I created an AI Nobody character using Flux 2 Klein to mostly see how some of my previous amalgamated datasets and previous captioning, to see what would result with this newer model. This model was more of a test to see how much fidelity/accuracy I can get with lower bit quantization which the results surprised me how well it learned even at low bit.
I'm still playing around w/ this Flux 2 Klein but thought I'd share this given that it's a fabricated persona and with it being a 4B it can be used without BFL licensing restrictions.
What can it be used for? Well aside from the obvious full character consistency purposes, which let's face it - there are plenty of those at this point and it's easy enough to maintain a character's facial identity with several i2v models or through i2i techniques, not to mention since this is a nobody character, there's no recognition affinity, so then why add to the wash of AI characters already out there?
well... since this is a Klein model, you can use transpose characteristics very easily with the edit feature.
Want to add a consistent BBC on an existing image for fun and giggles?
Edit Prompt: replace his penis <lora:this_lora>
Bam! (Mr. Biggz?)..
Want to transform a lanky twink to a muscle bound Gigachad with a consistent body type?
Edit Prompt: transform his body <lora:this_lora>
Shabazz .. er, Shazam!
Description
Flux Klein 9B version
FAQ
Comments (7)
Hello there 🙌 mind sharing some training insights for F2K? This is all very new and any help would be appeciated if you agree to share what you've learnt. Thank you
I use AI Toolkit and RunPod (usually a RTX 5090 which seems to be the most cost efficient). I'm still tinkering with settings and once I'm happy with training settings and dataset parameters that predictably produce consistent characters I may get around to creating an article on my approach. In the mean time, there are a several published articles on here that cover the subject at fair length that are good enough to at least start with and you can go from there.
I was going to look at your loras and make recommendations based on my observations, but it doesn't look like you have any - do you post them elsewhere?
@jrewingwannabe947 how much do you actually spend in runpod to train your lora?
@jay_santos Training time is usually a little over an hour and a RTX 5090 is about $0.89/hour -so usually a buck for a character based LoRA. Style and concept LoRAs usually take a little longer, but not much, since I tend to put more into those datasets .. so maybe $1.50 max.
@jrewingwannabe947 thank you, I will try to create lora soon, now just enjoying using other user's lora.
Any chance of an XL version? Or could you provide the training data and I can take a stab at it?
I guess I could feed the dataset to the trainer on here if you want. It's been a while since I've made an XL, and even longer since I've trained one on here.. so if there's Civit training params you'd prefer, just DM me and I'll let you know what the buzz cost it comes up w/ for the training run and you can decide from there.