version 1.0 was trained over 16 epochs
version 1.0b was the same training but stopped at 10 epochs.
1.0b is somewhat more flexible. But both create good results, so I decided to publish both.
This LoRA turn things into clusters of greebles and generally small geometric particles.
Basicly stuff you can create with the MoGraph Cloner inside of Cinema 4D
Again I trained on a bunch of Dall E 3 generations. This time, the prompt was:
"3d stacked, multilayered, infinite mograph, cluster knolling [subject]"
main trigger: m0c
all words used in training: (some work better than others)
m0c
infinity
face
transformers
cluster
tunnel
starwars
weed
cat
lobster
crustacean
microscopic
DNA
fractal
office
eye
emoji
stack
city
themepark
mushroom
explosion
dandelion
neon
cube
tech
DVD
pyramid
layer
tree
cottoncandy
rock
coral
chameleon
sunset
ring
square
I suggest playing with all levers there are. CFG scale, Clip Skip, LoRA strength, sample count.
And play around with weighting trigger words of course.
Description
version b is just stopped at epoch 10 as opposed to version 1.0 is trained through 16 epochs.
FAQ
Comments (5)
I was going to write that it(v1.0) is so strong and hard to control, but i just saw that you've uploaded another version(v1.0b). I will try that too. Thanks for sharing.
You have created some really neat stuff with it!
Yeah. I didn´t test it thoroughly enough before publishing it to be honest.
But the main issue was that I have about 200 training images. So about 20k of learining steps would probably be necessary. But CIVITAI training is limited to 10 steps. I tried to counter that by increasing the LR, but I still think that didn´t go so well.
I guess I should reduce the image set or train it locally but I didn´t get any local method to work properly.
Do you have any suggestions?
Thank you!
v1.0b version seems better to control. I'm not an expert on creating LoRAs but i usually keep the step count something like 2500. So when the training image count is high i use less repeats. Never tried something like 20k.
This subject is a bit challenging to control i think. The clones in your training images are too different (lines, spheres, cubes, random blocks etc) and i realized you used some additional tags for them which is good to use (and i think those shape tags should be in all of them).
But in some images (if that captions were final) only the trigger word exist and the shapes are so random on those images. So maybe (again, i'm not an expert) adding some extra tags to those images, or completely removing them may help to get better results.
And lastly, great LoRAs you create and share, thanks for the hard work and sharing with the community!
Any way to make one for SD 1.5? I dont got a strong enough gpu to do xp unfortunately, if not, I'll understand.
Great Job. Please improve this lora to be capable of creating full shape of any animal, birds or object. Though I have not tried all the trigger words, I found the lora is limited to face.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.