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 pass. It uses the same model and conditioning, with 8 steps, CFG 1, euler sampler, simple scheduler, fixed seed, and denoise 0.5. This allows the workflow to refine the upscaled latent without completely destroying the first-stage composition. The final image is decoded through VAEDecode, previewed through PreviewImage, and saved through SaveImage with a high-resolution Krea2 prefix.

Main features:

• Krea2 two-stage high-resolution refinement workflow

• First-stage base image generation

• Latent 1.5x upscale refinement route

• Second-stage denoise 0.5 polishing pass

• krea2_turbo_bf16.safetensors support

• qwen3vl_4b_fp8_scaled.safetensors text encoder

• qwen_image_vae.safetensors VAE

• ConditioningKrea2Rebalance enhancement

• Custom 12-layer Rebalance weights

• Multiplier 2.5 balanced control

• 8 steps in Stage 1

• 8 steps in Stage 2

• CFG 1 setup

• euler sampler

• simple scheduler

• Fixed seed for repeatable comparison

• FluxResolutionNode size control

• 9:16 vertical output route

• VAEDecode final decoding

• PreviewImage and SaveImage output

• Suitable for posters, character art, cinematic covers, and high-resolution Krea2 refinement

Suggested workflow:

Start with a prompt that already produces a strong base composition. The first stage should solve subject, scene, lighting, and overall image direction. Do not rely on the second stage to fix a bad composition. Use the second stage mainly for detail, scale, and polish. If the final image becomes too different from the first-stage idea, reduce the second-stage denoise. If the image still looks soft, strengthen detail words in the prompt or slightly increase the refinement pressure. This workflow works best when the first render is already good, but you want a cleaner and more finished high-resolution result.

⚙️ RunningHub Workflow

Try the workflow online right now — no installation required.
👉 Workflow: https://www.runninghub.ai/post/2072204149030211585?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/BV1FWT76KENn/

☕ 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/post/2072204149030211585?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/BV1FWT76KENn/

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。