectiveness.
Critical weaknesses
  1. Insufficient isolation of velocity-based adaptation. The proposed method combines temporal modulation with velocity alignment, but the experiments do not sufficiently isolate their individual contributions. Although several existing guidance schedules and matched-mean CFG baselines are considered, direct comparisons against time-only variants of the same formulation or shuffled/constant alignment signals are missing. The improvements over monotone schedules are also relatively modest on several metrics, making the independent benefit of velocity alignment unclear.
  2. Unclear trade-off between preservation and editing effectiveness. VAGS significantly improves background preservation, but the CLIP similarity of the edited region decreases compared with the corresponding baselines. For example, on PIE-Bench, FlowEdit+VAGS reduces CLIP-Edited from 23.80 to 23.08. The results therefore do not conclusively establish whether VAGS improves the overall editing quality or primarily shifts the preservation–editability trade-off. More comprehensive trade-off analysis or human evaluation would strengthen the conclusions.
  3. Limited cross-backbone validation. All experiments are conducted using Stable Diffusion 3.5 Large. Since VAGS is presented as a general plug-and-play guidance strategy, additional evaluations on other flow-based architectures, such as FLUX, would better demonstrate its generalizability.
  4. Incomplete comparisons with recent adaptive guidance methods. Several closely related recent approaches, such as CFG-Ctrl (CVPR 2026) and CFG-MP (ICML 2026), are not included in the experimental comparisons. These methods also aim to improve flow-based guidance and have been evaluated on Stable Diffusion 3.5. Comparisons with these approaches would help establish the advantages of VAGS over existing guidance strategies.
  5. Limited justification of velocity alignment as a guidance reliability measure. The method assumes that velocity alignment provides a meaningful signal for adjusting guidance strength. However, source-target velocity disagreement may indicate either successful semantic editing or undesirable structural distortion, while conditional-unconditional velocity alignment does not necessarily imply generation quality. More direct analyses connecting the proposed cosine similarity to editing success and generation reliability would strengthen the methodological justification.
Questions for authors
  1. Contribution of velocity alignment: What is the independent contribution of velocity alignment beyond temporal modulation? A direct ablation comparing the full method with a time-only variant of the same exponential formulation would help clarify this.
  2. Preservation–editability trade-off: How does VAGS balance improved background preservation against the observed reduction in edited-region CLIP similarity? A more comprehensive trade-off analysis or human evaluation would be helpful.
  3. Validity of velocity alignment: What empirical evidence supports the assumption that velocity cosine similarity reliably indicates guidance compatibility, rather than merely reflecting differences between conditional predictions?
Reviewer_X1qoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces VAGS, a method for per-step adaptive guidance in flow-based image editing and generation. It replaces a fixed CFG scale with a scheduler determined by both the current timestep and the cosine similarity of velocity vectors. For editing, the cosine similarity measures the alignment between the source-conditioned and target-condition velocity. For generation, it assesses the cosine similarity between unconditional and conditional velocities. The paper argues that this signal respectively reflects the compatibility between preservation and transformation in editing and the alignment with the data prior in generation. VAGS improves the quality of editing and generation on various benchmarks without requiring any additional NFE.

Critical strengths
  1. The method is simple and introduces minimal extra computational cost. VAGS can be integrated into existing true CFG pipelines without increasing the NFE.
  2. The experiment scope is extensive. The paper conducts experiments across both editing and generation tasks on multiple benchmarks, and compares their approach against various guidance schedulers and CFG control methods, demonstrating comprehensive evaluation.
Critical weaknesses
  1. The novelty and design evidence of the proposed method are not sufficient and solid. [1] provides a systematic analysis of how CFG weight varies with the timestep is better, demonstrating that monotonically increasing schedules can consistently improve CFG performance. [2] observes that using cosine similarity to determine a CFG policy lacks a robust basis, and furthermore it explicitly suggests that similarity may not be a good control signal of adaptive guidance. The paper proposes three design constraints for adaptive guidance: smoothness, symmetric modulation around a fixed CFG value, and regime-dependent interpretation. However, these appear more like engineering intuitions since they lack theoretical derivation, nor are there any ablation studies to justify why these constraints are necessary.
  2. The claims of the adaptivity of model dynamics and compatibility are not supported by analysis or experiments. In the introduction, the cosine similarity term for editing tasks is interpreted as a "local compatibility between preservation and transformation", while for generation tasks is described as an "analogous measure of compatibility with the data prior". However, neither of these concepts is subsequently defined rigorously or verified directly. More importantly, as shown in Figure 4 and Appendix Fig. 5, the adaptive guidance scale follows a very smooth, timestep-dependent exponential trend. This strongly suggests that the sample-specific cosine term sis_i has virtually no impact on the guidance scale, and thus the guidance is determined solely by an exponential function of the temporal factor 2σi−12\sigma_i - 1. This does not address the judgment requirement for "the value of any guidance strength depends on whether the guided velocity is consistent with the model’s current dynamics or working against them" in the abstract.

