Video - https://youtu.be/gxNilmJ9dK0

Overview

The Krea 2 Text-to-Image Model is now supported in ComfyUI - be sure to update to the latest version of Comfy though as this one is hot off the presses and won't work with out-of-date versions!

Krea 2 Turbo is an 8-step distilled checkpoint built for fast, high-quality text-to-image generation. It’s distilled from a fully post-trained RL checkpoint, so it produces higher quality than RAW but with lower variation between outputs.

Turbo typically generates in 8 steps, though with reference images then 4 steps is also an option.

It covers a wide aesthetic and stylistic range, from photographic portraits to anime, and can also generate reasonable text.

The model was trained on short, medium, and long natural-language prompts. Long, detailed prompts generally give the best quality.

Krea 2 OSS is a 12B dense DiT. It uses the Qwen Image VAE and a Qwen3-VL text encoder with multi-layer feature aggregation.

The Krea 2 Raw checkpoint is not recommended for inference use. Instead, use it as a good base for finetuning or post-training for your own needs and domains. Train a LoRA on RAW, then apply it on Turbo. Diverse outputs and a tunable base give you room to push the model toward specific styles or subjects.

For LoRA training, https://github.com/ostris/ai-toolkit is one great option!

The ComfyUI Repo also features many LoRAs - https://huggingface.co/Comfy-Org/Krea-2

Original Repos

The original repos can be found at:

Turbo - https://huggingface.co/krea/Krea-2-Turbo

Raw - https://huggingface.co/krea/Krea-2-Raw

Workflow Information

With this workflow you can do:

Text-2-Image

Image-2-Image

Image Inpainting

Image Upscaling

1-3 Reference Images

Custom Nodes

Links to custom nodes are provided in the workflow:

ConditioningKrea2Rebalance is very new, so that one may not yet be accessible via Manager, so git clone as usual to install for the time being.

Install all custom nodes to ensure all features work.

Models

This workflow features a multi-model refiner so you can use your existing favourite models alongside Krea-2. By default for me these are Lens, ERNIE and Z-Image.

You can, of course, change the loading section of the refiner to your own choices!

Model Loading

Model Loading has options for LoRAs, sage attention and Model Shift:

Control Panel

The Control Panel has your various options ready to go. Load Image(s), Conditioning Rebalancing, Latent Size, Upscaling and more!

Load Images

This isn't an Edit model like Qwen Image Edit and so any images provided are only references. You may be able to see the influence of this image (along with 2 others) in the title image at the top of this page! Think of these images more like suggestions.

Prompting

The prompt section features a negative prompt, but this won't be used at CFG 1. Change the CFG or use the RAW model and the RAW model settings for it to impact your generations:

An AI Prompt Enhancer is also available. Flip the switch to enable, and update the system prompt if you like!

KSampler

Krea-2 uses Euler / Simple by default, but lots of other options work too. Play around and see which you prefer!

When using more than 1 reference image, I personally found things worked better with a lower denoise. For example:

With the full denoise value, results tend to be side-by-side rather than integrated into the scene in a cohesive way. Values will vary depending on scheduler.

1 image works just fine as normal though!

Upscaling

There are many options for upscaling:

Krea-2 will change the image more than other models, such as ERNIE, so I'd suggest starting at a denoise of around 0.25 and work up from there.

The 2x VAE is the easiest way to get a nice, large image which is identical to the original scale one.

Enjoy! :D