wtf is comfyui-krea2-ostris-edit?

This project adds a small set of custom nodes to ComfyUI, the visual pipeline tool many people use for running AI image generation and editing models. Specifically, it lets ComfyUI work with a family of image editing LoRAs built on the Krea 2 model, ones that were trained using the ai toolkit's edit mode. There are two nodes included. The first is a text encoder node that takes a text prompt along with up to three reference images, and turns them into the conditioning signal the model needs. It uses Krea 2's own text encoder, called Qwen3-VL, and lays out the prompt and images the same way they were arranged during training, with each image marked as Picture 1, Picture 2, and so on. If a VAE is connected, each reference image is also encoded into latent space so the model can use it as a visual reference rather than only reading it through text. Images are resized to match how the model was trained, shrunk to fit within roughly 384 by 384 pixels for the text encoder and up to about one megapixel for the reference latents, though they are never enlarged. The second node patches the Krea 2 model itself so that it actually pays attention to those reference latents, something the standard Krea 2 setup in ComfyUI does not do on its own. It adds the reference images into the model's token sequence and treats them as fixed context rather than something being denoised. An optional caching mode can speed things up by computing the reference tokens once and reusing them at every sampling step, but this only works correctly if the LoRA being used was itself trained with that same caching option turned on. Installation is a simple git clone into ComfyUI's custom_nodes folder, with no additional dependencies to install. The project is aimed at people already using ComfyUI for image generation who want to run Krea 2 edit LoRAs specifically, and it is used by wiring the new nodes into a graph alongside the standard model loader, LoRA loader, and sampler nodes.