Guiders

Guiders are components in Modular Diffusers that control how the diffusion process is guided during generation. They implement various guidance techniques to improve generation quality and control.

For more details, refer to the Hugging Face Diffusers Documentation.

BaseGuidance

Source

Base class providing the skeleton for implementing guidance techniques.

Methods

cleanup_models

Source

Cleans up the models for the guidance technique after a given batch of data. This method should be overridden in subclasses to implement specific model cleanup logic. It is useful for removing any hooks or other stateful modifications made during prepare_models.

from_pretrained

Source

Instantiate a guider from a pre-defined JSON configuration file in a local directory or Hub repository.

Tip: To use private or gated models, log-in with hf auth login. You can also activate the special “offline-mode” to use this method in a firewalled environment.

Parameters:

  • pretrained_model_name_or_path (str or os.PathLike, optional): Can be either:
    • A string, the model id (for example google/ddpm-celebahq-256) of a pretrained model hosted on the Hub.
    • A path to a directory (for example ./my_model_directory) containing the guider configuration saved with save_pretrained.
  • subfolder (str, optional): The subfolder location of a model file within a larger model repository on the Hub or locally.
  • return_unused_kwargs (bool, optional, defaults to False): Whether kwargs that are not consumed by the Python class should be returned or not.
  • cache_dir (Union[str, os.PathLike], optional): Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used.
  • force_download (bool, optional, defaults to False): Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist.
  • proxies (Dict[str, str], optional): A dictionary of proxy servers to use by protocol or endpoint, for example, {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}. The proxies are used on each request.
  • output_loading_info (bool, optional, defaults to False): Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
  • local_files_only (bool, optional, defaults to False): Whether to only load local model weights and configuration files or not. If set to True, the model won’t be downloaded from the Hub.
  • token (str or bool, optional): The token to use as HTTP bearer authorization for remote files. If True, the token generated from diffusers-cli login (stored in ~/.huggingface) is used.
  • revision (str, optional, defaults to "main"): The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git.

get_state

Source

Returns the current state of the guidance technique as a dictionary. The state variables will be included in the __repr__ method.

Returns:

  • Dict[str, Any]: A dictionary containing the current state variables including:
    • step: Current inference step
    • num_inference_steps: Total number of inference steps
    • timestep: Current timestep tensor
    • count_prepared: Number of times prepare_models has been called
    • enabled: Whether the guidance is enabled
    • num_conditions: Number of conditions

new

Source

Creates a copy of this guider instance, optionally with modified configuration parameters.

Example:

# Create a CFG guider
guider = ClassifierFreeGuidance(guidance_scale=3.5)

# Create an exact copy
same_guider = guider.new()

# Create a copy with different start step, keeping other config the same
new_guider = guider.new(guidance_scale=5)

Parameters:

  • **kwargs: Configuration parameters to override in the new instance. If no kwargs are provided, returns an exact copy with the same configuration.

Returns:

  • A new guider instance with the same (or updated) configuration.

prepare_models

Source

Prepares the models for the guidance technique on a given batch of data. This method should be overridden in subclasses to implement specific model preparation logic.

save_pretrained

Source

Save a guider configuration object to a directory so that it can be reloaded using the from_pretrained class method.

Parameters:

  • save_directory (str or os.PathLike): Directory where the configuration JSON file will be saved (will be created if it does not exist).
  • push_to_hub (bool, optional, defaults to False): Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the repository you want to push to with repo_id (will default to the name of save_directory in your namespace).
  • kwargs (Dict[str, Any], optional): Additional keyword arguments passed along to the push_to_hub method.

ClassifierFreeGuidance

Source

Implements Classifier-Free Guidance (CFG) for diffusion models.

Reference:

CFG improves generation quality and prompt adherence by jointly training models on both conditional and unconditional data, then combining predictions during inference. This allows trading off between quality (high guidance) and diversity (low guidance).

Two CFG Formulations:

  1. Original formulation (from paper): 
    x_pred = x_cond + guidance_scale * (x_cond - x_uncond)

    Moves conditional predictions further from unconditional ones.

  2. Diffusers-native formulation (default, from Imagen paper): 
    x_pred = x_uncond + guidance_scale * (x_cond - x_uncond)

    Moves unconditional predictions toward conditional ones, effectively suppressing negative features (e.g., “bad quality”, “watermarks”). Equivalent in theory but more intuitive.

