S2-Guidance: Stochastic Self Guidance for Training-Free Enhancement of Diffusion Models

Chubin Chen, Jiashu Zhu, Xiaokun Feng, Nisha Huang, Meiqi Wu, Fangyuan Mao, Jiahong Wu, Xiangxiang Chu, Xiu Li

Introduction

Diffusion models have become the cornerstone of modern generative AI, yet they often struggle with a fundamental trade-off between image quality and prompt adherence. While Classifier-Free Guidance (CFG) is the standard tool for steering these models, it is prone to over-extrapolation, leading to artifacts, semantic incoherence, and a loss of fine-grained detail.

$S^2$-Guidance (Stochastic Self-Guidance) introduces a training-free mechanism to mitigate these issues. By leveraging the inherent redundancy in Diffusion Transformers (DiTs), the method extracts a corrective signal from the model itself, steering the generation process toward higher fidelity without requiring external models or additional training.

The Problem: The "Dirty Secret" of CFG

Classifier-Free Guidance works by pushing the model's output away from an unconditional prediction and toward a conditional one. However, this extrapolation often moves the sampling trajectory into low-probability regions of the data distribution. This results in "mode shifting"—where objects are misplaced or distorted—and a general degradation of textures and physical plausibility.

Previous attempts to solve this, such as Autoguidance, required training a separate "weak" version of the model to serve as a reference point. For massive, state-of-the-art models like SD3.5 or Flux, training such auxiliary models is computationally prohibitive and often impossible for closed-source systems.

Methodology: Stochastic Self-Guidance

$S^2$-Guidance operates on a simple yet powerful observation: modern DiTs are highly redundant. The authors propose that a "weak model" already exists within the full model in the form of its sub-networks.

1. Stochastic Block-Dropping

During each denoising step, $S^2$-Guidance randomly deactivates a small subset of the transformer blocks (typically around 10%). This creates a temporary sub-network that is structurally similar to the full model but slightly less capable.

2. Generating the Self-Guidance Signal

The difference between the full model's prediction and this stochastic sub-network's prediction identifies areas of "epistemic uncertainty." Where the sub-network fails or produces artifacts, the full model provides a corrective vector. This signal is used to steer the final sampling trajectory away from the errors identified by the sub-network.

3. Temporal Integration

While computing multiple sub-networks at every step would be costly, the authors demonstrate that one random block-drop per timestep is sufficient. Because diffusion is an iterative process, the stochastic noise from these individual samples is naturally smoothed out over the course of the 20–50 sampling steps, resulting in a stable and high-quality corrective path.

Experimental Results

The researchers evaluated $S^2$-Guidance across a variety of benchmarks, including ImageNet (class-conditional), HPSv2.1 and T2I-CompBench (text-to-image), and VBench (text-to-video).

  • State-of-the-Art Performance: On ImageNet ($256 \times 256$), $S^2$-Guidance achieved a record-breaking Inception Score of 259.22 and a state-of-the-art FID of 2.03.
  • Complex Composition: In T2I-CompBench, the method showed significant improvements in handling complex attributes like color and shape, outperforming standard CFG and other advanced strategies like CFG++.
  • Video Fidelity: When applied to Text-to-Video models (Wan-1.3B and 14B), $S^2$-Guidance corrected common physical errors, such as distorted human faces and unnatural motion (e.g., "sideways sliding" vehicles), leading to more physically plausible generations.

Key Insights

  • Redundancy as a Feature: The success of $S^2$-Guidance suggests that the over-parameterization of modern DiTs can be directly exploited for inference-time refinement.
  • Bayesian Grounding: Beyond empirical success, the authors provide a formal derivation showing that $S^2$-Guidance is a principled approach to uncertainty-aware correction from a Bayesian perspective.
  • Training-Free Implementation: The method requires zero extra parameters and zero fine-tuning, making it a "plug-and-play" enhancement for any DiT-based architecture.

Limitations and Future Work

The primary trade-off for $S^2$-Guidance is computational overhead. The extra forward pass for the sub-network increases inference time by approximately 40%. While this is non-trivial, the authors argue the quality gains are more significant than what could be achieved by simply increasing the number of sampling steps.

Future research may explore applying this self-guidance mechanism to older U-Net-based models or investigating whether similar internal sub-networks can be used to improve the faithfulness and reasoning of Large Language Models (LLMs).

Summary updated

BibTeX
@inproceedings{chen2026s,
  author = {Chubin Chen and Jiashu Zhu and Xiaokun Feng and Nisha Huang and Meiqi Wu and Fangyuan Mao and Jiahong Wu and Xiangxiang Chu and Xiu Li},
  title = {{S2-Guidance: Stochastic Self Guidance for Training-Free Enhancement of Diffusion Models}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=9RyRIyzgT5},
}
Figure