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.

The latent section uses FluxResolutionNode with a 1.4 megapixel target, 9:21 Ultra Tall aspect ratio, and divisible-by-64 alignment. The batch size is set to 4, making the workflow useful for generating multiple vertical image candidates in a single run. This is practical for RunningHub app previews, poster selection, thumbnail experiments, style tests, and prompt comparison.

The example prompt focuses on a highly stylized oriental fantasy scene: a woman in crimson ceremonial robes standing in a sacred hall, surrounded by huge ivory scrolls covered in calligraphy that spiral through the air like living ribbons. The scene emphasizes theatrical motion, painterly color contrast, traditional architecture, sacred performance, floating sparks, divine wind, and literary magic.

Main features:

• Krea2 merged-model implementation workflow

• merge50.safetensors pre-merged UNET model

• ComfyUI-Krea2T-Enhancer model-side enhancement

• Enhancer enabled with strength 1

• qwen3vl_4b_fp8_scaled.safetensors text encoder

• qwen_image_vae.safetensors VAE

• ConditioningKrea2Rebalance control

• Custom 12-layer conditioning weights

• Multiplier 2.5 balanced prompt pressure

• 20-step DDIM sampling route

• CFG 1 setup

• sgm_uniform scheduler

• Fixed seed for repeatable testing

• Batch size 4 multi-candidate generation

• 1.4 megapixel resolution route

• 9:21 Ultra Tall vertical format

• VAEDecode final image decoding

• PreviewImage and SaveImage output

• Suitable for merged model testing, vertical fantasy art, poster generation, and Krea2 style comparison

Suggested workflow:

Use this workflow when you want to test a pre-merged Krea2 model rather than repeatedly building a merge inside the graph. Start with a strong visual prompt and generate a batch of four candidates. Compare whether the merged model improves texture, color, style consistency, and composition compared with the base Krea2 Turbo model. Since the workflow uses 20 DDIM steps, it is better suited for deliberate testing than ultra-fast preview. Keep the prompt focused on one main visual direction, then adjust the merged model, Rebalance strength, or sampling settings only after the subject and composition are already stable.

⚙️ RunningHub Workflow

Try the workflow online right now — no installation required.
👉 Workflow: https://www.runninghub.ai/ai-detail/2073373450613383169?inviteCode=rh-v1111

If the results meet your expectations, you can later deploy it locally for customization.

🎁 Fan Benefits: Register to get 1000 points + daily login 100 points — enjoy 4090 performance and 48 GB super power!

📺 Bilibili Updates (Mainland China & Asia-Pacific)

If you’re in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown.
📺 Bilibili Video: https://www.bilibili.com/video/BV1oUTt6ZEEp/

☕ Support Me on Ko-fi

If you find my content helpful and want to support future creations, you can buy me a coffee ☕.
Every bit of support helps me keep creating — just like a spark that can ignite a blazing flame.
👉 Ko-fi: https://ko-fi.com/aiksk

💼 Business Contact

For collaboration or inquiries, please contact aiksk95 on WeChat.

⚙️打开下方链接即可在线体验，无需安装。
👉 工作流： https://www.runninghub.ai/ai-detail/2073373450613383169?inviteCode=rh-v1111
如果觉得效果理想，你也可以在本地进行自定义部署。

🎁 粉丝福利： 注册即送 1000 积分，每日登录 100 积分，畅玩 4090 体验 48 G 超级性能！

📺 Bilibili 更新（中国大陆及南亚太地区）

如果你在中国大陆或南亚太地区，可以通过下方视频查看该工作流的实测效果与构思讲解。
📺 B站视频： https://www.bilibili.com/video/BV1oUTt6ZEEp/

我会在 夸克网盘 持续更新模型资源：
👉 https://pan.quark.cn/s/20c6f6f8d87b
这些资源主要面向本地用户，方便进行创作与学习。