Use use_original_formulation=True to switch to the original formulation.

Parameters:

  • guidance_scale (float, defaults to 7.5): CFG scale applied by this guider during post-processing. Higher values = stronger prompt conditioning but may reduce quality. Typical range: 1.0-20.0.
  • guidance_rescale (float, defaults to 0.0): Rescaling factor to prevent overexposure from high guidance scales. Based on Common Diffusion Noise Schedules and Sample Steps are Flawed. Range: 0.0 (no rescaling) to 1.0 (full rescaling).
  • use_original_formulation (bool, defaults to False): If True, uses the original CFG formulation from the paper. If False (default), uses the diffusers-native formulation from the Imagen paper.
  • start (float, defaults to 0.0): Fraction of denoising steps (0.0-1.0) after which CFG starts. Use > 0.0 to disable CFG in early denoising steps.
  • stop (float, defaults to 1.0): Fraction of denoising steps (0.0-1.0) after which CFG stops. Use < 1.0 to disable CFG in late denoising steps.
  • enabled (bool, defaults to True): Whether CFG is enabled. Set to False to disable CFG entirely (uses only conditional predictions).

ClassifierFreeZeroStarGuidance

Source

Classifier-free Zero* (CFG-Zero*):

This is an implementation of the Classifier-Free Zero* guidance technique, which is a variant of classifier-free guidance. It proposes zero initialization of the noise predictions for the first few steps of the diffusion process, and also introduces an optimal rescaling factor for the noise predictions, which can help in improving the quality of generated images.

The authors of the paper suggest setting zero initialization in the first 4% of the inference steps.

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance. Higher values result in stronger conditioning on the text prompt, while lower values allow for more freedom in generation. Higher values may lead to saturation and deterioration of image quality.
  • zero_init_steps (int, defaults to 1): The number of inference steps for which the noise predictions are zeroed out (see Section 4.2).
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions. This is used to improve image quality and fix overexposure. Based on Section 3.4 from Common Diffusion Noise Schedules and Sample Steps are Flawed.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default, we use the diffusers-native implementation that has been in the codebase for a long time.
  • start (float, defaults to 0.01): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which guidance stops.

SkipLayerGuidance

Source

Skip Layer Guidance (SLG): Stability AI SD3.5

Spatio-Temporal Guidance (STG):

SLG was introduced by StabilityAI for improving structure and anatomy coherence in generated images. It works by skipping the forward pass of specified transformer blocks during the denoising process on an additional conditional batch of data, apart from the conditional and unconditional batches already used in CFG, and then scaling and shifting the CFG predictions based on the difference between conditional without skipping and conditional with skipping predictions.

The intuition behind SLG can be thought of as moving the CFG predicted distribution estimates further away from worse versions of the conditional distribution estimates (because skipping layers is equivalent to using a worse version of the model for the conditional prediction).

STG is an improvement and follow-up work combining ideas from SLG, PAG and similar techniques for improving generation quality in video diffusion models.

Additional reading:

The values for skip_layer_guidance_scale, skip_layer_guidance_start, and skip_layer_guidance_stop are defaulted to the recommendations by StabilityAI for Stable Diffusion 3.5 Medium.

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • skip_layer_guidance_scale (float, defaults to 2.8): The scale parameter for skip layer guidance. Anatomy and structure coherence may improve with higher values, but it may also lead to overexposure and saturation.
  • skip_layer_guidance_start (float, defaults to 0.01): The fraction of the total number of denoising steps after which skip layer guidance starts.
  • skip_layer_guidance_stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which skip layer guidance stops.
  • skip_layer_guidance_layers (int or List[int], optional): The layer indices to apply skip layer guidance to. Can be a single integer or a list of integers. If not provided, skip_layer_config must be provided. The recommended values are [7, 8, 9] for Stable Diffusion 3.5 Medium.
  • skip_layer_config (LayerSkipConfig or List[LayerSkipConfig], optional): The configuration for the skip layer guidance. Can be a single LayerSkipConfig or a list of LayerSkipConfig. If not provided, skip_layer_guidance_layers must be provided.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.01): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which guidance stops.

