What this workflow does

Starting with a 300x300 pixel image of a face upscale it to 1536x1536 whilst preserving identity.

How to use this workflow

Load an image of a face into both Load Image nodes.

Describe the image in the top prompt box and the style in the bottom prompt box.

Optionally load a different image to be upscaled with a similar face shape into the lower Load Image node. If it is already at the target size, you can skip the upscaler.

Tips about this workflow

This workflow is set to 1536x1536 as this is the largest image that can be decoded on GPU with 6GB of VRAM. If you have 6GB of VRAM or less use the --fp8_e4m3fn-text-enc --fp8_e4m3fn-unet flags to start ComfyUI or you may get memory errors. In fp8 mode this workflow may only run once before ComfyUI requires a restart due to a KSampler error.

For non-western faces currently OpenDalleV1.1 is the best checkpoint for preserving the face shape during upscale.

You could copy these nodes to use after your favourite 512x512 people image generation pathway (e.g. SDXL-turbo-DPO).

The example images are taken from the public domain and were cropped to a square.