Rule-Guided Diffusion

• Rule-guided diffusion is a method that dynamically applies adaptive guidance signals during the reverse diffusion process to align generated samples with target distributions.
• It extends classifier-free guidance by replacing fixed weights with a learnable network that optimizes for both fidelity and diversity, as measured by FID and CLIP scores.
• The approach supports reward-guided sampling, enabling enhanced image-text alignment and richer control in diverse generative modeling tasks.

Rule-guided diffusion refers to the modification of the diffusion model’s sampling process by adaptively applying guidance signals—such as conditioning, reward functions, or other constraints—in order to steer the generation toward samples that satisfy explicit rules, conditions, or target distributions. This approach generalizes and improves upon classical classifier-free guidance (CFG) by dynamically learning or modulating the way guidance is injected, allowing the model to achieve better alignment with the conditional target distribution, improved perceptual quality, or additional reward-driven objectives.

1. Classifier-Free Guidance and Its Limitations

Classifier-free guidance (CFG) is a fundamental technique in diffusion models for improving conditional generation quality. In CFG, for each reverse diffusion step, the denoiser is evaluated both with and without the conditioning signal (e.g., a text prompt or class label). The two outputs are linearly combined with a guidance weight ω: $$\tilde{\epsilon}_\theta(x, t, c) = \epsilon_\theta(x, t, c) + \omega[\epsilon_\theta(x, t, c) - \epsilon_\theta(x, t)]$$ Here, $\epsilon_\theta(x, t, c)$ is the noise prediction with conditioning and $\epsilon_\theta(x, t)$ is the unconditional prediction.

The scalar guidance weight $\omega$ controls the strength of conditioning. A higher $\omega$ yields sharper, more conditionally faithful samples (for example, images better aligned to a text prompt); however, a fixed $\omega$ introduces trade-offs: overly high weights may distort the distribution, reduce diversity, or cause misalignment with the true conditional target distribution $p(x_0|c)$.

2. Learning Adaptive Guidance Weights

Rather than using a static, hand-tuned guidance weight, adaptive rule-guided diffusion introduces guidance weights $\omega_{c, (s,t)}$ that are continuous functions of the conditioning $c$, the time pair $(s, t)$ (i.e., the source and destination times between which denoising occurs), and, implicitly, the sample state.

A learnable guidance network (typically a small neural network) is trained to output the optimal guidance weights that minimize the mismatch between the distribution of samples generated by the guided process and the target conditional distribution. One central principle is distributional self-consistency: the distribution obtained by denoising a noised conditional sample with the adaptive guidance should closely match the true process at each intermediate time.

A typical loss for training such a guidance network is

$$\mathcal{L}_{\ell_2}(\omega) = \mathbb{E}_{x_0, c, s, t} \|\tilde{x}_s(\omega_{c,(s,t)}) - x_s\|^2_2$$

where $x_s$ is a sample from the true noising process at time $s$, and $\tilde{x}_s$ is the result of denoising from $(x_t, t)$ to $s$ using the guided process with the learned weight.

This dynamic approach allows the guidance to adjust as needed across denoising time, conditioning context, or sample state, leading to better sample fidelity and less over-concentration around “mode” outcomes.

3. Distributional Alignment and Theoretical Rationale

The key theoretical motivation for adaptive rule-guided diffusion is to reduce the mismatch between the guided sampling distribution and the true conditional distribution $p(x_0|c)$. With a fixed guidance weight, the sampled distribution generally drifts from the target—even if the perceptual quality or condition satisfaction seems high—since the correction is not context- or time-aware.

By making $\omega = \omega_{c,(s,t)}$ a learned, continuous function, one can correct imperfections in the base denoiser more responsively, ensuring that both diversity and conditional accuracy are retained. Empirically, image synthesis tasks such as unconditional and class-conditional ImageNet or CelebA generation show that adaptive guidance consistently produces lower Fréchet Inception Distance (FID) scores than static baselines, indicating improved distributional alignment and realism.

4. Reward-Guided Sampling Extensions

The rule-guided diffusion framework naturally extends to reward guidance, where the objective is not just to match a conditional distribution, but also to maximize an explicit reward $R(x_0, c)$ evaluated on the clean data and associated conditioning. The guidance network is optimized to bias sampling toward high-reward regions, thereby tilting the generated distribution toward desirable semantic or structural properties.

For example, in text-to-image tasks, the reward function may be the CLIP score $R(x_0, c) = \mathrm{CLIP}(x_0, c)$, measuring textual alignment. During guidance weight learning, the objective includes an additional reward term, making the model more likely to sample $x_0$ that yield higher $R(x_0, c)$. This promotes images that not only look realistic but are also better aligned with the input text, as quantified by both FID and CLIP alignment scores.

5. Experimental Results and Applications

Empirical investigations on both synthetic (low-dimensional) and real-world (high-dimensional) data highlight the value of rule-guided diffusion. On standard image datasets such as ImageNet and CelebA, adaptive guidance leads to lower FID (improved image realism and variety) versus static $\omega$ strategies. In large-scale text-to-image experiments (e.g., MS-COCO), adding a reward term to guide the sampling process using CLIP results in improved image-text alignment, with a measured, sometimes minor trade-off in FID due to enhanced conditional concentration.

This approach is applicable wherever sample quality, diversity, and adherence to specified rules or conditionings are simultaneously required—for instance, in high-fidelity visual synthesis, semantic image editing, artistic image generation, and any generative modeling task with complex or evolving rule sets.

6. Implications, Limitations, and Future Directions

Replacing static guidance parameters with learned, dynamic guidance weight functions in rule-guided diffusion models provides a principled, empirically validated means for improved sample quality and conditional controllability. The method supports better trade-offs between perceptual sharpness and distributional coverage while incorporating flexible, potentially user-defined or reward-based guidance.

Prospective research includes formal analysis of the guidance learning objective, integration with advanced sampling schemes (e.g., Sequential Monte Carlo or Markov Chain Monte Carlo corrections), and the development of more expressive or task-specific reward functions for richer control. The generalization to structure-constrained, sequential, or discrete data domains is a promising area for further paper.

---

Guidance Mechanism	Adaptivity	Evaluation Metric	Application Domain
Static CFG (scalar $\omega$)	None	FID, CLIP alignment	Image/text synthesis
Learned $\omega_{c,(s,t)}$	Conditioning, time	FID, distributional distance	Image/text synthesis
Reward-guided $\omega_{c,(s,t)}$	Conditioning, time	FID, CLIP (or general $R(x_0, c)$)	Reward-guided sampling

Rule-guided diffusion thus generalizes and subsumes earlier practices such as CFG, providing a unified approach to adaptive, controllable generation under complex or evolving user-specified rules (Galashov et al., 1 Oct 2025).