Krea-2-Turbo_clear
Compare to Krea-2-Turbo (Original)
Krea-2-Turbo_clear | Krea-2-Turbo (Original)
A fine-tuned version of the Krea-2-Turbo model that reduces noise for a clearer look and ensures vivid color expression, especially for anime illustrations.
The output from the Krea-2-Turbo_clear model has higher saturation compared to the original, and adds a slight reddish tint, particularly to skin tones.
BF16 vs INT8_ConvRot_HQ
BF16 | INT8_ConvRot_HQ
INT8_ConvRot is a new format that ComfyUI officially supported on 2026.7.1. It uses about half the capacity of BF16 but achieves precision comparable to the original BF16.
INT8_ConvRot_HQ is a custom High Quality format that keeps the first and last layers in BF16, which have a particularly strong impact on image quality.
Recommended Settings
Sampler: euler
Scheduler: normal
Steps: 8-12
The PNG workflow can be downloaded from Hugging Face.
Tips (External site)
License
Description
FAQ
Comments (5)
First, thanks a lot for the work on this K2 model.
Would it be possible for you to add informations such as the recommended sampler, scheduler, CFG value, VAE (Qwan or Wan 2.1?), picture size (1152*896, 1216*832, 1280*768, etc...)?
Thanks for your comment!
This model is a fine-tuned version of the Krea-2-Turbo model. The Clip, VAE, and basic settings are the same as Krea-2-Turbo.
I personally use the following, but feel free to change them based on your preference.
- qwen3vl_4b_bf16.safetensors
https://huggingface.co/Comfy-Org/Krea-2/tree/main/text_encoders
- qwen_image_vae.safetensors
https://huggingface.co/Comfy-Org/Krea-2/tree/main/vae
- euler
- normal
- 12 steps
As for aspect ratio, I believe any value works as long as the total is around 1 megapixel.
@easygoing0114 Thanks for the clarification. 👍
Thank you for sharing this fine-tuned Krea 2 Turbo model. I really like the cleaner output and was considering using it as the base for training a character LoRA.
However, as I understand it, Krea 2 LoRAs are normally trained on the Raw model and then used with the Turbo model for inference, rather than being trained directly on the distilled Turbo checkpoint.
Would it be technically possible to apply the same “clear” fine-tuning to the Krea 2 Raw model as well? If so, do you have any plans to create and release a corresponding Krea-2-Raw_clear checkpoint for LoRA training?
I think a matching Raw version would be very useful for users who want to train LoRAs while preserving the cleaner texture and color characteristics of this Turbo fine-tune.
Thank you for your comment!
As you pointed out, the Krea-2-Turbo model can be thought of as a distilled version of the Krea-2-Raw model, adjusted to produce more stable outputs. This "clear" model is a further adjustment on top of the Krea-2-Turbo model.
What I did was a layer-wise adjustment, which isn't strictly distillation, but I think it moves in the same direction in the sense that it reduces diversity.
I believe the reason training on the Krea-2-Raw model is recommended is that it retains maximum diversity, so a Krea-2-Raw_clear model probably wouldn't be very well suited for training purposes.
There's also the issue that the Krea-2-Raw model's outputs aren't stable to begin with, which makes it inherently difficult to apply this kind of adjustment.