ComfyUI Node: 🔍 LTX Latent Upsampler (Tiled)

🔍 LTX Latent Upsampler (Tiled):

LTXVLatentUpsamplerTiled is a specialized node designed to enhance the upscaling process of latent images by addressing common issues such as color shifts and distortion, particularly in scenarios involving extreme aspect ratios. This node serves as a spatially-tiled drop-in replacement for the LTXVLatentUpsampler, utilizing a method that processes the upscale in spatial tiles with cosine-windowed overlap blending. This approach ensures that the output remains virtually identical to non-tiled methods in safe-aspect cases while eliminating instability in extreme cases. By dividing the image into tiles, the node effectively manages the upscale process, maintaining the integrity of the image's colors and details, and providing a more stable and reliable output.

🔍 LTX Latent Upsampler (Tiled) Input Parameters:

samples

This parameter represents the latent samples that need to be upscaled. It is a required input and serves as the primary data that the node processes to produce an upscaled output.

upscale_model

The upscale_model parameter specifies the model used for the upscaling process. It is a required input that determines the method and quality of the upscaling, directly impacting the final output's resolution and detail.

vae

This parameter refers to the Variational Autoencoder (VAE) used in the process. It is a required input that plays a crucial role in the normalization and reconstruction of the upscaled latent samples, ensuring the output maintains high fidelity to the original data.

tile_size

The tile_size parameter defines the size of each tile used in the upscaling process. It is an optional integer parameter with a default value of 24, a minimum of 8, and a maximum of 128. Adjusting this value affects the granularity of the tiling process, with smaller tiles potentially offering more precise control over the upscaling but requiring more computational resources.

overlap

This parameter specifies the amount of overlap between adjacent tiles during the upscaling process. It is an optional integer parameter with a default value of 8, a minimum of 2, and a maximum of 32. The overlap helps in blending the tiles smoothly, reducing visible seams and ensuring a cohesive final image.

max_size_for_no_tile

The max_size_for_no_tile parameter determines the maximum size at which the node will process the image without tiling. It is an optional integer parameter with a default value of 32, a minimum of 8, and a maximum of 256. Images smaller than this threshold are processed without tiling, which can be more efficient and produce faster results.

rotate_for_landscape

This boolean parameter, with a default value of False, indicates whether the image should be rotated for landscape orientation before processing. Enabling this option can optimize the tiling process for images with a landscape aspect ratio, potentially improving the quality of the upscaled output.

debug

The debug parameter is a boolean option with a default value of False. When enabled, it provides additional output and logging information during the upscaling process, which can be useful for troubleshooting and understanding the node's behavior.

🔍 LTX Latent Upsampler (Tiled) Output Parameters:

LATENT

The output parameter is a latent representation of the upscaled image. This output retains the enhanced resolution and detail achieved through the tiling and upscaling process, providing a high-quality latent image that can be further processed or decoded into a visible image format.

🔍 LTX Latent Upsampler (Tiled) Usage Tips:

• To achieve optimal results with extreme aspect ratios, consider adjusting the tile_size and overlap parameters to find a balance between processing time and output quality.
• Use the rotate_for_landscape option for images with a landscape orientation to ensure the tiling process is optimized for the image's aspect ratio.
• Enable the debug mode if you encounter unexpected results or need to understand the node's processing steps in more detail.

🔍 LTX Latent Upsampler (Tiled) Common Errors and Solutions:

Non-uniform scale detected

• Explanation: This warning indicates that the detected upscale ratio is not uniform across dimensions, which might lead to distortion.
• Solution: Ensure that the upscale model is configured correctly and consider adjusting the tile_size and overlap parameters to achieve a more uniform scale.

Tiling triggered unexpectedly

• Explanation: This message appears when the image dimensions exceed the max_size_for_no_tile threshold, triggering the tiling process.
• Solution: If tiling is not desired, increase the max_size_for_no_tile parameter to accommodate larger images without tiling, or adjust the image size accordingly.