SmoothedEnergyGuidance

Source

Smoothed Energy Guidance (SEG):

SEG is only supported as an experimental prototype feature for now, so the implementation may be modified in the future without warning or guarantee of reproducibility. This implementation assumes:

  • Generated images are square (height == width)
  • The model does not combine different modalities together (e.g., text and image latent streams are not combined together such as Flux)

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • seg_guidance_scale (float, defaults to 3.0): The scale parameter for smoothed energy guidance. Anatomy and structure coherence may improve with higher values, but it may also lead to overexposure and saturation.
  • seg_blur_sigma (float, defaults to 9999999.0): The amount by which we blur the attention weights. Setting this value greater than 9999.0 results in infinite blur, which means uniform queries. Controlling it exponentially is empirically effective.
  • seg_blur_threshold_inf (float, defaults to 9999.0): The threshold above which the blur is considered infinite.
  • seg_guidance_start (float, defaults to 0.0): The fraction of the total number of denoising steps after which smoothed energy guidance starts.
  • seg_guidance_stop (float, defaults to 1.0): The fraction of the total number of denoising steps after which smoothed energy guidance stops.
  • seg_guidance_layers (int or List[int], optional): The layer indices to apply smoothed energy guidance to. Can be a single integer or a list of integers. If not provided, seg_guidance_config must be provided. The recommended values are [7, 8, 9] for Stable Diffusion 3.5 Medium.
  • seg_guidance_config (SmoothedEnergyGuidanceConfig or List[SmoothedEnergyGuidanceConfig], optional): The configuration for the smoothed energy layer guidance.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.01): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which guidance stops.

PerturbedAttentionGuidance

Source

Perturbed Attention Guidance (PAG):

The intuition behind PAG can be thought of as moving the CFG predicted distribution estimates further away from worse versions of the conditional distribution estimates. PAG was one of the first techniques to introduce the idea of using a worse version of the trained model for better guiding itself in the denoising process. It perturbs the attention scores of the latent stream by replacing the score matrix with an identity matrix for selectively chosen layers.

Additional reading:

PAG is implemented with similar implementation to SkipLayerGuidance due to overlap in the configuration parameters and implementation details.

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • perturbed_guidance_scale (float, defaults to 2.8): The scale parameter for perturbed attention guidance.
  • perturbed_guidance_start (float, defaults to 0.01): The fraction of the total number of denoising steps after which perturbed attention guidance starts.
  • perturbed_guidance_stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which perturbed attention guidance stops.
  • perturbed_guidance_layers (int or List[int], optional): The layer indices to apply perturbed attention guidance to. Can be a single integer or a list of integers. If not provided, perturbed_guidance_config must be provided.
  • perturbed_guidance_config (LayerSkipConfig or List[LayerSkipConfig], optional): The configuration for the perturbed attention guidance.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.01): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 0.2): The fraction of the total number of denoising steps after which guidance stops.

AdaptiveProjectedGuidance

Source

Adaptive Projected Guidance (APG): 2410.02416

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • adaptive_projected_guidance_momentum (float, defaults to None): The momentum parameter for the adaptive projected guidance. Disabled if set to None.
  • adaptive_projected_guidance_rescale (float, defaults to 15.0): The rescale factor applied to the noise predictions. This is used to improve image quality.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.0): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 1.0): The fraction of the total number of denoising steps after which guidance stops.

AutoGuidance

Source

AutoGuidance:

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • auto_guidance_layers (int or List[int], optional): The layer indices to apply skip layer guidance to. Can be a single integer or a list of integers. If not provided, skip_layer_config must be provided.
  • auto_guidance_config (LayerSkipConfig or List[LayerSkipConfig], optional): The configuration for the skip layer guidance.
  • dropout (float, optional): The dropout probability for autoguidance on the enabled skip layers (either with auto_guidance_layers or auto_guidance_config). If not provided, the dropout probability will be set to 1.0.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.0): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 1.0): The fraction of the total number of denoising steps after which guidance stops.

TangentialClassifierFreeGuidance

Source

Tangential Classifier Free Guidance (TCFG):

Parameters:

  • guidance_scale (float, defaults to 7.5): The scale parameter for classifier-free guidance.
  • guidance_rescale (float, defaults to 0.0): The rescale factor applied to the noise predictions.
  • use_original_formulation (bool, defaults to False): Whether to use the original formulation of classifier-free guidance as proposed in the paper.
  • start (float, defaults to 0.0): The fraction of the total number of denoising steps after which guidance starts.
  • stop (float, defaults to 1.0): The fraction of the total number of denoising steps after which guidance stops.