CIVITAI / Workflows

Krea2 Two-Stage High-Resolution Refinement Workflow

Watch the full video first if you want to understand how this Krea2 two-stage high-resolution refinement workflow works in practice. The video shows how a fast Krea2 Turbo image generation route can be extended into a cleaner high-definition pipeline by adding latent upscaling and a second refinement pass. This ComfyUI workflow is designed for Krea2 two-stage image generation and high-resolution polishing. Compared with a simple one-pass Krea2 workflow, this version first creates the base image, then enlarges the latent, and finally runs a second sampling pass to improve structure, detail, texture, and overall visual clarity. It is useful when a normal Krea2 result is compositionally good but still needs more sharpness, scale, and final polish. The workflow uses krea2_turbo_bf16.safetensors as the main model. The text encoder route uses qwen3vl_4b_fp8_scaled.safetensors with the Krea2 CLIP type, while qwen_image_vae.safetensors is used for final decoding. This keeps the workflow compact, fast, and suitable for RunningHub online use. The first stage starts from an EmptyLatentImage controlled by FluxResolutionNode. In the uploaded setup, the resolution route is configured for a 9:16 vertical layout, making it suitable for vertical posters, character covers, mobile-first artwork, social-media thumbnails, and short-video platform visuals. The first KSampler uses 8 steps, CFG 1, euler sampler, simple scheduler, fixed seed, and full denoise. This stage builds the main composition, subject placement, atmosphere, lighting, and overall visual direction. The prompt conditioning is strengthened through ConditioningKrea2Rebalance. It uses 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 2.5, which gives the image stronger prompt response without becoming as aggressive as the multiplier 4.0 workflows. This makes it more balanced for high-resolution refinement, where the goal is not just intensity, but cleaner structure and better final detail. After the first sampling stage, the latent is enlarged through LatentUpscaleBy with a 1.5 scale factor. This is the key difference from a normal Krea2 baseline workflow. Instead of decoding immediately, the workflow keeps the result in latent space and increases its scale before refinement. The second KSampler then performs a high-resolution polishing pa