Producing Detailed Enlargements with Tiled Image-to-Image Upscaling

A stronger upscaling workflow combines an upscaler model with low-denoise image-to-image processing. Rather than enlarging the entire image in one pass, Ultimate SD Upscale divides it into tiles, regenerates detail within each tile, and stitches the tiles back together.

This method can produce substantially finer hair, facial features, armor, and lighting than a direct 4x model upscale. The improvement comes with a trade-off: because image-to-image is generating content, small details can change.

Ultimate SD Upscale is installed as a custom node through Manager and requires a ComfyUI restart. For the upscaling stage, it replaces the ordinary KSampler.

The node receives:

Model from the checkpoint.

Positive conditioning.

Negative conditioning.

VAE from the checkpoint.

A loaded input image.

An upscale model such as 4x-UltraSharp.

If a 512 by 512 image is enlarged by two, the process uses a two-by-two grid of four sections. Each section is processed with image-to-image. The four regenerated sections are then stitched into one 1024 by 1024 image.

The critical control is denoise. A low value, such as 0.2 or even 0.15, retains most of the original tile while allowing the model to generate finer detail. This is why the result can be sharper than a direct upscaler without becoming a completely new image.

Tile width and tile height define the dimensions used for each tile. The starting recommendation is to match them to the dimensions of the original input. For a 512 by 512 source, the tile width and height are set to 512.

Mask blur and tile padding affect how the processed tiles blend when they are reassembled. The default values are retained unless visible lines appear between tiles. In that failure case, those settings can be adjusted to improve the seams.

A fourfold enlargement can be produced in one 4x pass or by chaining two 2x Ultimate SD Upscale nodes. The first converts 512 by 512 to 1024 by 1024. The second receives that result and enlarges it again. Because its input is now 1024 by 1024, its tile width and height are updated to 1024.

Ctrl+Shift+V is useful here: duplicating the first Ultimate SD Upscale node with preserved connections automatically reconnects the shared model, prompts, VAE, and upscaler model. The original image connection is then removed and replaced with the output of the first 2x stage.

A direct 4x upscaler and a two-stage tiled workflow can produce images with the same final resolution while differing sharply in useful detail. Resolution alone therefore does not measure output quality.

The tiled method also makes large-image processing manageable by operating on sections. More importantly, each section receives fresh image-to-image detail generation rather than only geometric enlargement.

Build a tiled upscaling graph:

1. Install Ultimate SD Upscale through Custom Nodes Manager and restart ComfyUI.
2. Load a previously generated 512 by 512 image.
3. Add Ultimate SD Upscale.
4. Connect the checkpoint model, positive prompt, negative prompt, and VAE.
5. Add an Upscale Model Loader and select 4x-UltraSharp.
6. Connect the loaded image.
7. Set the enlargement factor to 2.
8. Set denoise to 0.2 as a starting point; reduce it to approximately 0.15 when stronger preservation is required.
9. Set tile width and tile height to 512.
10. Leave mask blur and tile padding at their defaults unless seams appear.
11. Connect the output to Preview Image and queue the workflow.

To reach a 4x enlargement through two stages:

1. Duplicate the Ultimate SD Upscale node with Ctrl+Shift+V.
2. Remove the duplicate's original image input.
3. Connect the first node's 2x output to the second node's image input.
4. Set the second node to another 2x enlargement.
5. Change its tile width and height to 1024 because its input is now 1024 by 1024.
6. Preview the second output and compare it with a direct 4x model upscale.

The tiled result can produce more convincing facial hair, eyebrows, eyes, nose lighting, crown detail, and armor detail than the direct 4x result.

Ultimate SD Upscale does not preserve every pixel. A low denoise value still allows the model to alter details. For example, the subject's gaze can shift slightly between the direct upscale and the tiled result.

When exact preservation of the original image is the priority, a conventional model upscaler is the safer choice. When sharper and more generated detail is more important than exact fidelity, tiled image-to-image upscaling is the stronger option.

Tile seams can appear. Mask blur and padding address that failure mode; start with their default values and adjust them when seams remain visible.

Two-stage upscaling creates more graph complexity and computation than a single direct pass. The additional work is justified only when the detail improvement matters.

Tiled upscaling improves detail by applying low-denoise image-to-image processing to sections of the image. Use it when added detail is worth small changes to the original; use a direct upscaler when exact preservation matters more.