acceptodds
Under review as a conference paper at ICLR 2027

Time-Warping Guidance in Flow Matching

Abstract

Inference-time guidance methods for diffusion and flow-matching image generators, from Classifier-Free Guidance (CFG) to recent autoguidance variants, form contrastive velocities between a model prediction and a less-informed baseline. We introduce TWiG (Time-Warping Guidance), which augments CFG with a second contrastive term whose probe lies at a different point in latent-time space, namely the conditional branch evaluated at a time-warped query that approximates a forward-process sample at a different sampling time. The warped query is built from a Tweedie noise estimate cached from the previous step, with a closed-form proxy and an explicit error bound. The schedule that selects the warped time is governed by a primary calibration scalar, together with two clip bounds. We hypothesize on geometric grounds that the calibration scalar is near and observe a corresponding empirical alignment peak on five backbones. We evaluate TWiG using automatic metrics spanning distribution, quality, and diversity (FID, CMMD, IS, Precision/Recall, CLIP, Image Reward, Aesthetic, Vendi, LPIPS, CVS), the compositional benchmarks GenEval and GenEval2, and a blinded pairwise human study. Across two evaluation backbones, TWiG improves CFG and shows favorable results against the evaluated baselines.

Primary area:
generative models
Keywords:
Diffusion models, rectified flow, text-to-image generation, classifier-free guidance, inference-time guidance, time-warped guidance, sampling diversity, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rKorOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Time-Warping Guidance(TWIG) , an inference-time guidance method for flow matching models. TWIG augments classifier-free guidance(CFG) with a second contrast velocity term using a transformed latent and time. The transformation is based on cached noise estimate from previous steps and an empirically identified time schedule. The method introduces a calibration parameter, hypothesized to target an optimal time.

Experiments report improvements over CFG and competitive results against existing guidance methods, including favourable human preferences, on two image-generation backbones. However, this comes at the price of significantly higher NFE or latency-- which is a key overhead to pay.

Critical strengths
  1. The paper presents a novel geometric view about time-shifted guidance in diffusion/flow models, offering a conceptually straightforward way to improve inference-time guidance. TWIG is simple to implement and requires no additional auxiliary model. Experiments have validated the intuition.

  2. The paper evaluates multiple quality, diversity, and compositionality metrics, as well as human preferences, and reports consistent gains over CFG across two backbones.

Critical weaknesses
  1. The novelty lies mainly in the jointly time-warped latent construction and its scheduling, however the advantages over existing approaches needs to be better demonstrated and explained more clearly. Please address the questions in the "Questions for Authors" section.

  2. TWIG's improvement on performance against baselines is modest, while requires significantly higher NFE. The main evaluation Table 1 only provides equal-NFE baseline on CFG, and most two-pass baselines use 100 NFE. In the meantime, TWIG uses 150 NFE. As a result, TWIG's advantage under equal NFE setting remains unclear.

Questions for authors
  1. please explain the insight behind Lemma 2 on TWIG's performance? This lemma itself bounds the reconstructed latent - which is the network's input, on noise estimate ∣∣ϵ^−ϵ∣∣||\hat{\epsilon}-\epsilon||. Two follow-ups questions on this lemma are: (1) How can we obtain the error bound on noise estimate? (2) How does the error bound on network input propagates to network output, i.e., the velocity prediction?

  2. Since the method relies on Tweedie's formula for noise estimation, which works best for multi-step sampling scenario, how well does TWIG perform under few-step sampling regime?

  3. Could the authors provide (at least some) results for equal NFE evaluation on two-pass baselines?

  4. Per Eq.10, vtoraclev_t^\text{oracle} should approach vtCFGv_t^\text{CFG} as t→0t \rightarrow 0. But Figure 2 and Figure 4 show on the contrary. Please explain this.

  5. Given TWIG requires one more forward run in each step compared to the vanilla CFG, please explain what role the second contrast term in TWIG provides.

