Deep under a mountain lives a sleeping giant, capable to eighter help humanity or create destruction...
A Colossus arise...
This is my first try to create a costom Flux Kontext model...
If you like it, feel free to give me some feedback. Also if you want to support me you can do this here. I have spend some good money to build a computer that is capable to actually train Flux-models.. Also training and testing takes also a lot of time and electricity..
https://ko-fi.com/afroman4peace
Version V1.0 Experimental
This version is based on my latest Colossus Project Flux models https://civarchive.com/models/833086?modelVersionId=1964437 and a version of Flux Kontext DEV I have modified to get less anatomical fails. If you are interested I may upload this version as well..
Parts of the model are merged via blockmerging and parts are merged via merging in some of my loras I trained myself to create Colossus Project V10.
In short: this checkpoint is a sort of Frankenstein-Checkpoint
Usage:
I will soon publish my workflow for this model. At the moment my current Colossus Project Flux model beats this model here by far for image creation...
For image editing Flux Kontext is just amazing :-)
Here is a good example: I changed out the text and added cat ears
Is this version better than the normal Flux Kontext?
Well this version isn't much tested yet. A real training could be also better. let me know what you think of it :-)
Here are is a comparison between Flux 1 original and my experimental checkpoint..
Flux Kontext (Standart)
Colossus Kontext V1.0
Licence: This model is only for non commercial usage.
Credits:
As always a great shoutout for Black Forest Lab who created Flux Kontext
Description
Clip_L for Colossus Project KONTEXT
FAQ
Comments (16)
does this work like normal flux with normal flux clip and vae? or do i need comfyUI and new clip
This is based on the new flux kontext. But it uses the same clip and vae and sampler as the old flux.1
hi, yes this works similar to the normal Flux models.. I think you need to update Comfy UI though. The main difference is the usage of this model.. Flux KONTEXT is meant to use for edit images. For example you can load a image and then just type.. Change the text to: X. or change the haircolor to red and so on. Tomorrow I will write an detailed article.. with workflow. You can download meanwhile one of the images from the gallery. This includes the WIP workflow.
@paleopterix I have added a IMG to IMG example in the gallery. (I LOVE CATS)
@Afroman4peace so I'm guessing i cant plug it into Forge and use the flux tab? I'm not sure my system will be able to run it in comfyui
@SLACK69 I am not sure. I don't know if Forge can handle it yet. Look into the latest updates. I am gonna plan to release this as well for nunchaku. (int4 or fp4)
@Afroman4peace alright, thank you
Hey there, excited to see your take on Kontext! 🙌
It says "Checkpoint Merge" on the tin, so did you train and merge specific LoRAs to expand on Kontext's editing ability or did you try merging regular Flux Dev LoRAs?
This is checkpoint is more or less a proof of concept. For now I experimented with block merging regular flux models into Kontext checkpoints. Basically I tested all blocks and observed what each block does and how compatible it is with the regular flux models. I also merged own trained regular Flux Loras into the checkpoint.
The result was a complete different checkpoint compared to the original flux Kontext checkpoint. It's more or less a Frankenstein checkpoint between a Kontext model and a hybrid de-distilled model. You can get more details and less blur. It's not well tested yet.
Well as I wrote in the description this doesn't count or replace training in the slightest. It's like slamming the knowledge how to cook fancy dishes into a fast food restaurant and expecting that they can cook something good with the same old ingredients.
However this helped me to understand some things that will get handy in the future. The next step will be training this thing ;-)
@Afroman4peace Sounds like a promising approach. Do you think full finetune training makes sense for it? Afaik it's using before/after image pairs during training, effectively training on ~2k x 1k resolutions.
Given how few Flux Dev finetunes there are due to resource requirements, I could imagine Kontext taking even more would make it less viable.
hi. Feel free to leave some feedback or images. This thingy isn't very tested. So every feedback helps me a lot.
Your sample images don't help since we don't know what you started with. 🤷
hi, besides the woman holding the sign, I started without any input image for the showcase. You can download the images and look inside the workflow.
But you are right. The Kontext model is mainly for editing images. I should add more images that shows its image editor features.
Hello, I've put some comparative images in the gallery. How did you go about making a custom kontext? Because someone who knows a lot about it told me that it was almost impossible. He tried very small modifications, on 2 simple blocks without touching the double, and it destroyed some functions.
hi. The modifying and merging between the Flux_DEV and Kontext is indeed very difficult. I have created countless versions during this process. I am not giving away all my secrets but I can say that you need a special node for it called "Expert Flux .1 merge" than you need to select the mode "add difference". You also can just merge the final layer and the inputs this also has a effect. Another thing I did is merging in some big loras I have created for flux.1 dev. happy experimenting :-)
Afroman4peace ok thanks ;)