ComfyUI Node: Hildegard References Split

Hildegard References Split:

The Hildegard References Split node is designed to enhance the process of image upscaling by dividing an upscaled image into smaller tiles and generating three distinct types of reference latents in a single operation. This node is particularly beneficial for AI artists looking to refine textures and spatial details in their artwork. It creates a tile_latent, which is a VAE-encoded crop of each tile that can optionally reduce high-frequency details, allowing the model to creatively enhance textures. Additionally, it generates a position_latent, a VAE-encoded 3x3 position map that places the current tile at the center with its eight neighboring tiles surrounding it, providing spatial context. Finally, it produces a global_latent, a VAE-encoded thumbnail of the entire upscaled image, offering a holistic view. This comprehensive approach allows for detailed texture refinement and spatial coherence in the final artwork.

Hildegard References Split Input Parameters:

image

This parameter accepts the upscaled image from the Hildegard Plan. It serves as the primary input for the node, providing the base image that will be split into tiles and processed to generate the reference latents.

dac_data

This input carries the tile grid, overlaps, and ordering information from the Hildegard Plan. It ensures that the node processes the tiles in the correct sequence and respects the spatial arrangement defined by the plan.

vae

The VAE (Variational Autoencoder) used to encode the three reference image types into latents. For the FLUX.2 Klein model, the Klein VAE is recommended. This parameter is crucial for transforming the image data into a latent space that the model can work with.

tile

This integer parameter determines which tiles to process. A value of 0 processes all tiles in the grid, resulting in outputs that are lists of entries for each tile. A value of 1 or higher processes only the specified tile, producing a list with a single entry. The index follows the linear or spiral ordering chosen by the Hildegard Plan. The default value is 0, with a minimum of 0 and a step of 1.

cell_size

This integer parameter defines the side length in pixels of each cell in the 3x3 position map. The final position image is cell_size x 3 on each axis. A larger cell size provides more spatial context detail and results in a larger position latent. The default value is 512, with a minimum of 64, a maximum of 2048, and a step of 16. This setting is crucial for balancing detail and computational efficiency.

Hildegard References Split Output Parameters:

tile_latent

The tile_latent is a VAE-encoded crop of each tile, which may include optional high-frequency reduction. This output is essential for allowing the model to invent and refine textures within each tile, enhancing the overall quality of the upscaled image.

position_latent

The position_latent is a VAE-encoded 3x3 position map that places the current tile at the center with its eight neighboring tiles surrounding it. This output provides spatial context, ensuring that the model maintains coherence across the entire image.

global_latent

The global_latent is a VAE-encoded thumbnail of the entire upscaled image. This output offers a holistic view of the image, allowing the model to consider the overall composition and balance when refining individual tiles.

Hildegard References Split Usage Tips:

• To achieve the best results, ensure that the cell_size is set according to the level of detail you desire. Larger values provide more context but require more computational resources.
• Use the tile parameter to focus on specific areas of the image that need refinement, which can save processing time and resources.
• Experiment with different VAE models to see how they affect the texture and detail of the output latents.

Hildegard References Split Common Errors and Solutions:

Error: "Invalid tile index"

• Explanation: This error occurs when the tile parameter is set to a value that exceeds the number of available tiles.
• Solution: Ensure that the tile parameter is within the valid range, starting from 0 up to the total number of tiles minus one.

Error: "VAE model not found"

• Explanation: This error indicates that the specified VAE model is not available or incorrectly configured.
• Solution: Verify that the correct VAE model is installed and properly referenced in the node settings. Ensure compatibility with the image data being processed.