I'm excited to share a project born from a passion for high-quality image generation. As a huge fan of Afroman4peace's work, I've long relied on his Colossus Project Flux V12 "Behemoth" (AIO) for its incredible consistency. I decided to see what magic would happen when it was merged with the unique aesthetic of Flux Krea, a model known for its exceptional realism and ability to overcome common AI generation issues like oversaturated textures.
The result is a truly versatile model that brings the best of both worlds. It maintains the stable and predictable results of Behemoth while offering the enhanced creativity and realistic artistic output of Krea. It's a fantastic all-in-one solution for anyone looking for reliability without sacrificing imagination.
A few notes.
1) It's large. It takes a long time to load. 86.67 seconds just to load the checkpoint on my system. You need patience. On the same note, generating a single image takes about 26GB of VRAM. It's not a batch cook model.
2) It doesn't like lighting instructions in the Prompt. If I had to guess I think this is the Krea in it. Everything else is fine, but giving it lighting instructions results in the plastic skin effect.
I've also uploaded a decent workflow for the model here:-
https://civarchive.com/models/1962578?modelVersionId=2221421
Description
FAQ
Comments (13)
Way too big man...even loading this on an H100, it's taking minutes. But once loaded, the model is AMAZING!!!!!!! worth the wait!
It's only going to get bigger. The next stage is to train it on Hands and Feet. Poses after that, and maybe clothing. Not sure if clothing is going to be it's own seperate LORA, because I want to use the FAshion CAptioning Dataset, and that is fucking huge.
Can I use this with 32gb ram / 24 gb vram ..? Was wondering
I just tested it using a very simple workflow. You might struggle.
only the 5090 player can offord it
nvidia will likes you.
Admittedly it’s a model for the future.
Seems to be running on my 3090 (24 GB). While the model is quite large, its output isn’t especially impressive. It might handle sexy female figures well but tends to produce anatomical errors when generating average-looking men. And no, I don't mean cocks, when talking about anatomy. The second thing, prompt adherence does not appear to be improved at all.
Fair criticism. Working out what to train it with in the next version and male anatomy might be it. As for prompt adherence just want to make sure you're using a dual clip loader and using prompts in both clip_l (up to 77 tokens) and t5xxxl (up to 512 tokens)? Just want to establish if the limitations are with the model or with the prompting.
@Project_SERA Thanks for the reply. Yes, I was using a dual clip loader and using prompts for both.
@brownbagel0 Thanks. Obviously it's the model itself. I'll play around with some fine tuned clip_l and t5xxl models see what I can bake in.
I was using various versions of the "Behemoth" (AIO) and loved it. I have now switched to Krea and just found this, you made me very happy, so thank you for that and for your work and time :)
I am just running it on my rig and it is superb ❤️
I'll post some images soon.
My rig specs are:
Intel Ultra 9 285K (Liquid Cooled by Corsair)
Gigabyte Z890 Auros Elite Wi-Fi7 Mainboard
Gigabyte RTX 5090 Water Force
128Gb CORSAIR VENGEANCE RGB DDR5 6400
2Tb Crucial NVMe Gen5 (Boot Drive)
2 x 4Tb Samsung NVMe for BIG games & AI
18TB Western Digital HDD
20TB Seagate Barracude HDD
Sadly, after doing multiple images it does seem to struggle with hands which is bizarre as I never had that trouble in any Flux system.