ComfyUI Node: LTXV FDG Parameters

LTXV FDG Parameters:

The FDGParameters node is designed to configure Frequency-Decoupled Guidance (FDG) for ComfyUI, particularly in the context of LTXV video generation. This node leverages the principles outlined in the paper "Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales" to enhance the quality of generated content by decoupling guidance into frequency components. By adjusting parameters such as low and high-frequency guidance scales, users can achieve a balance between global structure and fine detail, optimizing the fidelity of the output. The node's primary goal is to provide a flexible and powerful tool for AI artists to fine-tune their generative models, ensuring high-quality results even at low CFG scales.

LTXV FDG Parameters Input Parameters:

modality

This parameter specifies the type of content being generated, such as video. It is crucial for determining the appropriate guidance settings for different media types. The default value is VIDEO.

cfg

The cfg parameter represents the base scale for Classifier-Free Guidance (CFG), which is used when FDG is disabled or for other guidance types. It influences the overall strength of the guidance applied to the generation process. The default value is 1.5, with a range from 0.0 to 100.0.

fdg_enabled

This boolean parameter determines whether Frequency-Decoupled Guidance is active. Enabling FDG allows for more nuanced control over the frequency components of the guidance, enhancing the quality of the output. The default setting is True.

w_low

The w_low parameter controls the low-frequency guidance scale, which affects the global structure of the generated content. Lower values promote diversity, while higher values improve condition alignment. The default value is 1.5, with a range from 0.0 to 100.0.

w_high

This parameter sets the high-frequency guidance scale, impacting the fine details of the output. Adjusting this value allows for control over the sharpness and clarity of the generated content. The default value is 4.0, with a range from 0.0 to 100.0.

frequency_levels

The frequency_levels parameter determines the number of frequency levels used in the guidance process. A default value of 2 provides a good balance between detail and computational efficiency.

decomposition_method

This parameter specifies the method used for frequency decomposition, with options such as "laplacian". It influences how the frequency components are separated and processed.

use_projection

A boolean parameter that, when enabled, applies an optional APG projection to improve color accuracy in the generated content. The default setting is False.

projection_weight

The projection_weight parameter adjusts the strength of the projection applied when use_projection is enabled. It helps fine-tune the color correction process. The default value is 1.0.

stg

This parameter controls the scale of the standard guidance applied alongside FDG. It affects the overall influence of the guidance on the generation process. The default value is 0.0.

perturb_attn

A boolean parameter that, when enabled, allows for attention perturbation, which can enhance the diversity of the generated content. The default setting is True.

rescale

The rescale parameter adjusts the scale of the rescaling applied during the guidance process, influencing the overall balance of the output. The default value is 0.0.

modality_scale

This parameter sets the scale for modality-specific guidance, allowing for fine-tuning based on the type of content being generated. The default value is 1.0.

skip_step

The skip_step parameter determines the number of steps to skip during the guidance process, which can affect the speed and quality of the generation. The default value is 0.

cross_attn

A boolean parameter that, when enabled, applies cross-attention during the guidance process, potentially enhancing the coherence of the generated content. The default setting is True.

LTXV FDG Parameters Output Parameters:

GUIDER_PARAMETERS

The output parameter GUIDER_PARAMETERS contains the configured guidance parameters, encapsulating all the settings specified in the input. This output is crucial for applying the configured guidance to the generative model, ensuring that the desired effects on frequency components and overall quality are achieved.

LTXV FDG Parameters Usage Tips:

• For high-quality results at low CFG scales, consider setting w_low between 1.0 and 1.5, and w_high between 3.0 and 7.0.
• To achieve the best quality, adjust w_low to between 3.0 and 5.0, and w_high to between 7.0 and 12.0.
• Use the frequency_levels parameter to balance detail and computational efficiency, with a default value of 2 recommended for most cases.

LTXV FDG Parameters Common Errors and Solutions:

Modality {modality} already exists in parameters

• Explanation: This error occurs when the specified modality is already present in the parameters dictionary, indicating a duplication.
• Solution: Ensure that each modality is unique within the parameters dictionary. If necessary, modify the modality name or remove the existing entry before adding a new one.