Krea 2 v1: Introducing the latest addition to my AI Nobody character roster, Jon Javier.
Jon's character construct is a African American dude, aged somewhere to his mid 20s to 30s, with a sometimes-skips-leg-day-big-pecs athletic build, sparse & trimmed chest hair, and a big long slightly vascular dick attached to a pair of low-hangers. He is from somewhere cool, hates it when that happens and loves to [0.38291, -0.77104, 1.29483, 0.00412, ...].
Okay, character description backstory finished. You can alter and fill in the rest. Pics in the gallery.
Two main reasons why I created this AI Nobody character: 1) Test a dataset with large resolutions and 2) create a Krea 2 type version for one of my older Consistent Character Cocks inpaint LoRAs but this time able to also create a full character for more LoRA use versatility.
To expand a bit on the first nerdy point which is really what drove the creation of this model. I wanted to test how Krea 2 handles very large images in the dataset. The general previous guidance of Flux 1 / 2 x / Klein , Qwen, Z-Image, etc have been that really should be in the 1024px space - where one side of the image doesn't extend beyond 1024px, or much beyond, otherwise it'll essentially just split it up during training and learn sections of the image. I haven't been able to find documentation that specifies dataset image size guidance directly from the Krea team. I also know that Krea 2 can natively handle much higher resolutions than the other models mentioned, so I thought I'd train on a dataset that consisted of images that mostly fell within a 1280px up to 1536px dimension range and a few really big ones that were greater than 2048px. Training time fell generally within the same window as before on a RTX 5090.
The results I think are fairly noticeable in a positive sense- I haven't tested the outputs mathematically but from a eye-balling it visual inspection the outputs are much better quality than the exact same dataset but at the typical pixel range I used for Flux 2 Klein and Z-Image.
I'm going to test this some more with previous datasets so I can do some more in depth side-by-side comparisons, but this result is encouraging so I thought I'd share it. Let me know your thoughts.
Enjoy.