Classifier-Free Diffusion Guidance

Enables conditional image generation in diffusion models by jointly training on both conditional (text-to-image) and unconditional (unconditional noise) data, then interpolating between conditional and unconditional score estimates at inference time using a guidance scale parameter. This eliminates the need for a separate pre-trained classifier network, reducing computational overhead and training complexity compared to classifier-based guidance approaches that require gradient computation through an external classifier.

Implements a post-training inference mechanism that interpolates between conditional and unconditional score estimates using a scalar guidance weight (w), enabling real-time control over the quality-diversity tradeoff without retraining. The interpolated score is computed as: s_guided = s_conditional + w * (s_conditional - s_unconditional), allowing practitioners to dynamically adjust sample fidelity from pure diversity (w=0) to maximum fidelity (w>1) at inference time.

Implements a training procedure that simultaneously optimizes a single diffusion model on both conditional and unconditional objectives by randomly dropping the conditioning signal during training (with probability ~10-50%), forcing the model to learn both conditional and unconditional score functions within a shared parameter space. This approach avoids training two separate models while enabling the guidance mechanism to interpolate between learned conditional and unconditional behaviors.

Implements the mathematical mechanism for combining conditional and unconditional score estimates at inference time through weighted linear interpolation in score space. Given pre-computed score estimates from both conditional (s_θ(x_t|c)) and unconditional (s_θ(x_t)) models, the guided score is computed as: s_guided = s_θ(x_t|c) + w·(s_θ(x_t|c) - s_θ(x_t)), where w is the guidance scale. This approach operates entirely in the score function space without requiring classifier gradients or additional model components.

Implements the training objective that enables a single diffusion model to learn both conditional score functions (∇log p(x_t|c)) and unconditional score functions (∇log p(x_t)) through a unified denoising objective. During training, the model receives either a conditioning signal (text embedding, class label, etc.) or a null/masked signal with equal probability, forcing it to learn robust score estimates for both cases. The model learns to predict noise residuals that are consistent with both conditional and unconditional distributions.

Implements the inference-time sampling procedure that uses interpolated guided scores to generate conditional samples with controlled fidelity. During the reverse diffusion process (from noise to image), at each timestep the model computes both conditional and unconditional score estimates, interpolates them using the guidance scale, and uses the guided score to determine the next denoising step. This enables real-time control over sample quality without retraining, by adjusting the guidance scale parameter.

Implements the training mechanism that randomly replaces conditioning signals with null/masked tokens during training, forcing the model to learn unconditional score functions. With probability p (typically 0.1-0.5), the conditioning signal is replaced with a special null token or zero vector, causing the model to predict noise based only on the noisy image and timestep. This simple masking approach enables joint conditional-unconditional training without requiring separate data streams or model branches.

Provides the mechanism for empirically selecting optimal guidance scale values through inference-time experimentation. Practitioners can generate samples at multiple guidance scales (e.g., 1.0, 3.0, 7.5, 15.0) and evaluate quality-diversity tradeoffs without retraining. The guidance scale parameter directly controls the strength of the unconditional score contribution: higher values increase fidelity but reduce diversity, while lower values increase diversity but reduce fidelity.

Implements the application of classifier-free guidance to text-to-image diffusion models, where conditioning signals are text embeddings (from CLIP or other encoders) and guidance enables high-quality image generation from text prompts. The model learns both text-conditioned and unconditional score functions, then uses guidance to interpolate between them at inference time, enabling users to control image quality and diversity through the guidance scale parameter.

Implements the mechanism for learning high-quality unconditional score estimates (∇log p(x_t)) within a conditional diffusion model through conditioning dropout during training. The model learns to predict noise residuals when the conditioning signal is masked, effectively learning the score function of the marginal data distribution. These unconditional scores are then used at inference time to compute guided scores through interpolation with conditional scores.