Krea-2 in ComfyUI: reference images and refiner tutorial

This tutorial compares Krea-2 image workflows in ComfyUI for local AI generation. It follows an earlier workflow guide and focuses on reference-image controls, prompt adherence and refiner passes.

The speaker tests CFG-Zero-Star and Krea2T-Enhancer on a complex rodent scene. CFG-Zero-Star makes a modest difference in that example. The enhancer at strength 1 produces stronger colors but merges characters and disrupts text. Reducing its strength to 0.04, or 0.03 alongside CFG-Zero-Star, limits those problems. Simpler prompts fare better at strength 1, while a mixed-style scene loses the requested separation of styles.

The refiner examples expose another limitation: enabling the enhancer while enlarging the image produces colored blocks. Keeping the resize factor at 1 avoids them in the demonstrated workflow. The speaker also compares Z-Image Turbo and ERNIE Turbo, preferring ERNIE's text repair and sharper detail in these tests.

Four reference methods receive comparisons: Qwen Image Edit Plus, an image-edit rebalance node, a text-encode node and image batches passed through AI prompting. The rebalance node accepts four images with individual weights and an optional negative prompt. The text-encode option accepts more than four images, though its masks did not work as the speaker expected. Lower reference strength and a second refiner pass improve lettering in the examples, sometimes at the cost of reference details.