KREA2 v2: I wanted to test out how Krea2 handles character consistency beyond the body and facial characteristics, such as tattoos and naughty bits. So I chose this character for a previous dataset test run, because of the image dataset variety and general quality of the images. Hole-ly moley is it good. The results have much higher quality in both prompt adherence and image outputs.

The bad part is that it takes quite a bit more time to train than previous base models and since I rent GPUs to train, my out of pocket cost to retrain all of my previous characters would be higher than I'd like to entertain (otherwise I'd do it). So, I'm just going to retrain a few of my favorites and for the others wait until I find a more optimized training approach or the costs to rent beefy GPUs goes down by a bit more. I wouldn't hold your breath.

Anywayz.. back to this big guy. The use of bypass LoRAs and/or text encoder Comfy nodes aren't necessary to get spicy stuff going but it does help. Example images with prompts provided in the gallery.

Enjoy.

Flux 2 Klein 9B v1: I created this character LoRA, I'm naming Victor Hernandez, based off some previous datasets that I image restored / upscaled and pieced bits and pieces together, to further test my current character creation workflow. Results were pretty good so I thought I'd share.

Character consistency traits: Latino, dark hair and eyes, upper arm tats, hung (cut), with low-hangers and a tight fit bod. I made it where these traits shouldn't be hard to modify with prompting to suit the particular need of the character composition in mind.

Image gallery uses basic default settings for Flux 2 Klein 9B with no additional LoRAs or post edits/upscaling.

Enjoy