After a while I finally got the hang of it. It's actually much simpler than I expected. In this case here, you do not really need to describe enviroments very well,
just adding detailed decoration and detailed room should suffice. I thought I had made a mistake by not training it with same types of rooms instead of
general and varied environments. Actually that was what make it works easier than expected.
And as for people, no need to really add 1boy or 1girl, unless that's exactly the number you want, just adding mini people suffice. I forgot that I
added as caption such tag as well as mini world. Also even if it won't be visible, adding a background, as forest background improves the quality a lot. So,in short,
this LoRA works differently from many isometric LoRAs (or not?)
Alright, this is my second attempt at training a Low Rank Adaptation model for ISOMETRIC STYLE for Noob AI EPS 1.1
It was really trained on NAI EPS 1.1. I wish I could train it for V-Pred but I have no resources neither skill to train for it. Sorry!
It DOES WORk with V-Pred 1.0 TOO! YAY! Although V-Pred base model will always add a white background.
So it's a bit lame, but fine-tuned models perform quite well. I have a lot of cute images which I might add here later. So far I added those 3 new samples above, which show
that there's no need to describe everything - actually you can lazy prompt as mentioned above.
Then I deleted those extra samples below, as they were nice when I finished training, and I was happy for the LoRA to be working, but I was not so sure about what I was doing there.
Even now I would have to learn. Thing is NAI works quite differently from SDXL. And then those quality and negatives MUST be used.
For fine-tuned models, Euler A usually works quite well. For NAI EPS 1.1, you can try others. Raising CFG beyond 6, and too many steps does not seem to improve much.
So about 5 to 6 CFG, and 20 to 25 steps. 30 steps if you're rerolling he same seed and want to check if it improves.
For NAI V-Pred it works better when it's outdoors and dimension is 832x1216, actually just like my Kawaii Chibi Avatar, V-Pred allows immense freedom,
although it requires prompting (reads thinking, unless you're super used to it and how to pick perfect keywords out of tons of booru tags,) and some "emotional personal tags and senteces,"
as NAI models are very focused on character subjects,and if you play well around them, the model opens the doors to unusual and unthinkable things related to dioramas, isometrics
miniatures and all the likes.
NAI base models such as any others (even SD 1.5) are able to create isometrics from their base training. Yet they need to understand the meaning of some extra tags such as
"isometrics" per se (if they ever do! lol.) Then, if you're used to NAI V-Pred booru-thinking, you might have a lot of fun.
If not, whether just pick a V-Pred merge (depends on what had been merged,it won't work) or EPS fine-tuned and or merges (again not all work the same way.) All those furry/anthro focused models tend to work quite well with all my LoRAs, not exactly that I do anything better than anyone else, but it's on their end, as
they tend to be excellent versatile models with high capacity of understanding and adherence to input.
isometric view, from isometric,from above,tree,forest background,
houses,detailed mini village,forest background,plant,boat,river,
harbor, mini anime +2boys, mini passengers, 3d,salaryman,
building, shop,,water,,, very aesthetic, absurdres, newest,
perfect quality, original, amazing quality, masterpiece, very awa
isometrix (10).png
isometric view, isometric,motor vehicle,beach umbrella,ice cream,beach background,
chair,truck,ice cream cone,shop,sand,cooler,
beach background, tourists, mini people,food stand,
very aesthetic, absurdres, newest, perfect quality, original,
amazing quality, masterpiece, very awa,
isometrix (8).png
isometric view,isometric,from above, indoors, indoor pool, club, banners, water, tiles,
detailed decoration,no humans, window,day, sky background,,
very aesthetic, absurdres, newest, perfect quality, original,
amazing quality, masterpiece, very awa,
isometric view, stairs,isometric,tree,forest background,fence,sign,window,scenery,house,mini village,
forest background,plant,rail, train, train station, mini anime people, mini dog, mini passengers, , railing,door
building,,,bush,water,chimney,bench,,, very aesthetic, absurdres,
newest, perfect quality, original, amazing quality, masterpiece, very awa,
MSPA8Y5G2H1PCFBEN53QQ3E2E0.jpeg
Reason? Maybe I have not searched enough (I'm lazy so...) but I have not seen any isometric LoRA for NAI, maybe there's only one for Illustrious. My guess is that IL/NAI are
a pain to train styles not related to characters. For instance, I trained Miami Beach Hologram for NAI EPS and it was a total disaster. On V=Pred images were so dark and messy
that I rage-del it. That famous 1980s car on the water, in the sky, everything so messy, The only good thing it had was being able to write Miami properly, but hey! IL/NAi are not
that bad at small words, especially those they don't know as they always print them as watermark or artist name.
Well, NAI is able to create isometric view on its own. But it does not work all the time and I don't really know which setting should be used.
Then let's say this is an experimetal LoRA, which is offered as is and it's user's responsability. IT should mainly be used for research and educational purposes.
I have not tested it very well yet, but it seems to work well with Euler CFG ~5 and Steps 20-ish.
The first attempt was a huge... disaster. NAI learnt simply NOTHING from it. And for this one I used a more aggressive setting and it seems it worked you can see in the sample images.
Those images where from the training preview. There's nothing extra, no hires and no cfg boosts neither any other LoRA.
If it will work on Civ, only time will tell.
It was trained with generated images from (as usual) Flux and SDXL itself (which is amazingly good at isometric on its own.)
Good luck! :3
Ah right, it's licensed under CC 4.0
Trigger words
You should use isometric style to trigger the image generation.
Download model
Weights for this model are available in Safetensors format.
