KSampler (Spectrum + SPD / SPEED):
The SpectrumSPDKSampler is a sophisticated node designed to enhance the sampling process by integrating the SPEED sampler with spectral expansion techniques. This node is particularly beneficial for AI artists looking to achieve high-quality image generation through a multi-resolution progressive diffusion approach. It operates by initially processing a low-resolution SPD prefix, which is uncached, and then applying spectral expansion at a specified transition point to achieve full-resolution outputs. The node is optimized for use with the Euler sampler, ensuring that the geometry of training and inference remains aligned without requiring manual adjustments. This alignment is crucial for maintaining the integrity of the generated images, especially when using SPD-trained LoRA models. By automatically configuring the SPEED sampler based on embedded schedule metadata, the SpectrumSPDKSampler simplifies the workflow for users, allowing them to focus on creative aspects rather than technical configurations.
KSampler (Spectrum + SPD / SPEED) Input Parameters:
model
The model parameter refers to the AI model that will be used for the sampling process. It is essential for defining the architecture and capabilities of the image generation process. There are no specific minimum or maximum values, as this parameter is typically a pre-trained model object.
steps
This parameter determines the number of steps the sampler will take during the diffusion process. More steps generally lead to higher quality outputs but require more computational resources. The default value is often set to 30, with no strict minimum or maximum, but practical limits depend on available computational power.
window_size
Window size defines the size of the window used during the spectral expansion process. It affects the granularity of the expansion and can impact the final image resolution. The default value is 2.0, with flexibility depending on the desired output quality.
flex_window
Flex window is a parameter that allows for slight adjustments in the window size to accommodate variations in the image generation process. It is set to a default of 0.25, providing a balance between flexibility and stability.
warmup_steps
Warmup steps specify the initial steps taken to stabilize the model before full-resolution sampling begins. This parameter helps in achieving smoother transitions and is typically set to 6.
tail_actual_steps
This parameter defines the number of steps taken at the tail end of the sampling process, ensuring that the final image details are refined. The default is 3 steps, which can be adjusted based on the desired level of detail.
blend_w
Blend weight is used to control the blending of different spectral components during the expansion process. A default value of 0.3 is used to maintain a harmonious balance between components.
cheby_degree
Chebyshev degree determines the degree of the polynomial used in the spectral expansion, affecting the smoothness and accuracy of the expansion. The default is set to 3, which is a common choice for balancing complexity and performance.
ridge_lambda
Ridge lambda is a regularization parameter used to prevent overfitting during the spectral expansion. It is set to a default of 0.1, providing a moderate level of regularization.
history_size
History size defines the number of past steps considered during the sampling process, which can influence the stability and consistency of the output. The default value is 100, allowing for a comprehensive history without excessive computational overhead.
enabled
This boolean parameter determines whether the SpectrumSPDKSampler is active. When set to true, the node performs its functions; otherwise, it is bypassed.
one_sampler_only
This parameter, when set to true, restricts the process to using only one sampler, which can simplify the workflow and reduce computational complexity.
verbose
Verbose is a boolean parameter that, when enabled, provides detailed logging information during the sampling process. This can be useful for debugging and understanding the internal workings of the node.
KSampler (Spectrum + SPD / SPEED) Output Parameters:
patched
The patched output is a modified version of the input model, incorporating the spectral expansion and SPD techniques. This output is crucial for generating high-quality images with enhanced resolution and detail.
KSampler (Spectrum + SPD / SPEED) Usage Tips:
- Ensure that the model parameter is correctly set to a pre-trained model compatible with the SpectrumSPDKSampler to achieve optimal results.
- Adjust the steps parameter based on the desired quality and available computational resources; more steps generally yield better results but require more processing time.
- Utilize the verbose parameter for detailed insights into the sampling process, which can aid in troubleshooting and optimizing the node's performance.
KSampler (Spectrum + SPD / SPEED) Common Errors and Solutions:
"SPEED/SPD re-spaces σ mid-loop and is Euler-only; ignoring requested sampler"
- Explanation: This warning indicates that the node is designed to work exclusively with the Euler sampler, and any other sampler specified will be ignored.
- Solution: Ensure that the Euler sampler is selected when using the SpectrumSPDKSampler to avoid this warning and ensure proper functionality.
"No schedule metadata found in selected LoRA"
- Explanation: This error occurs when the selected LoRA model does not contain the necessary schedule metadata for automatic configuration.
- Solution: Use a LoRA model that includes schedule metadata, or manually configure the SPD scale and sigma parameters to align with the model's requirements.