[1] Analysis of Classifier-Free Guidance Weight Schedulers.
[2] How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models.

Questions for authors
  1. Please clarify the technical advancements and significance of VAGS relative to existing adaptive CFG approaches.
  2. How much does sis_i actually contribute to the final adaptive scale and experiment results? How to demonstrate the improvements in the results are not simply due to the use of the monotonically increasing exponential scale scheduler as advocated in [1]?
  3. What is the necessity of the "adaptive rule" / "three constraints" proposed in line 215? The current paper simply lists "smoothness", "symmetric modulation", and "regime-dependent interpretation" without providing any theoretical justification or experimental validation.
  4. Paper writing requires improvement. Some of the main figures are of low resolution, such as those in Figure 2 and 4. Furthermore, some concepts are introduced using strong, assertive words, yet they lack clear further definitions, derivations, or experiment evidence.

[1] Analysis of Classifier-Free Guidance Weight Schedulers.

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

This paper proposes VAGS, a training-free dynamic CFG method that gradually adjusts the guidance scale by combining the denoising progress and the cosine similarity of velocity fields. In inversion-free image editing and text-to-image generation, this method achieves good experimental results on PIE-Bench, DIV2K, COCO17, CUB-200, and Flickr30K, with almost no additional inference overhead.

Critical strengths
  1. Low computational cost and easy integration. It requires no additional forward computation or training and can be directly applied to existing CFG sampling pipelines.

  2. The experimental results are competitive. The paper evaluates their method on PIE-Bench, DIV2K, COCO17, CUB-200 and Flickr30K, covering image editing and text-to-image generation tasks.

  3. This paper provides extensive ablation studies and analyses. They compare different guidance scheduling strategies and the modulation strength on generation quality and editing fidelity. Many ablation studies can help clarify the behavior and applicable scope of the method.`

Critical weaknesses
  1. Technical contribution is incremental. There has been much research on dynamic CFG scheduling and geometric adjustments of guidance vectors [1][2][3][4]. The main contribution of this paper is combining time progress with the cosine similarity of velocity fields and applying this combination to image generation and editing.

  2. The effectiveness of the velocity alignment mechanism. In figure 10, the cosine similarity between unconditional and conditional velocity is tightly concentrated around 1, and the average guidance scale varies little across different samples. I would like to ask whether the performance improvement mainly comes from time scheduling rather than velocity alignment.

  3. The geometric assumption lacks sufficient support. In eq 9, the authors adjust guidance using the cosine similarity between unconditional and conditional velocity, but the CFG update direction is determined by their difference in eq 1. These two geometric relationships are not equivalent.

  4. There is an inconsistency in the time modulation term. Appendix D.1 claims that the absolute value of the time factor during the noisy stage is small. However, the absolute value of (2σi−1)(2\sigma_i-1) in the eq 10 approaches 1 at the noisy endpoint. This explanation is inconsistent.

  5. The results between fidelity and editing performance have different trends in image editing. VAGS substantially improves structural fidelity, but the Whole CLIP and Edited CLIP metrics on PIE-Bench decline in Tab 1. I would like to see some analysis of preservation and editability.

  6. Generalization remains insufficient. The experiments are mainly conducted based on Stable Diffusion 3.5 Large, with a lack of evaluation in different model architectures.

References:

[1] Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models. ICLR 2025

[2] CFG++: Manifold-Constrained Classifier Free Guidance for Diffusion Models. ICLR 2025

[3] Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking. NeurIPS 2025

[4] Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance. CVPR 2024

Questions for authors

Most of my questions are listed in the weaknesses, and I aslo have some questions for authors:

  1. More failure cases. I would like to see some failure cases to show the boundary of the proposed method.

  2. Is the editing task sensitive to the source guidance scale? The paper uses a fixed source guidance score in the editing task, I would like to ask how the performance will change under different source guidance scores.

  3. If using a different ODE solver, can the VAGS still achieve a stable improvement?

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

Related papers

Discussion (0)

Sign in to comment.