Video - https://youtu.be/k6ZU1PuRvLM

Overview

Lots more stuff in this week's Krea-2 workflow power-up, which follows on from last week's Krea-2 workflow!

A few more folks have released their own nodes for using reference images, so now there are 4 different options in total for using references:

Qwen Image Edit Plus

Krea-2 Image Edit Rebalance

Text Encode Krea-2

Batched Images via the AI Prompting

For extra details, the main KSampler group now features the Detail Daemon Sampler which can also help with getting proper text.

Talking of details, CFG-Zero-Star and Krea2T-Enhancer are now in the model chain as experimental options, should you wish to try those. Krea2T-Enhancer tends to work better with simpler prompts, but CFG-Zero-Star will take whatever you throw at it.

The AI Prompt Enhancer can now see all 3 images (if loaded) and the modified the system prompt is slightly better at followinf your instructions. However, poor old Qwen really isn't quite as bright as Gemini-4 when it comes to actually being able to follow instructions, so you may need to be very specific in your prompting. Qwen will often ignore what you ask - or do the opposite.

Workflow Information

To load a workflow in ComfyUI you can either drag the .json file into your browser window, or use file -> open.

Models

You will need to download the model files if you don't already have them. Click the links on the note in the workflow to download file to the specified locations.

The workflow uses the standard Krea-2 model files from Comfy, along with Lens, ERNIE and Z-Image Turbo for the default refiner choices. Feel free to change refiners to your favourite models!

If you've used last week's workflow then you'll already have all the model files required.

Custom Nodes

Custom nodes can be installed using ComfyUI Manager or git clone in your custom nodes directory. There are a few more custom nodes in this release compared to last week, as detailed in the workflow note:

Model Loading

Note that various model sizes are also available and can be loaded here. This includes the larger FP8 Krea model or the smaller FP8 QwenVl model. If you've got a lower VRAM system, then the smaller models are recommended. If you have plenty of VRAM you can go all-out with the larger options!

The model loader now features CFG-Zero-Star and Krea2T Enhancer options (disabled by default):

Krea2T-Enhancer can enhance prompt comprehension... or mess it up! As detailed in the note, Krea2T-Enhancer can really over-saturate colours and blend prompt elements into one, big mess... which is why I've set the default value down. On the flip side, it improve results if you've got a fairly simple prompt (something like "a face" or "a portrait of a flower"). If you do have a simple prompt, then you may like to crank that value all the way up to 1 (or higher).

Note: When doing upscale refining it will add artefacts if the "upscale image by" isn't 1.

It is disabled by default for the refiner steps:

It can be enabled if using the KREA model for upscaling (Option D):

Any other model will simply skip using the node.

CFG-Zero-Star is much more chill and can help prompt adherence without messing stuff up. 0 is typically the best value, but you can go up to 2 (if using the default 8 steps).

Control Panel

The Control Panel has also been updated with more options. Check the in-workflow note for more info if needed.

3 different encoding options are now available, so reference images can be seen by one of Qwen Image Edit Plus, K2IE or Krea2-Text-Encoder.

Qwen Image Edit Plus is just the same as before, and has no extra options.

K2IE is my current favourite, purely because you can easily alter the strength of each image:

Imagex_tokens has 4 options (low, normal, high, max) so if you want a particular image to have more or less influence, you can totally do that now.

I've got 3 image groups attached by default, but if you want 4 then it is simple enough to copy an image loader group and attach that should you wish.

It does also feature negative prompt support, so that's a thing!

The Krea2-Text-Encoder has support for even more images:

Just keep copying the image loader groups, update the SET node names, and you can attach 4,5,6,7,8+ images. It also features mask inputs too, but in my testing it still appears to describe the entire image anyway. Your mileage may vary.

Vision Position allows before or after, with before generally being better, so far as I can tell.

As a way to partially replicate the "tokens" functionality of K2IE, the ConditioningAverage node value can be lowered in order to reduce the impact of Krea2-Text-Encoder.

Detail Daemon

The KSampler now has Detail Daemon by default. Setting the detail_amount value to 0 will disable it.

Typical values for detail_amount would be up to ~1.0, but can go higher if you also change the start, end, bias, etc. 5 is the max.

AI Prompt Enhancer

I've moved the AI Prompt Enhancer up into the "dragons" area if you're looking to change the system prompt:

It can now see all 3 images (if loaded), but you'll need to be very careful with your prompting as it can take things very literally... Experiment with your own system prompts, but this is one more way to use your reference images - and perhaps with a bit more flexibility.

Enjoy :)