For me download NMKD stable diffussion off itch.io, get it all set up , go to the settings and download what you need to. Put my LoRA file into the lora folder. Place a good CHKPT file in the chkpt file. Now, set up the positive prompt with crash bandicoot, as one prompt unless you wanna get detailed with your prompt. Honestly, there's no real secret or trick with my LoRA models.

First, see the create tab to the left of the download button? PRess that an look for a option that will allow you to use the LoRA model online. Now, I use NMKD Stable Diffussion Gui from itch.io. To to use that download the program, Go to the wrench icon and download the upscalers and junk. Put, the Lora in the (root folder)/ models/ LoRA folder. An place a chkpt file in the model/ checkpoint folder. Close, and reopen the program. Go back to the wrench icon an look for the option to convert a chkpt to safetensor option. This will make renders take alot less time. When, the conversion is done. Select the safetensor in the chkpt wind on the left hand side. In the positive window put crash bandicoot,

Now, in the negative prompt put in the negative inversions like badhands, . Than, in the LoRA window click the crash bandicoot box. Now set the resolution and other settings. Than press render. Depending on your gpu a 720 p image may take 30-40 minutes if you use 20-30 epooch steps.

I hope that helps. I know its kinda confusing but, thats how ai art is for local hard ware users .

The first thing I do it upload a zip folder with a character or style I wanna go for. Grab my hugging face access key than do the basic steps and custom steps on this google colab https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/kohya-LoRA-trainer-XL.ipynb?authuser=1#scrollTo=_u3q60di584x

I train a model for about 4-5 hours. I think the SDXL Lora training is ckeckpoint safetensor based. But, converts the model to a LoRa format. The program I use for my lora models is called NMKD stable diffussion gui on itch.io. The setup is simple enough. But, you need to use the wrench tool to convert any chkpt models you have in the chkpt folder from safetenser to a safetensor.diff file . That way the rendering proccess can go faster. The fid model will essentially act as the brain. When you put the lora model in the lora file. Close the program and reload it. The model will now show up in the lora files. When click this will allow the lora model to be usable. If you use crash bandicoot as a positive prompt . The program will try to draw him . Based on 5 images I used as refference.

Because, I don't have a cuda gpu I'm using a invoke ai methoid to render images. Honestly, the proccess I use to make models is very particular. So, sometimes I spend 3-4 hours on e621 , rule34, or deviantart looking for a base image and a consistant art style. Otherwise, the model doesn't turn out right.

Personally I can't use or make LyCoris models. So, I can't really test how the model would work on server based programs. Since, all art is rendered with local gpu and ready boosting. That way I have more control over what images get rendered. When I render stuff on google colab . Thats generally when images turn out worse. Sorry I know thats alot but, my proccess of doing ai art is very specific XD I would try to make a more easier model for other programs to use. But, I would have to start drumming donations or something to get a more moddern computer to train models for different formats.

But, ease of use. I just stick with LoRa models and recommend ai beast furry mix and indigofurrymix for the base chkpt models to use along the LoRa models I make.

@Jcbq Thats odd on my computer I use SD 1.5 and SDXL models and my LoRa models work just fine with both model types. Mybe it's just how some programs or servers handle models that make them act or work differently. Than again I have been able to do things that really shouldn't because, I usually edit my computers like crazy to get some extra milage out of programs XD

I might have to try LyCORIS one of these days . Is it even good?

@Jcbq Wouldn't it just come down to . What art you used, how detailed or stylized . Because, I've seen people have issues trying to make character specific models with sd , lora, and other models only to get bad results. So, unless LyCORIS holds more training and image data. I don't think it would be much of a improvement. Unless it's visual scaling stuff which SDXL already does.

<3 glade you like him. I been meaning to do crunch and tiny. But, never got around to those models for some reason XD