V.3.0
Merged the the two old models ( different ratios ) with a new mech based model i'm currently training to see what would happen.
Will add more pics later. For now, just reused the prompts from V.2.0 for comparison.
e/: Pff, cross-posting fked again. Seems to be an issue with merged LoRAs and A1111. Just another thing i can't do anymore. HASH ID doesn't match and i even re-downloaded the file from Civit.
Any other LoRA's used are still cross posted. CivitAI generated images should be fine to.
V2.0
Just a little test to see if the fp8 training setting in Kohya works... well it does. I personally don't need that setting, but memory usage went from 11.8GB down to like 8.7GB. I don't use the memory optimization option, only xformers, gradient checkpoint and cache latents. I guess with memory optimization it could go under 8GB.
My own base model is getting a bit to "all knowing" when it comes to strange stuff, so i decided to use the common models again, at least for image reproduction purpose.
More pics added ( especially mechas )
Bit more flexible ( not there yet, don't have much time at the moment for generating, training or Photoshop )
It's not really intended for robots/mechs/cyborgs...well, wasn't intended to do the stuff which it is doing right now in the first place, but it should add some variety to normal outputs.
V1.0
Not really were i want it to be, need to change a few things.
Strength of 0.8 might be to high in some cases, 0.7 or 0.6 should to or it will overpower everything. Of course that depends on how long your prompt is.
Even though i use Tank a few times, its not trained on pictures of tanks as a whole, only the plating of armored vehicles, planes an everything that could count as armor.
As always i will test it here in there.
Was actually intended to do something else, so it tends to do close ups, depending on the subject.
Already made a few images while testing in my gallery.