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

This paper proposes Time-Warping Guidance (TWiG), a training-free method that extends CFG with a contrastive term based on a time-warped latent and timestep. Experiments on Flux2-Klein and Z-Image demonstrate improvements in image quality, diversity, and compositional alignment over standard CFG.

Critical strengths
  1. The paper presents an interesting extension of CFG by jointly modifying the latent and timestep for cross-time guidance. The proposed construction is mathematically well motivated and requires no additional training.
  2. The experimental evaluation is comprehensive, covering multiple guidance baselines, quality and diversity metrics, compositional benchmarks, and human preference comparisons.
Critical weaknesses

The main concern is that the source of TWiG's improvements is not sufficiently established. While the matched latent-time construction is mathematically motivated, Appendix K shows that the exact warped prediction is better aligned with the oracle, whereas the estimated prediction used in practice is less aligned. This suggests that the approximation error may play an important role in the guidance mechanism. Without more targeted comparisons, it remains unclear how much of the improvement comes from the proposed geometric construction rather than the approximation introduced during sampling.

Questions for authors
  1. Could the authors compare TWiG with a simpler variant that changes only the timestep while keeping the original latent unchanged?
  2. Appendix K shows that the exact warped prediction has better oracle alignment than the current conditional prediction, while the estimated warped prediction has worse alignment. Does using the exact matched query also improve the final guidance update, or does the improvement depend on the approximation error introduced by the cached noise estimate?
  3. The target timestep (t⋆=0.5t^\star=0.5) is motivated by the observed peak in velocity alignment. How does generation quality change when different target timesteps are used?
  4. TWiG requires an additional model evaluation at each sampling step. Although it outperforms compute-matched CFG, SG-prev achieves competitive results with fewer model evaluations. Could the authors further clarify when the additional computational cost of TWiG is justified relative to SG-prev, particularly given the relatively small gains on some metrics?
Reviewer_BsCMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Time-Warping Guidance for image generation with flow-matching models. It adds an extra model prediction at a changed latent and time to standard classifier-free guidance. The method needs no further training. The experiments show better image quality, diversity, and prompt alignment, at the cost of more model evaluations.

Critical strengths
  1. TWiG gives a simple and practical extension of CFG. It gives a clear way to construct an additional prediction by changing both the latent and the query time. It uses the same pretrained model and needs no further training or architecture changes. This makes the method easy to add to existing sampling pipelines and gives the contribution practical value.

  2. The experimental results support for its benefits. The paper evaluates image quality, diversity, and prompt alignment on two backbones, and includes a blinded human study. It also compares against CFG using the same total number of velocity-field calls. This comparison is especially useful because the gains cannot be explained simply by giving TWiG more model calls. Together, these results provide reasonable evidence that the proposed query is a useful way to spend additional sampling computation.

Critical weaknesses
  1. The theory does not explain why TWiG improves CFG. Lemmas 1–2 describe the warped query and its error, but do not show why moving away from its prediction gives a better sampling direction. The useful sign of the correction is supported mainly by empirical diagnostics. Since this correction is the central idea, the missing justification limits understanding of when TWiG should help or hurt.

  2. The benefit of the main new component is not clearly isolated. Existing Self-Guidance already uses predictions across time. TWiG also changes the latent, time schedule, normalization, and guidance strength. The reported image-quality ablations do not compare a time-only query using the same schedule, normalization, weight, and computation. Therefore, it remains unclear how much improvement comes specifically from the matched latent-time construction. This makes its contribution beyond existing guidance methods harder to assess.

Questions for authors
  1. Under what conditions does moving away from the warped prediction improve the CFG sampling direction? Can the analysis explain when this correction helps or hurts?

  2. Can you compare TWiG with a query that changes only the time, keeping the time schedule, normalization, guidance weight, and number of model calls the same? This would clarify the benefit of the proposed latent correction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.