Summary
Veteran AI walks through four advanced Krea2 workflows built on ComfyUI and hosted on RunningHub, starting from a high-definition base workflow and extending it into image replication, image editing, and reference-based character placement. The common thread across all four is a deliberate departure from Krea2’s default settings: swapping the standard Turbo model for the RAW model paired with a Turbo LoRA, and replacing the default Qwen Image VAE with the Wan 2.1 VAE to recover fine detail in skin, glass, fabric, and hair.
The HD base workflow adds two-stage sampling — rough composition first, then a low-denoising refinement pass — plus optional upscaling and face restoration. Image replication uses the TextEncode Qwen Image Edit Plus encoder node to accept a reference image, intentionally leaving the prompt blank, and produces outputs that match the original’s composition, pose, color relationships, and overall feeling without pixel-level copying. The editing workflow is identical but adds a natural-language action instruction (such as having a character raise a gun toward camera), while reference-based generation inserts a consistent character into entirely different scenes ranging from luxury storefronts to museum corridors.
The video includes side-by-side comparisons of preview versus full-release model outputs, showing clear quality improvements in the final Hy3 training. For ComfyUI practitioners working with Krea2, this is a practical blueprint for extracting significantly higher-quality outputs than the default pipeline delivers, with all node-level implementation details shown on screen.
📺 Source: Veteran AI · Published July 07, 2026
🏷️ Format: Tutorial Demo
