Update1
Tried to make it better with local training this time. I think it's definitely a step-up from the first one, as it resembles the original render style more closely and has more texture overall making it feel more like a 3d model by default.
Changed trigger word, so use tag: m@druga
That along with the typical ones: realistic, 3d, uncensored
This version recognizes Ellie better now (though I wouldn't use it without a separate lora): ellie \(the last of us\)
I only tested this one on Hana-4-Chrome, but the results are satisfying enough that I would definitely advise you to get this checkpoint, it's really cool!
Madruga3D style from some of his Blender made renders.
It's somewhat rough at the moment, as I could've definitely used more images in the dataset and definitely could've spent more time fixing up the tagging, so consider this more of a personal prototype. I'll possibly update it once I have more buzz and figure out a good setting set.
I'd say right now the lora is more focused on close shots and oral acts, but it's not baked in, so it's still flexible with other things.
Use: realistic, (3d, madruga3d:1.2)
I think the additional weight helps negate the 2d, painterly feel of most models.
I found it to work somewhat well with JankuV4, though I like better the outputs that Hana-4-Chrome gave me. I imagine more 3d focused checkpoints could give more stable 3d results.
Experiment with scheduler/sampler and cfg to find out what works best for you.
Description
Same dataset, but different approach.
One thing I did is to not scale the images down, which earlier caused them to be a bit blurry and added compression artifacts in. I'm still unsure how resolutions of above 2000 px in width/height affect training, but it seems to work well.
I trained this locally on AdamW8bit, and due to the local training this version was trained on lower resolution altogether (768 px). It's surely to cause a slight decrease in quality, but the lora overall turned out better than the first.
Overall, some things were gained and some things were lost. I think it resembles 3d model style more closely and resembles the artist's renders more closely. I added tags for Ellie from TLOU from the original dataset. On the other hand the untagged appearance of Chloe and Max from LiS is now somewhat broken, though it was never meant to be focused on recreating those anyway.
There's a few issues still, so expect another update soon enough.
!! Trigger word is changed to: m@druga
FAQ
Comments (3)
I imagine not all outputs with the model will be perfect, but if you're enjoying the results, then I encourage you to post some images, if you can.
It helps creators on the platform a lot and encourages them to continue making content.
Thanksđź‘‹
kohya-ss/sd-scripts does not supports prompt weighting for training (at least for now), so no need (and maybe harmful) to escape parenthesis (\( and \))
Oh, is that so? I was wondering about that myself.
I honestly never checked where the escaping of those characters even came from, so I thought it was the same thing with training.
I've trained a few models before with character names and they usually worked well, so I think it doesn't influence it too much (I suppose it just takes it as a literal string).