ComfyUI Node: 🎲 LTX Tiled Sampler

🎲 LTX Tiled Sampler:

The LTXTiledSampler node is designed to enhance the sampling process by dividing the latent space into smaller, manageable tiles, allowing for more efficient and refined processing. This node is particularly beneficial for tasks that require light denoising and spatial refinement, as it performs full sampling on each tile at the training-distribution token count and then blends them using a cosine function. This approach ensures that the output maintains high fidelity while optimizing computational resources. Additionally, the node supports audio passthrough, allowing audio data to be processed alongside video data without additional steps, making it a versatile tool for multimedia projects.

🎲 LTX Tiled Sampler Input Parameters:

noise

This parameter represents the noise input required for the sampling process. It is essential for generating variations in the output and is typically used in conjunction with the guider and sampler to produce the desired results.

guider

The guider parameter is responsible for guiding the sampling process. It helps in refining the output by providing additional context or constraints, ensuring that the generated samples align with the intended design or artistic direction.

sampler

The sampler parameter dictates the sampling method used during the process. It plays a crucial role in determining how the latent space is explored and sampled, impacting the diversity and quality of the output.

sigmas

Sigmas are used to control the level of noise and detail in the sampling process. They influence the smoothness and sharpness of the output, allowing for fine-tuning of the results based on the desired artistic effect.

latent_image

This parameter represents the initial latent image that serves as the starting point for the sampling process. It provides the base structure upon which the sampling and refinement are performed.

tile_axis

The tile_axis parameter determines the axis along which the tiling is performed. Options include "auto", "H" (height), and "W" (width), with "auto" automatically selecting the optimal axis based on the dimensions of the input.

n_tiles

This parameter specifies the number of tiles to divide the latent space into. It ranges from 1 to 8, with a default value of 2. Increasing the number of tiles can lead to more detailed refinement but may also increase computational load.

tile_overlap

Tile_overlap defines the amount of overlap between adjacent tiles. It ranges from 0 to a maximum value, with a default of 8. Overlapping tiles can help in blending the edges smoothly, reducing visible seams in the output.

max_size_for_no_tile

This parameter sets the maximum size for which tiling is not applied. It ranges from 8 to 256, with a default of 24. If the axis size is below this threshold, the process will run in a single pass without tiling.

audio_pass

The audio_pass parameter controls how audio data is handled during the sampling process. Options include "passthrough" and "tile_carrying", with "passthrough" being the default. This setting determines whether audio is processed alongside a specific tile or independently.

audio_carrier_tile

This parameter specifies which tile carries the audio data when the "tile_carrying" option is selected for audio_pass. Options include "first", "middle", and "last", with "first" as the default. It determines the tile that will be used for audio sampling.

debug

The debug parameter is a boolean flag that, when enabled, provides additional output and logging information during the sampling process. This can be useful for troubleshooting and understanding the internal workings of the node.

🎲 LTX Tiled Sampler Output Parameters:

output

The output parameter represents the final sampled latent image after the tiling and refinement process. It is the primary result of the node, showcasing the enhanced and denoised version of the input latent image.

denoised_output

Denoised_output provides a version of the output that has undergone additional denoising. This parameter is useful for applications where a cleaner and more polished result is desired, free from noise artifacts.

🎲 LTX Tiled Sampler Usage Tips:

• To achieve optimal results, adjust the n_tiles and tile_overlap parameters based on the complexity and size of your input. More tiles can provide finer detail but may require more processing power.
• Use the tile_axis set to "auto" for most cases, as it automatically selects the best axis for tiling based on the input dimensions, ensuring efficient processing.
• Enable the debug mode if you encounter unexpected results or need to understand the node's behavior better. It provides valuable insights into the sampling process.

🎲 LTX Tiled Sampler Common Errors and Solutions:

"axis_size=<value> ≤ <max_size_for_no_tile>"

• Explanation: This message indicates that the axis size is smaller than or equal to the maximum size allowed for no tiling, resulting in a single-pass process.
• Solution: If tiling is desired, increase the max_size_for_no_tile parameter or adjust the input dimensions to exceed this threshold.

"extracted tensor is <value>D, expected 5D"

• Explanation: This error occurs when the extracted tensor does not have the expected five dimensions, which is necessary for the tiling process.
• Solution: Ensure that the input latent image is correctly formatted and has the required dimensions before processing.