Meet Mort Nobody!
Mort is a Nobody, an AI OC, a character not based on any living person, in personality or looks. Mort may LOOK like somebody you know, but he's a Nobody.
Mort was a bartender when he was younger, and had a side hustle as a minor mobster. He still prefers nice clothes, especially pinstripe suits. He likes dogs.
He's happy to be the bad guy in the background, or the boring boss, or the grumpy father in law. He's also happy enough to take off his shirt, though I didn't include any NSFW images, so you'll have to ask him nicely.
Tested in many models, he works well in most of them, but I tend toward RealisticVision and Protogen x58. Anime, painting, photorealistic, comic, Mort can be whatever you need.
Description
First version. Still working out some kinks, he tends to put on a nice suit if you don't tell him to dress casual in some way. 2.1 version coming.
FAQ
Comments (21)
Looks like some Clancy Brown in there (first shot an couple others).
Ha! he does!
Have you written anywhere about how you mastered that char tuner style image here? Is the key controlnet + open pose? Or are you using a TI or LORA still?
Great work!
Not anywhere public other than a couple of reddit comments. I keep meaning to set up a blog to write stuff out but am having more fun making stuff. XD I'm using Controlnet AND CharTurner; controlnet holds the poses, charTurner makes sure it's all the same character in the same outfit.
Thanks, I like the description!
I should try it when I have some free time ;)
Very good ! More TI realistic characters !
Any requests? Haven't decided on my next Nobody yet.
@mousewrites What about the family theme ? After the grandpa, a granma. Then father and mother. And last a brother and sister, to complete the Nobody family.
@ritcher1 Yeah, that was my rough plan. There's a whole bunch of Nobodys already, but most are young women, so I'll do the older people. Grandma, papa (heh), momma. I might make a couple of kids but I don't think I will release them, because people are WEIRD and I don't want to be part of it.
@mousewrites A family without kids would be weird. Let the siblings be 18+, since the platform have a strong policy about it.
Love it , keep it up. More People of Color and Varieties of People
Thanks! I agree, the lack of the wide spectrum of humans is a problem. Hopefully will release a new Nobody soon, though I suspect it'll be buried in the sea of conventionally attractive young women, (real and otherwise) but I will continue! Your comment means a lot; poor Mort didn't get a lot of love.
@mousewrites When you give Mort a relative, please record the process somehow so it can be used as learning material. @julieisdead sorry to highjack your post. :-)
How many images did you use to train it? And how did you manage to get consistency on generation and multiple poses (I'm assuming it is the turnaround) in order to be able to get enough material?
Final dataset was 35 images, but you're right in that I start with a turnaround, and cut it into individual shots (ie, 4 turns, =4 images), and then usually make a copy that is just head up of front, side, 3/4. Dataset is now 7. Train that, that's the v1. V1 is very limited, but you run like, 20 sets of images, pick out the 5-6 that are the closest, inpaint them to be what you want. now your dataset is 12. Train again, repeat, filling in 'holes' (ie, if you want it to be able to do photo and anime, it's best to add a few of each.
Note, I trained him mostly with transparent background pngs for the final set. This is so I didn't have to worry about captioning his background as much, as the 'loss mask' thing takes the non transparent as "learn this" and the transparent as "I don't care about this". If you train a LORA, you can't do that, so make sure your backgrounds are well captioned and VARIED, or you'll have a tendency to invoke the background environment.
All of that being said, a bigger dataset will reduce "preferences"; with 35 images, Mort likes his SanFranTokyo city, and his pinstripe suits, and dogs, which if I do a new version I need to address. :D
Thanks a lot for your reply. I'm just starting so I hope you don't mind about a few questions. Please forgive me if they sound too noob. The first one is, I'm aiming for just the face(right?) so, wouldn't it be better to train the nobody with shirtless portraits (closeups) so the model could take care of the clothing? I mean, if I train it using a suit and a tie, I believe it would be hard to remove it later, right? The second one is about the transparent background (can we use transparent png files for training or does it show the checkers pattern?), wouldn't it be also easier to use a plain flat black background so it would be easy to caption, and not used as noise by the model? I mean, I don't know if this is how it works but I think about that like alpha channels, so I'm sending #000 black I'm actually adding nothing, right? (I'm just trying to avoid one step, to remove backgrounds). Actually, I'm trying to find the right pipeline/creating openpose templates for the turnaround generation.
@Scofano We all started somewhere! :D
A1) Technically, in a perfect world, you should be able to just give it a single set of face turnarounds and it should be able to make that face consistently. However, in practice, if you only train on closeups of the face, you will struggle to get good pictures OTHER than close ups of the face. If you're going to use this model only to inpaint faces on images you've already made, this MIGHT work, but even so, it will 'pull' toward exactly the images you gave it, especially if they are all 'alike' in that way. Essentially, you're unintentionally teaching the AI to prefer close up face images when you call this embed. You can stop this some with captions, but even so, if every image in your dataset is a close up face, it will want to give you close up faces.
A2) I am not only training a face when I train a character. We are more than just our faces; body proportion, age, other physical aspects of us from the neck down do come into play. If you only do the face, you would have 0 consistency on height, weight, musculature, let alone body posture/attitude.
A?) with those points in mind, you need to choose if you want to have one set costume (ie, you always want this character in their super hero fighting outfit) in which case you don't caption it, and the ai learns that this person always wears X,Y, or Z. OR you put them in a bunch of outfits (no repeats), so it learns that the person is the "same thing" in each image, despite the outfit. OR you tag your outfits, if you have a small data set and you have repeated clothing themes you don't want (ie, character is wearing tshirt in 4 images, i'll tag "tshirt" so it doesn't try to pull toward that.
For SD, "Naked" is the same as "Outfit: no clothing" so you can't get around this just by making everybody nude and then asking for clothes. It will try to be naked if you do so.
B) PNG stuff: depending on how you save your PNG, it actually saves it as white or black (or the original background!) and has an alpha channel to control what you see. There are no "checkerboards" baked into the image, that's just photoshop and other program's way of visually telling you something is 'clear.' You 100% can just fill the background with black or white, but again, it will try to "pull" your images into having a solid black or white background.
Overall, the AI learns more than we expect it to. using the png loss mask thing HELPS, but identifying where your trained embedding is 'pulling' and fixing the dataset, the caption, or culling out images (ie, this image is great but it's influencing too much") is 90% of the work of getting a strong embedding/lora that is controllable as well as easy to use.
@mousewrites Super valuable information in here. Thanks a lot! Where can I find your Reddit comments/posts so I can learn/research a bit and stop annoying you here? :-)
@Scofano I'm Mousewrites on Reddit, too, though I post about more than AI. ;)
I keep meaning to start writing articles on here now that we have that option. Perhaps this weekend.
@mousewrites Following! ;-)
+1 for photorealistic non-supermodel people models