Classifier-Free Guidance

At its core, CFG works by linearly interpolating between a conditional prediction (guided by a prompt like "a robot picking up a blue block") and an unconditional prediction (guided by a null prompt). The formula is: guided_prediction = unconditional_prediction + guidance_scale * (conditional_prediction - unconditional_prediction). This amplifies the signal in the direction of the condition.

The guidance scale (often denoted as w or s) is a critical hyperparameter controlling the strength of conditioning.

• Scale = 1.0: Yields the standard conditional model output.
• Scale > 1.0: Increases adherence to the condition, improving sample quality and prompt alignment but potentially reducing diversity.
• Very High Scale (>7.0): Can lead to over-saturated, repetitive, or artifact-ridden outputs. Optimal scale is typically found between 5.0 and 15.0 for diffusion models.

A key advantage over earlier methods like Classifier Guidance is that CFG does not require training a separate noise-aware classifier model. This simplifies the training pipeline, reduces memory overhead during sampling, and avoids the challenge of training a robust classifier on noisy intermediate data. The model learns both conditional and unconditional distributions within a single network.

The model is trained to perform both tasks by randomly dropping the conditioning signal (e.g., setting the prompt to null) during a fraction of training steps (typically 10-20%). This teaches the same network to generate data both with and without guidance, enabling the blending trick during inference. The dropout probability is a key training hyperparameter.

In Vision-Language-Action (VLA) models, CFG is used to control action generation. For a language command like "Place the cup on the shelf," the conditional prediction generates actions aligned with the command, while the unconditional prediction represents a generic "do nothing" or neutral trajectory. The guidance scale sharpens the model's focus on the instructed task, improving the precision of the resulting motor commands.

CFG introduces a well-known trade-off:

• High Fidelity/Alignment: Achieved with a higher guidance scale, leading to samples that closely match the condition.
• Sample Diversity: Decreases as the guidance scale increases, as the unconditional "exploration" direction is suppressed. This is a fundamental limitation of the technique, often requiring tuning for the specific use case.

The precursor to CFG, classifier guidance uses a separately trained classifier model to compute gradients during the sampling process of a diffusion model. These gradients steer the denoising process toward outputs that maximize the classifier's score for a given condition (e.g., a class label).

• Mechanism: Requires training both a generative diffusion model and a classifier on noisy data.
• Drawback: Introduces complexity and can lead to adversarial artifacts, as the guidance signal is not natively aligned with the model's own training distribution.

The broad paradigm of training a model to produce outputs contingent on some input signal or context. Classifier-Free Guidance is a sampling-time technique for conditional models.

• Training Objective: Models are trained on paired data (e.g., image-text, state-action-goal).
• Conditioning Signals: Can be discrete (class labels, text), continuous (embedding vectors), or multimodal (image + language).
• Contrast with Unconditional Generation: The model learns the joint distribution P(output, condition) instead of just P(output).

A scalar hyperparameter (often denoted as w, s, or γ) that controls the strength of the conditioning signal in CFG. It is the critical knob for trading off sample quality against adherence to the condition.

• Formula: output = unconditional_output + guidance_scale * (conditional_output - unconditional_output).
• Typical Range: Values from 1.0 (weak) to 7.5+ (strong).
• Effect: Higher values amplify the condition but can reduce diversity and sometimes introduce saturation or artifacts.

A class of generative models that learn to reverse a gradual noising process. CFG was pioneered for and is most commonly applied to denoising diffusion probabilistic models (DDPMs) and score-based generative models.

• Core Process: Data is corrupted with noise over many steps, and a neural network is trained to predict the noise (or the clean data) at each step.
• Sampling: Generation occurs by iteratively denoising pure noise.
• Relevance for CFG: The iterative nature of diffusion sampling provides a natural interface for injecting guidance at each denoising step.

A foundational technique for controlling the randomness of a model's output distribution, commonly used in language and autoregressive action models. While CFG operates on the mean of a distribution, temperature scales the variance.

• Mechanism: The logits of a categorical distribution are divided by a temperature parameter T before applying softmax.
• Effect: T → 0 makes outputs deterministic (argmax). T → 1 uses the native distribution. T > 1 increases randomness and diversity.
• Complementarity: Can be used in conjunction with CFG for action decoding, where CFG steers toward a goal and temperature controls exploration.

A deterministic search algorithm used in autoregressive decoding (common in action sequence generation) to find high-probability output sequences. It contrasts with stochastic methods like CFG or temperature sampling.

• Process: Maintains a fixed number (beam width) of most likely partial sequences at each generation step.
• Use Case: Prioritizes likelihood and coherence over diversity. Often used for tasks requiring precise, verifiable sequences.
• Trade-off: Computationally more expensive than simple sampling but can yield higher-quality, more predictable plans for robotic control.