Technical Summary: Mitigating Diffusion Model Hallucinations with Dynamic Guidance

Problem Definition

Diffusion models, while dominant in high-fidelity image generation, suffer from "hallucinations"—samples containing structural inconsistencies (e.g., incorrect anatomy like extra fingers or missing body parts) that lie outside the support of the true data distribution. The authors attribute these hallucinations to mode interpolation.

Recent findings indicate that the learned score function in Denoising Diffusion Probabilistic Models (DDPMs) is excessively smooth in the low-density regions between data modes. While the true score function should be sharp in these intermediate regions to force the sampler to take larger, corrective steps, the learned smoothness causes the denoising process to "trap" samples, resulting in semantically invalid interpolations between incompatible modes.

Existing guidance methods (e.g., Classifier Guidance, Classifier-Free Guidance) typically steer samples toward high-likelihood regions based on a fixed condition provided at the start of sampling. The authors argue that if this fixed condition misaligns with the initial noise or the evolving trajectory, the guidance can push the sample into regions requiring large corrective steps, leading to overshooting or undershooting the target mode and generating hallucinations. Furthermore, current approaches often treat hallucinations as a post-hoc filtering problem or a text-image alignment issue, rather than a fundamental flaw in the sampling trajectory itself.

Methodology: Dynamic Guidance (DG)

The paper proposes Dynamic Guidance (DG), a method that mitigates hallucinations by adaptively sharpening the score function during the sampling process, rather than relying on static conditions or post-generation filtering.

Core Mechanism

Instead of committing to a single guidance target (class $c$) from the beginning, DG identifies the most probable mode $c^*$ for the current noisy sample $x_t$ at each denoising timestep. The guidance is then applied toward this dynamically selected mode.

The algorithm approximates a sharper score function $\hat{s}_\theta(x_t)$ as follows:
$$\hat{s}_\theta(x_t) := \sum_{c} \mathbb{1}_{c^*}(c) \nabla_{x_t} \log p_\theta(x_t | c)$$
where $c^* = \arg \max_{c \in C} \log p(c | x_t)$.

By recalculating the most probable class at every step, the guidance signal remains aligned with the local score direction, preventing the trajectory from being locked into paths that require large, hallucination-inducing interpolation steps.

Implementation Variants

1. Discrete Conditioning (Classifier Guidance): For datasets with discrete labels (e.g., ImageNet classes, shape types), DG uses a classifier to predict $p(c|x_t)$ and applies gradients toward the most likely class $c^*$.
2. Pseudo-Classes (Clustering): The method does not require pre-defined semantic classes. The authors demonstrate that clustering semantically coherent neighborhoods in the data distribution (e.g., using DINOv2 features) to create "pseudo-classes" is sufficient. This allows DG to isolate hallucination-prone directions without mapping to interpretable concepts.
3. Text-to-Image (Classifier-Free Guidance): For text-conditioned models, DG is extended by dynamically adjusting the text prompt. A Vision-Language Model (VLM) observes intermediate denoised predictions and, if the image matches a specific "branch" of a prompt (generated by an LLM to resolve ambiguities), the prompt is updated to that specific branch. This effectively sharpens the score function along the direction of the resolved ambiguity.

Selective Sharpening

A key theoretical insight is that DG selectively sharpens the score function. It increases the gradient magnitude only along directions associated with the selected guidance mode (where hallucinations occur), while leaving other directions (e.g., valid semantic variations like skin tone) unaffected. This preserves sample diversity while suppressing invalid generations.

Key Contributions

1. Dynamic Guidance Algorithm: A novel sampling strategy that adaptively selects guidance targets at each timestep to avoid low-probability interpolation regions.
2. Selective Score Sharpening: Demonstration that DG mitigates hallucinations by sharpening the score function specifically along hallucination-inducing directions, preserving diversity elsewhere.
3. First Generation-Time Mitigation: The first approach to address hallucinations during the sampling process, rather than through post-hoc detection/rejection or post-training.
4. Pseudo-Class Extension: A principled method to extend DG to scenarios without explicit class labels by using data-driven clustering.
5. Text-to-Image Adaptation: An extension of DG to latent diffusion models using dynamic prompt adjustment via VLMs and LLMs.

Experimental Results

Controlled Datasets

The authors evaluated DG on synthetic datasets where hallucinations are strictly defined:

• Single Shapes: Images with 1-3 instances of the same shape. Hallucinations are mixed shapes.
  • Result: DG achieved a 74.21% reduction in hallucination rate (HR), significantly outperforming Variance Filtering (11.58%) and Classifier Guidance (11.05%).
• Mixed Shapes: Images with up to one instance of each shape. Hallucinations are multiple instances of the same shape.
  • Result: DG achieved a 72.97% HR reduction, while Classifier Guidance actually increased hallucinations (-1.92%).
• Hands-11k: Human hand images. Hallucinations are anatomical errors.
  • Result: DG achieved a 45.12% HR reduction, outperforming Variance Filtering and Classifier Guidance.

Large-Scale Benchmarks (ImageNet-1k)

Since hallucinations are hard to define in ImageNet, the authors used proxy metrics: Precision (fraction of samples in the real data support) and Inception Score (IS).

• Performance: DG consistently outperformed Classifier Guidance across various guidance scales.
• Metrics: In the best configuration, DG improved Precision by 8 points and Inception Score by 5 points compared to the baseline, while maintaining diversity (Recall was not significantly reduced).
• Ablation: The authors found that applying DG only during intermediate timesteps $[T_1, T_2]$ (specifically 800 to 400 in their setup) yielded the best results, avoiding the rigidity of full-trajectory guidance.

Text-to-Image Generation

Applied to Stable Diffusion 2.1 with a user study of 18 participants:

• Hallucination Reduction: DG reduced the perceived hallucination rate for 100% of participants, with an average reduction of 14.1% (from 52.7% to 45.3%).
• Preference: Participants preferred DG-generated images over the baseline (38.7% vs 27.3% preference rate).

Significance and Claims

The paper claims that Dynamic Guidance represents a fundamental shift in addressing diffusion model hallucinations. By intervening directly in the generative process, DG prevents hallucinations from arising in the first place, avoiding the computational waste of generating and then discarding flawed samples.

The authors emphasize that their approach is principled: it relies on the mathematical necessity of sharpening the score function in low-density regions to prevent mode interpolation. Crucially, they demonstrate that this can be achieved without sacrificing the "valid" interpolations that contribute to sample diversity and generalization. The ability to use pseudo-classes and dynamic prompt adjustment suggests that DG is a scalable solution for complex, real-world generation tasks where rigid class definitions are insufficient.

The work concludes that DG provides a step toward more reliable diffusion models, offering a mechanism to control the trade-off between fidelity and diversity by targeting the specific geometric properties of the learned score function that lead to structural inconsistencies.