In-situ Autoguidance: Eliciting Self-Correction in Diffusion Models

TL;DR

Diffusion-model guidance often trades quality for diversity when using Classifier-Free Guidance (CFG). The paper introduces In-situ Autoguidance, a zero-cost method that generates a degraded prediction on-the-fly via dropout during inference to form a self-guidance signal, eliminating the need for an auxiliary model as in traditional Autoguidance. The approach grounds guidance as inference-time self-correction and demonstrates viability on ImageNet 512×512 with the EDM2-S backbone, achieving competitive Fréchet Inception Distance (FID) scores close to unguided baselines while preserving diversity. This work provides a strong zero-cost baseline for guidance and opens a flexible design space for future self-referential and perturbation-based guidance strategies.

Abstract

The generation of high-quality, diverse, and prompt-aligned images is a central goal in image-generating diffusion models. The popular classifier-free guidance (CFG) approach improves quality and alignment at the cost of reduced variation, creating an inherent entanglement of these effects. Recent work has successfully disentangled these properties by guiding a model with a separately trained, inferior counterpart; however, this solution introduces the considerable overhead of requiring an auxiliary model. We challenge this prerequisite by introducing In-situ Autoguidance, a method that elicits guidance from the model itself without any auxiliary components. Our approach dynamically generates an inferior prediction on the fly using a stochastic forward pass, reframing guidance as a form of inference-time self-correction. We demonstrate that this zero-cost approach is not only viable but also establishes a powerful new baseline for cost-efficient guidance, proving that the benefits of self-guidance can be achieved without external models.

Figures (3)

• Figure 1: Conceptual diagram of our In-situ Autoguidance method. A single model $D_\theta$ undergoes two evaluations for each sampling step. The Deterministic Evaluation produces a high-quality prediction $D_{\text{good}}$. The Stochastic Evaluation, with dropout activated, produces a degraded prediction $D_{\text{bad}}$. The difference between these two predictions from the same model forms the guidance signal, eliminating the need for any auxiliary network.
• Figure 2: A fractal-like 2D distribution experiment, adapted from Karras et al. (2024). (a) Ground truth samples. (b) Unguided sampling produces outliers. (c) Classifier-free guidance (CFG, $w=4$) eliminates outliers but severely reduces diversity. (d) Naive score truncation also harms diversity. (e) Original Autoguidance concentrates samples effectively. (f) Our proposed In-situ Autoguidance, while slightly less concentrated than (e), also successfully guides samples towards high-probability regions without a significant loss of diversity, demonstrating its viability at zero additional model cost.
• Figure 3: Qualitative results of In-situ Autoguidance. Each pair shows a ground truth image from the dataset (left) and our generated sample (right) for the corresponding class. The classes are (from left to right, top to bottom): Dog, Cat, Cuckoo, Snake. Our method produces recognizable and detailed instances of the target classes.