ComfyUI Node: 🎛️ Krea 2 Conditioning Control

🎛️ Krea 2 Conditioning Control:

The ConditioningKrea2Rebalance node is designed to enhance the conditioning process for Krea 2 and similar multi-layer-tap text-conditioned models. It operates on a stack of 12 Qwen3-VL hidden-state layers, which are packed into a single tensor. This node allows you to reshape the tensor to expose the layer axis and apply a per-layer gain, enabling you to emphasize specific layers that are important for your task. By doing so, it helps in maintaining the overall magnitude constant while applying a global multiplier, ensuring that the masks and pooled outputs remain untouched. This approach is particularly beneficial for recovering fine details without degrading the model's likeness, prompt adherence, and color fidelity. The node offers a refined method for rebalancing the conditioning layers, allowing for detailed control over the model's output by shifting tap ratios and holding the magnitude, thus preserving quality.

🎛️ Krea 2 Conditioning Control Input Parameters:

conditioning

This parameter represents the input conditioning tensor that the node will process. It is a crucial input as it contains the multi-layer text-conditioned data that needs rebalancing. The tensor is reshaped and adjusted based on the other parameters to achieve the desired output.

preset

The preset parameter allows you to select a named per-layer weight profile. Options include "balanced," "detail," "subtle," and "uniform," with "balanced" being the default. This parameter determines the initial weighting applied to each layer, providing a starting point for further adjustments. If set to "custom," the per_layer_weights parameter is used instead.

per_layer_weights

This parameter is a string of comma-separated values representing the gains for each conditioning tap. It is used only when the preset is set to "custom." The default is a balanced set of weights, and it allows for precise control over the emphasis placed on each of the 12 layers.

multiplier

The multiplier is a global gain applied after per-layer weighting. It has a default value of 1.0, with a range from -1000.0 to 1000.0. This parameter controls the overall amplification of the conditioning tensor. Keeping it around 1.0 with renormalize enabled ensures a quality-preserving rebalance, while values greater than 1 can amplify the entire tensor, potentially leading to oversaturation.

renormalize

This boolean parameter, defaulting to True, ensures that the input RMS (Root Mean Square) is maintained after per-layer weighting. It allows tap ratios to change without altering the overall magnitude, preserving the quality of the output. This is the recommended mode for maintaining the integrity of the model's output.

🎛️ Krea 2 Conditioning Control Output Parameters:

conditioning

The output conditioning parameter is the rebalanced tensor that results from the node's processing. It retains the original structure but with adjusted layer weights and global gain, reflecting the changes made by the input parameters. This output is crucial for further processing in the model, as it contains the refined conditioning data that enhances the model's performance and output quality.

🎛️ Krea 2 Conditioning Control Usage Tips:

• Use the preset parameter to quickly apply common weighting profiles that suit your task, such as "detail" for more emphasis on fine details.
• When using per_layer_weights, ensure the values are carefully chosen to avoid overemphasizing or underemphasizing specific layers, which could affect the model's output quality.
• Keep the multiplier close to 1.0 when renormalize is enabled to maintain a balanced output without oversaturation.
• Experiment with different preset options to find the best fit for your specific use case, as different tasks may benefit from different layer emphases.

🎛️ Krea 2 Conditioning Control Common Errors and Solutions:

ValueError: Invalid per_layer_weights format

• Explanation: This error occurs when the per_layer_weights string is not formatted correctly, such as containing non-numeric values or an incorrect number of weights.
• Solution: Ensure that the per_layer_weights string is a comma-separated list of 12 numeric values, with no extra characters or spaces.

AssertionError: Preset weights length mismatch

• Explanation: This error indicates that a preset does not have the correct number of weights, which should be 12 for Krea 2. - Solution: Verify that the preset weights are correctly defined with 12 values. If using a custom preset, ensure it matches the expected format and length.

RuntimeError: Tensor shape mismatch

• Explanation: This error can occur if the input conditioning tensor does not have the expected shape, which can disrupt processing.
• Solution: Check that the input tensor is correctly formatted and matches the expected dimensions for Krea 2, typically (B, seq, 12*D).