New model FLUX Kontext is Amazeballs

[Winnow]

Late last week Black Forest Labs released the local version of their latest model Flux Kontext.

It's the real deal. It can change image but keep the context. same pose (or change pose), add text, change text keep same style of text. It can change style "make realistic" "make the image a sketch" It can combine images . instant colorize images, etc etc.

It can do a shitload of things. I've been messing with it for 4 days straight and barely scratched the surface.

oldman-robot.png

The old man is sitting in a chair. The robot is looking at the old man. The robot is smiling

It takes about 60 seconds to generate an image (RTX 4090). It is able to manipulate the character/items but keep them consistant

oldman-hat-spang.png

The old man is weating the baseball cap from the second image. He is wearing a dirty old T-Shirt. He is sitting at a table with a wooden sign on it that says "Spang was Right!"

It is good with text. It kept the text on shirt same while making the t-shirt dirty. It added the additional text. Kept the hat the same.

combined.jpg

expand the image for more detail.

Change Snow White text on bottom to "Woke Garbage" keep same style.
Change Bowling Columbine text on bottom to "Spang Loves this Movie" keep same style.
make image realistic

It's is amazing with changing text. Look at how well it changed "Bowling for Combine" to the newt text keeping the exact formatting and even his feet on top of the text....60 seconds and a simple prompt to do this. These are not cherry picked. First result I got I used. I even typo'd "The old man is weating the baseball cap from the second image." and it still got it right.

"Male image realistic" is all you need to prompt for Kontext to convert any art style to realist (or vice versa changing styles of all kinds)

Optimally you want 24GB VRAM to run this but I've seen people do it with as little as 8GB VRAM but I'm sure it will be really slow instead of under a minute but at least you can actually use some form of it. Do not but any new GPU with less than 24 GB.

This model is AMAZING. AI is constantly improving but this kind of image manipulation on your own PC is a huge step. The recent emotion/inflection AI voices are amazing as well.

Way less need for LORAs. you can make full story comics keeping character consistency using a single image. (can change poses, clothes etc while keeping face etc consistent) I'll show that in another post.

[Winnow]

workflow.png

This is one of the workflows I use for Kontext multiple images. It's very straight forward. No special nodes needed (just make sure ComfyUI is up to date)

The nodes to the right of the sampler are mostly just to create that stitched together images for the examples I used. The bypassed pink node is a Lora to speed up the process from 20 steps to 8 steps with a slight hit to quality. Good for testing etc.

FinalImage_20250630-013611.jpg

This is just to give example of using a single image to create images with Flux Kontext. The lady in the pink sweater (AI created not real woman) in the snow is the original picture. (I think it's an SDXL image I made)

change her top to a red and white striped tank top. she is walking along the beach at sunset. She s wearing a sun hat. She is walking a dog.

she is hugging a man. She is wearing a strapless dress. In a park. Summer time. The wind is clowing her hair

typo'd blowing "clowing" but it still worked.

The other images are just showing that you can use one image and get consistent results. Hands and feet are very good (save for the kneeling facing away toes).

voicejack.jpg

remove cloth covering face. give him a cape and a red light sabre in one hand. In his other hand he has holding up a wooden sign that says "Jackass". Backround is a space station. Yellow test at the top says "Voice Actor's Last Stand"

It's a testament to how good this model is that it can manage to make the image with all my typos! "Backround is a space station. Yellow test at the top says"