For those interested in getting started with AI art generation, give this a whirl.
https://lykos.ai/This is the link to Stability Matrix. It makes it extremely easy to generate art locally and it is what I still use today, although I have done some advanced modification to my install. You cannot do what I do with the available options, I had to manually install a few bits. Again, one day I will write it up.
Here are some tips:
- Don't use any of the checkpoints they offer during the install. I use CyberRealistic because the default checkpoints have a tendency to shy away from nudity. There are other checkpoints that are built on porn images, but those go too far in the other direction. If you select a checkpoint during install, don't worry. You can have multiple checkpoints. Just go to the model browser and select checkpoints from the dropdown and check the box that says "NSFW" to see more options. Note, some checkpoints do anime and drawn art better. CyberRealistic doesn't do drawn art very well, it tends to make it look more real. The fairy in my prior post was done in CyberRealistic.
- For efficiency select Automatic1111 during install for the interface. You can select ComfyUI to have more control over workflows, but it can get pretty "flowchart-y". I like the simplicity and speed of Automatic1111. Again, you can install more than one interface and select which one you want to launch each time.
- Browse the LoRAs. I downloaded Breast Size Slider (Makes boobs bigger, but also makes hips wider), Breast Expansion Lora, Cleavage Bulge, Undersized Clothes, Button Gap, Unbuttoned Shirt, Super Gigantic Breasts, and Hyper Breasts. Note, Super Gigantic Breasts really doesn't make them very big. Hyper Breasts has a tendency to elongate your subjects torso unrealistically, so I use it at a lower weight. VERY IMPORTANT NOTE: Your checkpoint will be based on a certain model (SD 1.4, SD 1.5, XL, Pony, etc). Make sure your LoRAs are compatible with the checkpoint you want to use. You have to click on a LoRA you are interested in because many have versions compatible with different checkpoints, but the menu will only show what the most recent version is compatible with.
- This is something I wish I had known before I ran out of hard drive space. The interface will store your images in MULTIPLE locations. If you go into settings you will see a path you can change (note, completely exit the interface, shut it down, and close Stability Matrix and relaunch for it to take effect, reloading the interface doesn't always work). This is not the only place it saves. You need to also look in C:\Users\<username>\AppData\Local\Temp\gradio or wherever you have your user data pointed to. You would think the folder would empty itself at exit, but it doesn't. At least mine doesn't.
- Final tidbit is on weights. The further down the prompt something is the less weight it will have, so put what is important to you at the beginning of the prompt. Also, you can impact the weight by using different code. I will list some examples below.
So here are some prompt examples. Note, LoRAs have trigger words. Read the description of the LoRA in the model browser to find out what they are. Once you are in the Automatic1111 or ComfyUI interface you cannot see what the trigger words are. You can go to the model browser after launching your interface and you can even download new LoRAs to try while it is running. You just have to hit the refresh button near the "checkpoint" menu at the top or in the LoRA tab to see the freshly downloaded models.
Prompt: 24 year old woman, wearing a blouse and skirt, ((((flat_chest)))), breast expansion sequence, <lora:expansion:1>
Negative: hands
(This example and the output are the attached image.) Note the (((()))) around the words "flat_chest". This impacts the weight of what is inside. I really want her to start with a flat chest and Stable Diffusion has a tendency to make all women busty. There are LoRAs for flat chests, but I can't use LoRAs with animations or all frames will have her with a flat chest.
So weights can be done as any of the following: (((gigantic breasts))), (gigantic breasts)+++, (gigantic breasts)1.6
Following a word with - will reduce the weight, + will increase the weight. Or you can put the words in parenthesis and follow it with a weight, usually between -2 and +2, but you can go stronger. However, going to strong will result in anomalies. Multiple parenthesis also add weight, essentially like doing 1.1^x, x being the number of parenthesis.
This link has some helpful info.
https://getimg.ai/guides/guide-to-stable-diffusion-prompt-weightsLoRAs are always called in <> and have a weight. Some LoRAs do better at lower than 1 weight, but most are fine around 1. If your LoRA isn't taking effect you might have to increase the weight based on what other LoRAs are with it. Note the stronger the weight on the LoRA, the more you are going to see the training data that was used. There is a LoRA out there for big breasts that used a man wearing prosthetics for training data, and if you raise the weight too much you get his face on the subject. Ugh. In Automatic1111 you click the Lora tab (to the right of the Generation tab) and click on the LoRAs you want to use. It will add them to your prompt at whatever default weight you have them set to. You can either change the weight by typing over it in the prompt, or you can adjust the default weight on the LoRA tab by clicking the hammer and wrench icon in the upper right corner of the LoRA thumbnail. You can also add the trigger words so you don't have to type them every time you use the LoRA.
Final notes, pay attention to what sampling method and scheduler type your checkpoint prefers. For CyberRealistic I use DPM++2M and Karras, but others have different preferences. Like the LoRAs for more Asian women use Euler and ESRGAN 4X for upscaling. It makes a HUGE difference and maybe you will want to experiment. I like to try new checkpoints with different settings. Also, the width and height of your image can also impact how congruent it is. CyberRealistic works best in 512x512, 512x768, or 768x512. I have experimented with 512x1024 for my BE sequences and it seem to do fine. If you want it higher res you use the upscaler. I use ESRGAN_4X with 15 steps, adjust the denoising strength to 0.25-0.5, and upscale by usually 1.5-2. I try to keep it a good size for uploading to the forums without having to convert to JPG. HOwever, you can create larger images then just use a program to change it to JPG and shrink the file size.
Lastly, if you like a subject but want to tweak some settings, there are two icons next to the "seed" at the bottom, a die and a recycle symbol. The die randomizes the seed each time and the recycle takes the seed from the most recently generated image and uses it again, useful for if you want to upscale the last image and clean it up. Also, you can play with all the other tabs. Img2Img is the tab where you can do inpainting to clean up things you don't like. The painters palette at the bottom of the output image will send your pic over there for you to clean up. I won't do an inpainting tutorial because I am no good at it. I actually tried to get rid of an extra finger and inpainting put an entire hand on the end of the extra finger...
Good luck, one day I will write up how to do animations.
Punchy