What is krea2-remote-encoder?

krea2-remote-encoder is a setup that lets you run one part of an AI image generation pipeline on a second computer, freeing up memory on the main computer that does the heavy image generation work. Specifically, it moves the text encoder for Krea 2, an AI image generator built on ComfyUI, onto a separate machine over your local network, while the image generating model itself stays on the computer with the graphics card. Only a small piece of data describing what the text means gets sent across the network between the two machines. The README is unusually direct about what this actually gains you: about five seconds saved per image generated, which comes entirely from skipping the step where the main computer has to load and then remove the text encoder from memory each time. The speed at which each image generation step runs stays exactly the same either way, so on a single image the difference is barely noticeable. The real benefit shows up when your graphics card's memory is nearly full, since the encoder and the image generating model together do not fit in 16 gigabytes of memory, which normally forces the main computer to shuffle parts of the model in and out of memory. Moving the encoder to a second machine avoids that shuffling entirely. The author notes that an earlier version of this document claimed a much bigger speedup, but that this was not supported once properly measured, and corrects the record here. The project includes separate setup guides for the encoder machine and the graphics card machine, a list of six specific problems people are likely to run into with their exact error messages, a document explaining where the data goes at each step, a small set of tests, and a minimal example workflow. It relies on a separate, smaller open source project called comfy_remote_run to send data between the two machines. The author is upfront that this is a do it yourself setup rather than a polished tool, and that some debugging should be expected. It is intended for people who are already hitting memory limits while generating high resolution images, generating in large batches, or running multiple AI models at once, rather than for casual single image use. The contents of this repository are released under the MIT license, though the third party tools it depends on carry their own separate licenses.