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

Jul 16, 2026

Upload by Eden

IB-Flow proposes an information-theory-based approach to few-step text-to-image generation that targets the Classifier-Free Guidance (CFG) trajectory. Existing distillation methods use coarse blind injection with globally static guidance and indiscriminate supervisor timestep sampling, ignoring image generation as a dynamic entropy-reduction process and leading to severe over-conditioning artifacts. IB-Flow models distillation via an Information Bottleneck constrained mutual-information game and introduces adaptive instance-aware injection targeting plus an entropy-aware guidance schedule. Experiments show artifacts are eliminated and SOTA fidelity is reached under extremely stringent 2-step settings.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

Abstract

Introduction

arXiv :2607 .09133v1 [ cs .CV] 10 Jul 2026 IB-Flow: Information Bottleneck-Guided CFG Distillation for Few-Step Text-to-Image Generation Yiting Wang1 Jingyi Zhang2 Wenhu Zhang3 Ke Chao4 Yves Liang 1 Kun Cheng2 Kang Zhao2 ∗ 1Tsinghua University 2Wan Team, Alibaba Group 3HKUST 4Beijing Normal University 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 coarsegrained blind injection paradigm that perpetually enforces a globally static guidance strength while indiscriminately sampling the supervisor timestep. This stateagnostic design completely disregards the intrinsic nature of image generation asa 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 zero-overhead closed-form solution predicated on the local vector field norm. To regulate the injection strength, we introduce an entropy-aware schedule that dynamically decays alongside the Signalto-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 fundamentally eradicates over-conditioning artifacts, shattering the performance ceiling to achieve state-ofthe-art (SOTA) generative fidelity under extremely stringent 2-step configurations. 1 Introduction Recent advances in large-scale generative models, particularly those based on Diffusion Models [9,24] and Flow Matching [16], have achieved unprecedented success in text-to-image generation. By constructing continuous probability paths from simple prior distributions to complex data distributions, they enable highly realistic visual synthesis. However, these models heavily rely on iterative numerical solvers during the sampling phase, typically requiring dozens or even hundreds of Neural Function Evaluations (NFEs) . This sequential denoising nature leads to severe inference latency and massive computational overhead, fundamentally limiting their deployment in real-time interactive applications. To break the inference efficiency bottleneck of diffusion models, the core objective of few-step generation fundamentally necessitates a dual-dimensional compression paradigm, wherein step distillation ∗Corresponding author. Preprint. (1) A poster of a cat dressed as Napoleon Bonaparte, holding a yellow wedge of cheese. (2) A cat sitting next to a modern television that is currently displaying a nature documentary. (3) An orange tent pitched by a serene lake, with a circular gold ring lying on the autumn leaves. (4) A focused young woman with spectacles reading a thick book ata desk lit by a brass lamp. (5) A painting of a majestic lion with a speech bubble saying"meow" . (6) A painting of a feline super math wizard, surrounded by floating mathematical equations. Figure 1: Demonstrating extreme few-step image generation at only 2 NFE, our IBFlow seamlessly improves perceptually orthogonal qualities compared to Arcflow, which improves from: (1,2) structural fidelity,(3,4) textural

FAQs

Why do existing few-step CFG distillation frameworks struggle with inference quality?

They enforce globally static guidance strength and sample the supervisor timestep blindly, which ignores the dynamic entropy-reduction nature of image generation. This mismatch produces severe CFG over-conditioning artifacts and limits achievable compression quality.

How does IB-Flow address the injection target selection problem?

It uses an instance-aware selection mechanism that turns an intractable KL divergence constraint into a zero-overhead closed-form solution based on the local vector field norm.

How does IB-Flow schedule CFG injection strength during generation?

It introduces an entropy-aware schedule that decays alongside the signal-to-noise ratio (SNR), using maximal thrust early for structural anchoring and then smoothly returning to the natural manifold for micro-detail refinement.

Jul 16, 2026

Upload by Eden