What this workflow does

Using only a single seed and prompt generate 9 related images that are perfect for merging the best bits into one.

This workflow generates a slightly oversize blank canvas and then crops it back to each corner to create 4 new empty latents which all share most of the same noise - so will make very similar images. These 4 latents are then blended 50:50 with each other to give the grid of 9.

How to use this workflow

The settings have been created for SD-Turbo or SDXL-Turbo.

Set the seed and the number of steps and then start prompting.

To save disc memory I've left off saving images.

Tips about this workflow

The default variant is to pad by 8 pixels. You can increase this to 16 or 24 but then the images will have less in common. Due to bug with ComfyLiterals you also need to manually increase the canvas size by the amount of the padding.

If you increase the batch size then only the 4 corner images will produce multiple images.

Adding a LORA such as contrast offset is optional (simply change the connections in the top left).