- This topic has 39 replies, 12 voices, and was last updated 3 years, 5 months ago by
.
-
Posts
-
@warrant63 Thanks a lot 🙂
@ethiopia In it's current form I don't think that you could extract a riggable 3D model from it, because it's almost impossible to get every possible angles as source material for model training. So "the AI" is not 100% sure how it exactly looks like in full form - because you didn't "teach it" ( maybe a partial extract then Zbrush the rest could work, but I'm not sure in this part ). Btw this is the main reason that many DeepFake doesn't look that good from "extreme" side angles
I worked a lot with DFL, and I usually dig up every possible good material for my models (even from Russian torrent sites), but I couldn't cover every angle this way either 😀
This is the diagram from MVE for a few facesets ( the brighter blue means more material for that angle - it will look good // the white border on black means only a few - it will be so-so // and the full black is mostly 0 - it will look meh ) - The center of the diagram is standing face to face with your model, center top model looks up etc etc
Megan-Fox
Ana-de-Armas
Kate-Upton
And like at Stable Diffusion, DFL is still a bit low res to use it without any upscale ( a normal DF face model is usually 320x320 or max 384x384. A good 512x512 model is already on the RTX A6000 48GB's territory - if you dont want to train it forever 😀 ). If you trained a good model you can use it almost anywhere and on any media (ofc it depends on the destination material where you want to insert it, face size/shape/angle etc - it can't do a miracle if you want to insert a narrow face onto a wide shape).
Btw in my experience, only studio photos are not enough for this 🙁 ( The Kate Upton diagram is a great example for this, luckily she did a few movies/shorts )
It's a great detail quality booster tho ( because of the high res ), but you need a lot of pics with different expressions in different yaw/pitch ( + in different light conditions ), the more you get, the more it will add to your model's quality/usability.
IMO the best source in order: 1. TV series BluRays/Rips 2. Movie BluRays/Rips 3. HQ Interviews ( Vimeo is always better ) 4. Photoshoots/Getty ( again for detail quality ).It's good fun btw and it's not only for NSFW stuff ( a lot of people tend to forget this ) 😀
For anyone interested in a dead simple GUI install on their computer try out this one:
https://nmkd.itch.io/t2i-guiInstalls just like a regular program, you don't have to go through the more intricate 1111 method. It's a nice quick intro to AI so you can see quickly if your PC runs it well enough for you to want to bother installing 1111. Some people even prefer this GUI to 1111 or use both for specialized circumstances.
It's updated fairly consistently and while it's missing some of the bells and whistles of 1111 that mostly seems to be because the developer is waiting to see which of several features become must haves or have settled on a specific method of implementing that feature. It's the wild west right now so I don't blame them for not coding in every shiny new toy that might get changed radically or tossed within six months.
There are some celebs who are very particular about photography angles, favouring one over some others which adds to the problem.
🙂
It's interesting how with the exact same prompts and setting every image produced is different.
Yeah, that’s why I was wondering how it can create a series of say the same subject or objects. It feels a bit random atm.
Odd that you're getting different images from the exact same prompts unless the program is changing the seeds each time.
@diviniren Having played a bit more with the AI, you can get the same subject multiple times if you are referring to a famous person, or you could train the AI to reproduce always the same person.
However, each time they would have different clothing, since the AI makes that up each time so, if you are looking to make a story, it's definitely not for you until things change.
@jrnkmail strangely, that can occur. So far I can generate dozens of images from the same settings and prompts...so maybe there is a hidden factor (maybe the AI is assigning a different weight to each prompt each time).
@legolas18 Interesting. I wonder if it's an 1111 thing. I always get the exact same picture when using NMKD gui, unless I start tweaking the settings in some way.
I've been playing with SD for four days now and a word of warning - it's killed (hopefully/maybe temporarily) my interest in DS and the process of DL content, setting up scenes and waiting for them to render.
The thing is if you want to make something with SD that is specific and intentional as opposed to “Naked Vicky In A Temple With A Sword” it takes a lot of time and effort. Any idiot can and do generate a "nice picture", post it and say aren't I great. But after four days I've realised that the creative process is just as taxing as DS if one wants to produce something they set out to do, instead of images, while elaborate and very pretty, are nevertheless directionless soulless dross.
If I see another picture of some cutie princess covered in vines in a magnificent baroque dress floating in a forest I think I'll fucking vomit.I'm glad you mentioned "figures interacting with the scenes" because in a nutshell that's what SD is very bad at. The idea of taking DS scenes into img2img sounds very interesting. I'll see what I can do with that.
This Stable Diffusion looks interesting, can it (or any other software for that matter) do something like this: take input from a standard photo image, and alter the head position. Such as look to the side or look straight forward. To change head direction would have to use some kind of AI I suspect, since an invisible side of the head is unknown, and would have to be "assumed" what it looks like and then created before it could be produced.
@ethiopia I see, Thanks for the detailed explanation. Hopefully someday in the future this technology will exist without the need for trainers.
- You must be logged in to reply to this topic.
