Kontext: It's time

[Weatherby43]

https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev

[evanheckert]

Exciting! It's already in AI Toolkit, so it seems not to be a completely foreign architecture to implement.

[wzwow]

yes kontext pls!

[E2GO]

+1 for this.
May be we got some hope to get Kontext support for Forge with a help of @DenOfEquity and his marvelous dualPrompt extension. But no info on whether he'll ever be doing it)

[DenOfEquity]

Working with Flux is a PITA on my low-end hardware, but the implementation of Kontext is easy. You can experiment with https://github.com/DenOfEquity/forge2_flux_kontext.

[DrSid42]

Damn. It WORKS. You can have too much fun with this .. I mean it can kill you ..
Not only you can add mustaches to politicians. Not only you can gender swap them. You can change them into animals.

[E2GO]

Thanks a ton for putting this together!
Even on modest hardware (don't know details), Kontext + your extension feels great.
I’ll be testing it this week and will share any feedback or PRs that can help.
Your effort is a huge boost for the community—really appreciate it! 🙌

[yamfun]

sorry what's the link for forge2?

[DenOfEquity]

Updated. Original method was too simple, I'd overlooked that it didn't (and couldn't) set up some parameters used by the model correctly. New version shows higher quality results; mustachioed and gender-swapped politicans should now look even more believable and should work with more samplers too.

[E2GO]

@DenOfEquity Just wondering. Is it ok that swapping image gives kind of different results. In this particular examples swapping woman and puppy gives "generic" FLUX face to a character. While using woman first, puppy second - results is exactly how it should be. Same setting with fixed Seed.

Examples below:

[DenOfEquity]

Strange that it kept the clothing but changed the head and style. Tweaking the prompt might encourage retaining those features.
I'm testing some modifications that change how images are handled - might help too. Check the 'dev' branch, just uploaded.

[E2GO]

Great! Will check it asap)

[E2GO]

Dev branch. Ok now it's works more accurate for sure. No more generic FLUX faces after swap. Some micro changes into style happened after swapping (_could be improved to perfection? ;) _).

Settings tests. Overall - all outputs are completely the same for all settings. I got slightly blurred second one image. Just a little-tiny bit more realistic face of a woman now.

Reduce inputs to half width and height aren't working now? Tested on single image - same results and time.

1. BFL recommended settings.
   
   Swapped:
   
2. No change settings.
   
   Swapped:
   
3. To Output setting.
   
   Swapped:
   

Main branch to compare. Also slight blur on second one image. Could be Seed\Model problem - don't know here:

Swapped:

[DenOfEquity]

Thanks for tests. I'll update 'main' shortly as the improvement is clear. I don't know why the second image is slightly blurred, unless it's related to the old Flux issue with plain white backgrounds.
The reduce option is working for me. Resizing values are printed to console.
For the image sizes you've used, all methods should be the same. They could vary with different input/output sizes and aspect ratios. Generally the BFL option will be the correct choice. 'No change' might have value with high resolutions only.

[wzwow]

cool nice improvement!

[ride5k]

might be a stupid question, but is kontext flux-only?

[AnabelleKM]

Yes, Kontext is Flux-only.

[E2GO]

Inspired and based on @DenOfEquity script for Kontext did an advanced script with some additional stuff included: https://github.com/E2GO/forge-kontext-assistant
Didn't test it a lot so use on your own risk))