Video - https://youtu.be/NnOgH6BZGuw

Overview

I've been working hard through this British heatwave, testing out various new features for my Krea-2 workflow, because often using more than 1 image resulted in split-frames results. Yes, Krea-2 kinda does that a lot anyway even with just plain prompts, so it isn't just an image, but hey!

Whilst many of the image composite nodes I tried failed pretty hard (I'm looking at you Layer Forge and others!), almost by accident I found some of the native nodes to be just as much fun!

Being fed up with custom node packs which just plain didn't work, had broken features or needed older versions of the transformers library, I figured I'd stick with native ComfyUI nodes to make sure things would work.

I started off with the Porter-Duff Image Composite node mostly because it meant no fiddling around. Yes, I can't place a rodent on the left like other packs could do (if they worked)... but Krea being Krea, it kinda didn't care!

As those tests worked so well, I figured why not also use the Latent Blend node to see what IMG2IMG came out like. Krea-2 also didn't mind and was able to "pick out" parts of both images.

Things turned out nice and weird, so I figured I'd weave those options in. One example of the weirdness is shown below:

As combining images was fun, I then gave the Rebalance Guider a go for testing positive & negative prompting combined. That worked reasonably well, so just to make things even more weird, I split out the images for the +VE & -VE conditioning so that the -VE conditioning could take images without them going through into the positive conditioning as well.

For more effective negative prompting there is https://github.com/blue-pen5805/ComfyUI-krea2-negpip - which uses the old (negative:-1) (another negative:-1.1) format.

Note also that you'll need to git clone this one into your custom nodes directory as it isn't available via ComfyUI Manager. Maybe in the future, but it has already been a while...

As creating characters is useful for games like D&D (or just for fun!) I also added Identity Forge.

The whole list of updates is provided via the in-workflow notes:

I've been testing so many things I might have forgotten something, but it should all be there!

Workflow Information

To load the workflow you can either dragon the .json file into comfy, or use file->open.

As usual you should install all custom nodes & download the model files.

Custom Nodes

Some new custom nodes compared to last week are:

Identity Forge

Krea2 NegPiP

Full list provided via the in-workflow note:

Model Files

Model File links are also provided in-workflow:

Model Loader

The Model Loader group now provides NegPip and the Krea2T-Enhancer Advanced options:

Select which models you'd like to load (Krea-2 has NVFP4 up to BF16) here.

Control Panel

The Control Panel has a bunch of extra options. See the note for additional information.

The optional Rebalancing KSampler group can be enabled/disabled:

As this is an optional group, you'll need to change the latent used for the Standard Output from LAT_STD to LAT_REB to see the output in that group:

Or your could send it straight to the 2x VAE group if you want to have both KSampler outputs:

The QIEP -VE option allows you to separately send any images to the negative prompt conditioning, for use with the Rebalance KSampler or CFG>1:

Note: They don't seem change a whole lot from my testing - and I was rather hoping for more of an impact - but it is there if you want to try too!

Image Choices now includes an option for Image 1 Composite, which will replace the usual Image 1 option with the composite you create.

Enable BOTH Load Image 1 and Load Image 1 Composite to create the composite image:

You can change the mask value to have more of one image or the other, as well as changing the mode for image compositing:

Makes nicely strange images!

Note that when using the composite image, this will take precedence over image 1 for most options.

You now also have more latent choices:

Being a latent, you'll mostly need to change the Ksampler denoise value to see any actual change vs denoise 1.0 and an empty latent. Values below 0.8 are generally needed, with values less than around 0.6 needed for plain IMG2IMG.

The two new options are Image Composite and IMG1 & IMG2 Blend.

The Image Composite option created a latent from the Image Composite (so you'll need to have that image available).

IMG1 & IMG2 Blend makes a blended latent version of two images, but note that this node needs both images to be the same size (e.g. both at 1280x720). The stretch option in the image resize node is great if you have two different image resolutions.

You will, of course, need to load both Image1 and Image2.

There is also a Blend Factor available if using IMG1 & IMG2 Blend:

These latents can be used on their own for IMG2IMG generation, and can be mixed with any of the Image Encode options.

Note that the Image Encode nodes are now grouped alongside the image loaders for ease:

ID Forge

ID Forge, though not really AI Prompting, is included in the same fast group enable option:

This character creator has LOADS of options:

So if you want to make something like a bard, cleric, paladin portrait for your D&D game, this makes things really easy. Just select an archetype and go!

You can just leave everything on random too and keep on create different people, which is also fun in and of itself.

Enable Concat ID, and you'll mix the two prompts so your character prompt gets appended to your main prompt:

Optional Upscaling

The upscaling models are the same as last week, but if you're new to this workflow then you can either use the defaults provided or switch them out to whatever your favourite models are:

Upscale 1

For speed, the default settings in Upscale 1 and Upscale 2 have been change, so now it's only 2 steps using linear_quadratic:

Hopefully this UK heatwave will sort itself out soon, as the computer heat is making me cook!

Enjoy! :)