I uploaded this workflow to the  public, because of The Flux1-based Upscaler has come together quite nicely.

In testing, when the size of the original image was 1080pixel x 1080pixel, scaling up to 32K size was completed within 12GB of VRAM consumption for almost the entire process.

However, in the case of an RTX4070 12GB class GPU, the operation seems to be quite marginal and OutOfMemory error sometimes occurs, but the program should run completely after repeated runs.

For the function of analyzing the original image, we initially considered utilizing ollama, but it's slow in operation instead of having high performance, so we adopted the form of using WD14-tagger and Florence2.

※27 Jan 2025

Updated with Comfy-WaveSpeed.

Comfy-WaveSpeed contributes significantly to speeding up various Detailers even more than Kasmpler. Therefore, the latest version is equipped with Comfy-WaveSpeed, achieving a substantial increase in speed.