This workflow is for Krea2 RAW-Turbo high-noise and low-noise sampling. It is designed to test a more advanced Krea2 rendering strategy: using a RAW model route for early high-noise image formation, then using a Turbo model route for lower-noise refinement and final completion.
The workflow uses krea2_raw_fp8_scaled.safetensors and Krea2-Turbo_fp8_nsfw.safetensors in the same production chain. It also uses qwen3vl_4b_fp8_scaled.safetensors as the Krea2 text encoder and qwen_image_vae.safetensors as the VAE. The base resolution is controlled through ResolutionSelector and is set for a 21:9 ultrawide 2MP route, making it suitable for cinematic panorama images, fantasy key visuals, wide posters, and large-scene concept art.
The workflow is built around multiple RandomNoise, BasicScheduler, SamplerCustomAdvanced, VAEDecode, VAEEncode, and ImageScaleBy stages. The RAW part is responsible for early structure, atmosphere, and high-noise formation. The Turbo part is used for later refinement, cleaner surface, and final image convergence. This gives the creator more control than a normal one-pass Krea2 render.
The purpose of this workflow is not simply speed. It is for testing how Krea2 behaves when different noise levels and model phases are separated. High-noise sampling can help establish stronger global composition and visual direction, while low-noise sampling can help refine texture, detail, and final image quality.
Main features:
Krea2 RAW-Turbo high/low noise workflow
RAW model early structure generation
Turbo model low-noise refinement
krea2_raw_fp8_scaled.safetensors support
Krea2-Turbo_fp8_nsfw.safetensors support
Qwen3-VL Krea2 text encoder
Qwen Image VAE
RandomNoise and BasicScheduler control
SamplerCustomAdvanced multi-stage chain
VAE decode and re-encode workflow
21:9 ultrawide 2MP base resolution
Suitable for cinematic panoramas, fantasy scenes, and sampler testing
Suggested workflow:
Use this workflow when you want stronger control over the generation process than a normal single KSampler route. Start with a prompt that has clear subject, environment, lighting, and visual direction. Let the RAW phase establish the image foundation, then let the Turbo phase refine the result. If the image becomes unstable, simplify the prompt or reduce the number of visual concepts. If the result is too soft, strengthen texture and lighting details before changing the sampler chain.
⚙️ RunningHub Workflow
Try the workflow online right now — no installation required.
👉 Workflow: https://www.runninghub.ai/post/2074460408995471362?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/BV1nxM56cEta/
☕ 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/2074460408995471362?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/BV1nxM56cEta/
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。
Description
FAQ
Comments (1)
I would have thought you would want to do it the other way around as the high noise is the structure of the image which Krea2 Turbo is likely the best I have seen for consistently good anatomy, like sure there will be the rare seed with an extra finger etc or some body horror if your prompt is overly complex with the pose, but it's generally pretty great..
What Krea2 Turbo lacks however is subtly if you prompt something it will not be very subtle even if you prompt for to be subtle there is a floor where Krea2 turbo will not draw it any more subtly where as Krea2 RAW will go as subtle as to make things almost practically invisible if you prompt for it, which is most important in the final
~30% of the generation.
I can understand not running completely raw though as Krea2 Raw can be a bit too soft, so I would use a extracted turbo R256 lora and running it at ~0.2 strength, but yeh I would switch to running full turbo for the high noise part of the generation then do 16 almost raw steps with a turbo lora applied at 0.2. It will also be significantly faster for another thing as you should only need to run about 6-7 turbo steps for a total of 22 - 23 steps.
Also you don't need to run two separate models, you can just use the turbo extracted lora at full strength for the high noise pass and ~0.2 strength for the low noise pass.
Edit: yep I prefer my way, used FlowMatch Euler Discrete Scheduler (Custom) to get more control over the sigmas between passes, used 10 steps from 0 to 7 for the high noise pass, and 50 steps from 35 to 50 for the low noise pass so a total of 22 steps seems to give pretty great results.
This equates to straight 70% and 30% of the generation according to sigmas ramp anyway