This repository includes a custom node for ComfyUI for upscaling the latents quickly using a small neural network without needing to decode and encode with VAE. The node can be found in "Add Node -> latent -> NNLatentUpscale".

This node is meant to be used in a workflow where the initial image is generated in lower resolution, the latent is upscaled and the upscaled latent is fed back into the stable diffusion u-net for low noise diffusion pass (high-res fix).

Compared to VAE decode -> upscale -> encode, the neural net latent upscale is about 20 - 50 times faster depending on the image resolution with minimal quality loss. Compared to direct linear interpolation of the latent the neural net upscale is slower but has much better quality. Direct latent interpolation usually has very large artifacts.

Clone this repository in ComfyUI custom_nodes directory with: git clone https://github.com/Ttl/ComfyUi_NNLatentUpscale.git.

Dataset: COCO 2017 validation images center cropped to 256x256 resolution. The comparison image is linear upscale of the input image. All tests are done with fp32 precision and batch size 4.

VAE Upscale: VAE decode -> Linear interpolation -> Encode.

NN Upscale: Neural network upscale (This repository).

Latent Upscale: Linear interpolation of latent.

SDXL, 2x upscale:

SDXL, 1.5x upscale:

SD 1.5, 2x upscale:

SD 1.5, 1.5x upscale: