CIVITAI / Workflows

Krea2 Three-Stage High-Resolution Rendering Workflow

Watch the full video first if you want to understand how this Krea2 three-stage rendering workflow works in practice. The video shows how a Krea2 Turbo image can be built step by step: first creating the base composition, then refining the latent image after the first upscale, and finally polishing the result through a third controlled refinement pass. This ComfyUI workflow is designed for Krea2 three-stage image rendering. Compared with a simple one-pass Krea2 workflow, this version does not stop after the first generation. It uses a staged latent pipeline to gradually improve the image structure, scale, clarity, and final visual finish. This makes it suitable for creators who want a cleaner and more production-ready result from Krea2 without building a heavy external upscaling system. The workflow uses krea2_turbo_bf16.safetensors as the main generation model. The text encoder route uses qwen3vl_4b_fp8_scaled.safetensors with the Krea2 CLIP type, and the final decoding route uses qwen_image_vae.safetensors. This keeps the workflow compact, fast, and practical for RunningHub online use. The prompt route is strengthened through ConditioningKrea2Rebalance. In this workflow, Rebalance is set to the balanced preset with a custom 12-layer weight structure: 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 2.5, 5.0, 1.1, 4.0, 1.0 The multiplier is set to 2.0, with renormalize enabled. This makes the conditioning stronger than a plain Krea2 render, but less aggressive than the heavy multiplier 4.0 workflows. It is a balanced setup for staged refinement, where the goal is not only strong prompt response, but also cleaner composition and stable detail across multiple passes. The first stage starts from an EmptyLatentImage controlled by FluxResolutionNode. The active output format is set to 3:4 Golden Ratio, making this workflow suitable for vertical posters, character images, fantasy covers, mobile thumbnails, and stylized AI artwork. Stage 1 uses 6 steps, CFG 1, euler sampler, sgm_uniform scheduler, fixed seed, and full denoise. This stage creates the base image composition. After Stage 1, the latent is enlarged through LatentUpscaleBy using nearest-exact at 1.25x scale. Stage 2 then runs another 6-step KSampler pass with denoise 0.45. This stage refines the upscaled latent while preserving the main structure. The workflow repeats the same logic once more. The Stage 2 lat