Krea 3 is the image model Krea launched through ComfyUI Partner Nodes on October 15, 2026, with Comfy's official blog announcing the integration the same day. Unlike generating an image from scratch, Krea 3 is built for reference-driven generation and editing: start from a real product shot, portrait or room, describe in words what should change and what must stay the same, then try the next version quickly.
It targets the oldest problem in AI image editing
The most common failure when editing with AI is collateral change: the background gets swapped and the product label changes with it, or a new material quietly alters the object's shape. Krea 3's editing logic asks creators to name the keepers first: product labels, shapes and materials can be pinned in place while only the setting, lighting and composition move; when one element's fabric, finish or color changes, the room, the object's outline and the camera angle stay as they were. That fits e-commerce product imagery and the endless revision rounds of client review.
Multiple references, each addressed by number
Krea 3 accepts several reference images and refers to each by upload order: the person from image 1, the jacket from image 2, the backdrop from image 3, with one job per reference. A result can be fed back in as the next starting image to adjust framing or continue a look, and Krea suggests chaining selected frames into a character sequence for a downstream video model. Numbered references cut the cost of describing the same element again in every prompt, and reduce how often multiple sources contaminate each other.
How it runs in ComfyUI, and what it costs
The path has four steps: update ComfyUI to the latest version or open Comfy Cloud, find Krea 3 in the Node Library or load the official template, sign in to a Comfy account, then load a starting image plus references and run. Partner Nodes are billed in Comfy credits, and Krea 3's pitch is precisely lower cost and wait time per image, so second and third revisions happen inside the same session instead of rejoining a queue each time. For teams already working in ComfyUI workflows, the integration needs no new toolchain — swapping in the node is enough to try it. The practical habit is equally direct: state the keepers before the changes, and when a sequence drifts, restart from the strongest earlier result instead of piling edits onto a weak one.
