acceptodds
Withdrawn from ICLR 2027

IB-Flow: Information Bottleneck-Guided CFG Distillation for Few-Step Text-to-Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While large-scale text-to-image generative models have achieved unprecedented visual performance, their inherent reliance on multi-step iterative solvers incurs severe inference latency. Few-step distillation targeting the Classifier-Free Guidance (CFG) trajectory has emerged as the prevalent dual-dimensional compression paradigm. However, existing frameworks remain subjugated by a coarse-grained blind injection paradigm that perpetually enforces a globally static guidance strength while indiscriminately sampling the supervisor timestep. This state-agnostic design completely disregards the intrinsic nature of image generation as a dynamic evolutionary process characterized by progressive entropy reduction, which not only restricts the performance boundary of few-step compression but also precipitates severe CFG over-conditioning artifacts. To transcend these limitations, we re-examine the distillation procedure through the theoretical lens of Information Theory, formally modeling it as a dynamic mutual information game constrained by the Information Bottleneck (IB) principle. Specifically, we dismantle traditional blind assumptions via a dual-track adaptive framework. To determine the injection target, we propose an instance-aware selection mechanism that transmutes the intractable KL divergence constraint into a closed-form conditional optimum solution predicated on the local vector field norm. To regulate the injection strength, we introduce an entropy-aware schedule that dynamically decays alongside the Signal-to-Noise Ratio (SNR), applying maximal thrust for initial structural anchoring before smoothly reverting to the natural manifold to refine micro-details. Extensive empirical evaluations corroborate that our framework efficiently reduces the over-conditioning artifacts, shattering the performance ceiling to achieve state-of-the-art (SOTA) generative fidelity under extremely stringent 2-step configurations.

Primary area:
generative models
Keywords:
Few-Step Distillation, Classifier-Free Guidance, Text-to-Image Generation, Flow Matching, Information Bottleneck

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_11yXOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose a few-step generation method, called IBFlow. It improves few-step text-to-image distillation through adaptive teacher-supervision timesteps and SNR-based CFG scheduling. Motivated by the Information Bottleneck principle, it reports gains over existing two-step baselines across three model backbones.

Critical strengths
  1. the training objective is simple and easy to understand.
Critical weaknesses
  1. I have checked the offical metrics reported in TwinFlow and ArcFlow, and found that the data in Tab.1 are obviously lower than their offically reported data.
Questions for authors
  1. the authors characterize the current injection paradigm as “blind”. Could they clarify why selecting a specific teacher-supervision timestep should be preferable to sampling across multiple timesteps? Given the stochasticity of training, it is unclear why the proposed selection rule provides more reliable supervision than a sampling strategy that covers a broader range of timesteps.
  2. In the proof of lemma 1, the equality (19) cannot be obviously obtained. Can authors provide more details?
  3. the benefit of Dyn τCA⋆\tau^\star_{\mathrm{CA}} appears modest. Could the authors repeat this ablation with at least three random seeds and report the mean and standard deviation to assess whether the improvement is statistically significant?
Reviewer_66o2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes two schedules for the CFG-Augmentation term in few-step distillation: an instance-aware schedule that sets the CA supervisor timestep from the norm of the teacher’s CFG residual, and an entropy-aware schedule that decays the guidance scale with the signal-to-noise ratio. Both are presented as closed-form solutions of Information Bottleneck problems. The practical idea is simple and the 2-NFE gains are consistent in sign across different backbones.

Critical strengths
  • The authors provide sufficient theoretical proof and support, which makes the work sound and solid.
  • The proposed idea is evaluated across three different teachers and three alignment benchmarks. Each component is ablated separately.
Critical weaknesses
  • The experiments do not isolate the proposed mechanisms. The method modifies the CA term of Decoupled DMD, yet the direct control is not reported: Decoupled DMD with a static ω and randomly sampled τ_CA. Without it, the gains cannot be attributed to the proposed schedules rather than to differences in the base pipeline or the training setup.
  • In addition, the reported improvements are too small to be distinguished from noise and hyperparameter selection.
Questions for authors
  • The method is formulated as a change to the CA term of Decoupled DMD. The most direct control is therefore Decoupled DMD with static ω\omega and randomly sampled τCA\tau_{CA}. This comparison is needed.
  • In Table 5, Linear (87.82) and Cosine (88.16) both score below Constant (88.37). This suggests that the benefit may come from one specific curve rather than from decaying guidance in general. Can the authors give some explanation for this?
  • The performance gains are small. Is the performance variance large across different runs?
  • There is no user study and failure case analysis.

Minors:

  • Table 5 is not referenced in the main text.
  • DPG for IBFlow-Qwen is 88.68 in Table 2 while it is 88.67 in other tables.
Reviewer_FZSzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To solve two limitations of existing CFG distillation methods: random selection of teacher supervision timesteps and the use of static CFG guidance strengths,authors propose IB-FLOW, which introduce an instance-aware supervision target based on the vector field residual norm and an SNR-dependent guidance schedule.

Critical strengths

1 paper is well motivated

2 proposed method is simple and practical

3 proposed method works across different backbone

Critical weaknesses

1 If the conditional and unconditional predictions are identical (i.e., the CFG residual is zero), Eq. 9 would imply that the KL divergence between different timesteps is zero. However, the image distribution can still evolve from noise to clean images even without CFG. How can Eq. 9 hold in this case?

2 result doesn’t include FID and CLIP score.

3 The claim of zero-overhead adaptive target selection requires further clarification, particularly regarding whether the additional conditional and unconditional teacher predictions can be reused from existing computations.

Questions for authors

See weaknesses.

If authors solve weaknesses, I am willing to raise my score.

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

Related papers

Discussion (0)

Sign in to comment.