CIVITAI / Workflows

Krea2 Merged Model Implementation Workflow

Watch the full video first if you want to understand how this Krea2 merged-model implementation workflow works in practice. The video demonstrates a compact Krea2 image-generation route built around a pre-merged model file, combining model-side enhancement, conditioning rebalance, and a longer DDIM sampling setup for more controlled vertical image production. This ComfyUI workflow is designed for Krea2 merged-model image generation. Unlike a standard Krea2 Turbo workflow that directly loads the original base model, this version loads merge50.safetensors as the main UNET model. That means the model fusion has already been prepared before running the workflow. The graph itself does not need to perform a live model merge every time. Instead, it uses the pre-merged model as the starting point, making the workflow simpler and more practical for repeated generation. The model route loads merge50.safetensors through UNETLoader, then passes it into ComfyUI-Krea2T-Enhancer. The enhancer is enabled, with strength set to 1. This gives the workflow an additional model-side enhancement layer before sampling. The purpose is to make the merged Krea2 model respond with stronger texture, cleaner style behavior, and better visual presence. The text encoder route uses qwen3vl_4b_fp8_scaled.safetensors with the Krea2 CLIP type. The VAE route uses qwen_image_vae.safetensors for final decoding. This keeps the workflow aligned with the normal Krea2 ecosystem while allowing the main model itself to come from a merged checkpoint. The conditioning route uses ConditioningKrea2Rebalance 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 2.5, with renormalize disabled. This provides stronger prompt control than a plain conditioning route while avoiding the more extreme pressure of multiplier 4.0 setups. It is a balanced configuration for testing merged-model behavior, prompt response, color structure, composition, and stylized texture. The sampling setup uses KSampler with 20 steps, CFG 1, DDIM sampler, sgm_uniform scheduler, fixed seed, and full denoise. Compared with the common 8-step Krea2 Turbo baseline, this version gives the workflow a slower but more deliberate render path. It is suitable when the creator wants more stability, stronger structure, and more consistent testing across multiple outputs.