ComfyUI Node: H3 RefMod Step Curve

H3 RefMod Step Curve:

The MiniMaxH3RefModStepCurve node is designed to enhance the generation process by applying a per-step reference strength curve during the denoising phase. Unlike the Apply node, which bakes the frame curve into the reference latent once before sampling, this node re-mixes every reference latent at each denoising step. This approach allows for a dynamic adjustment of the reference strength, where early steps with high sigma values establish the global structure and identity, while later steps with low sigma values refine the fine texture. This method ensures that the direction, shape, and value envelope of the reference strength runs over the denoise timeline, providing a more nuanced and controlled generation process. The node is intended to be used between the model loader and the sampler, maintaining the same type of connection (MODEL -> MODEL).

H3 RefMod Step Curve Input Parameters:

model

The model parameter specifies the H3 model to be patched. It is returned unchanged except for the per-step reference mix applied during the denoising process. This parameter is crucial as it determines the base model that will undergo the reference strength adjustments.

curve_direction

The curve_direction parameter defines the envelope of the reference strength over the denoise timeline. Options include concept_at_end, concept_at_start, constant, concept_at_middle, and concept_at_ends. Each option dictates how the reference strength is distributed across the denoising steps, affecting the composition and identity setting in early steps and texture refinement in later steps. The default is concept_at_end.

curve_shape

The curve_shape parameter determines how the per-step strength transitions between its endpoints. Available shapes include ease, linear, sigmoid, tanh, quadratic, cubic, exponential, stair, elastic, bump, and dip. This parameter influences the smoothness or abruptness of the strength changes, with ease being the default for a smooth transition.

retention

The retention parameter is a float value ranging from 0.0 to 1.0, with a default of 1.0. It acts as a master reference strength multiplier, affecting each loader row's strength. Different retention levels, such as 1.0 for fully preserved and 0.0 for no reference, allow for varying degrees of reference influence on the generated output.

curve_value

The curve_value parameter sets the endpoint value for the reference strength, with a default of 1.0. It determines the strength at specific points in the denoising process, such as both endpoints for constant and concept_at_ends, or the start for concept_at_end. This parameter allows for precise control over the reference strength's impact.

graph_preset

The graph_preset parameter allows users to select a predefined graph preset, overriding the curve widgets. Options include none and any available presets. This feature provides a convenient way to apply complex strength curves without manually adjusting each parameter.

H3 RefMod Step Curve Output Parameters:

img

The img output parameter represents the rendered debug grid of the applied strength curve. This visual representation helps users understand how the reference strength is distributed across the denoising timeline, providing insights into the node's effect on the generation process.

H3 RefMod Step Curve Usage Tips:

• Experiment with different curve_direction and curve_shape combinations to achieve the desired balance between global structure and fine texture in your generated outputs.
• Utilize the retention parameter to control the influence of the reference on the final output, adjusting it based on whether you want to preserve the original identity or focus on attribute transfer.
• Save frequently used configurations as graph_preset for quick application in future projects, streamlining your workflow.

H3 RefMod Step Curve Common Errors and Solutions:

Graph preset not found or invalid

• Explanation: This error occurs when the specified graph preset cannot be found or is invalid.
• Solution: Ensure that the graph preset name is correct and that it exists in the list of available presets. If necessary, create or save a new preset with the desired configuration.

Reference strength curve not applied

• Explanation: This issue arises when the reference strength curve is not correctly applied during the denoising process.
• Solution: Verify that all input parameters are correctly set and that the node is properly connected between the model loader and the sampler. Double-check the curve_direction and curve_shape settings to ensure they align with your intended outcome.