- This topic has 13 replies, 6 voices, and was last updated 1 year, 5 months ago by
.
-
Posts
-
I wanted to play around with some renders and transform them with the img2img function of AI to show you how you could enhance your creativity.
You might want to open these images in a new tab to see them better. 🙂
Enjoy!
dForce X-Fashion Summer Girl Outfit for Genesis 9
and finally a random render I did some time ago:
Very good demonstration of how to turn a Daz doll into a real looking human and the hair/clothing look 100% better. I gather for best results you would use prompts similar to the original image, pink shirt, denim shorts etc. Though prompting may not even be necessary at low denoising levels.
My DS renders are usually a bit more complicated than portraits but img2img can transform them, if you know how to use the GUI parameters. For this I had to use several positive and negative prompts for it to work properly.[click image for full size]
Hey! That's what >I< wanna do!
Too cool. Great results. How neat.
Man, what kinds of prompts/non-prompts and settings are you guys using.
And is there any online-hosted equivalent to your local installs?
(Guessing I'll have to kick down and install locally to approach what is happening here.)
Super interesting and a completely different realm and level to using DS.@spats I am running a local install of Automatic1111, because of all the extensions I can use in it. There are other User Interfaces of course, such as ComfyUI (node based) and Invoke 3.0.
In terms of prompts/settings that's a huge topic, and I cannot give you specific settings because of the semi-random nature of the process (in other words, I do have to try out several settings every time, depending on the image).
However, images created by Automatic1111 tend to have all their settings embedded in them, so you could start those as a starting point.
@frank22 That looks pretty good! especially because it's a more complex image.
At this point I think that AI could be a better render engine than any traditional ones.
There are already hybrid ones to better calculate/predict extremely complex elements like gasses (those always gave me nightmares in DAZ) that also work super fast and with no noise.
I agree too, the result is amazing and pretty nice.When will we be able to do such images in DAZ ?!!!
Is it just a problem of PC Power, the iray limits or what else ?My stuff is mostly related to NSFW/Anime and/or "celebrity" renderings. The only real way to see how well your work actually works is by using a well-known character and see if it looks realistic. Anyone can produce some random gal, but only by a comparison with a universally known real-human can test the limits of the uncanny valley paradox. SD can achieve that but not DS as its limitations are tied to polygon limitations, poor lighting, poor clothing and the overall 3D sameness of all its mung models.
Unless DS adopts some elements of AI I don't see it lasting much longer."...images created by Automatic1111 tend to have all their settings embedded in them..." Yep that's the truth. Even with no negative or positive prompts a checkpoint will produce some sort of random woman because of its training.
@Spats - as with DS or any other prog there is no substitute other than getting you hands dirty and experimenting with it. Setting up a local SD installation is easy to do and with 8GB of VRAM you can do pretty much anything up to about 4K, even larger if you know what you're doing. You'll be surprised how you'll get the hang of how it works. Lots of random stuff happening but hey that's life isn't it ☺.This looks pretty cool. Indeed ..time DS implemented such rendering capabilities.
gues I will have to start looking for an " AI for Dummies" 🙂This is quite interesting, on topic:
✨.Daz studio to ai stable diffusion... - YouTube
h**ps://www.youtube.com/watch?v=zYCv6oC3xf0At first all I saw and thought was click bait, since the results were similar to what I'd been getting with hosted SD. But then the multi-step, multi-filter, multi-generate and refine and inpainting masking became clear. Finally getting what you guys were saying in this thread and others. It takes more than just a single generate and go mentality. Refining is still true in img2img, or any other aspect. (A lot of the specific tick boxes and numbers seemed important throughout, but the utilization of the seed number copied across for the inpainting at the end seemed key and made sense.)
Still don't have a local install, but am now itching to use the public hosted in different and more extensive ways to see how close it comes.
Neat.(P.S.: Post-watch thought somebody here could have easily posted this video.)(Then thought AI could have, too.)
P.P.S.: Also noticed that in the Negative Prompts in the vid there was "(((no cap)))" included. Qs: Does this uncap the processor to, in effect, unlimit processing power and duration? Isn't a "no cap" in the Negative Prompts field considered a double negative? Subsequently limiting processing power? "No, no caps = Caps"
(Not relevant to my online purposes and thus not included or used in the tests I'm doing. But, just curious.)Thanks Spats. I gather that had more to do with ControlNet posing rather that actually making the Daz models look better. Even with ".. most beautiful fitness girl in the world" LOL, they ended up looking just like Daz dolls.
Not making headway with webui versions of SD.
Reminds me of DS newbie days and "not getting" things and being frustrated and overwhelmed. Thinking that progress is happening, then realizing it's not.
Fear will have to bite the bullet and go local. (Which in my and perhaps other's circumstance is disk-dependent.)- STABLE DIFFUSION COMMIT EQUATION:
"DAZ Commit": The condition in which the end user cannot, or is not willing to, quit or leave the DAZ Studio platform due to the amount of disk space devoted to the runtime installation and environment. Alt: The size of DAZ Studio runtime on disk as a percentage of usable disk space (DAZ Commit Size).
"Stable Diffusion Hesitance" (SDH): The hesitation or inability to install, locally, the Stable Diffusion platform due to limited disk space, often due to a bloated or overgrown DAZ Studio runtime.
"Stable Diffusion Commit" (SDC): The willingness or commitment to alter disk allocation or size, or DAZ runtime allocation or size, to change DAZ Commit Size enough to make space for local install of Stable Diffusion.Given SDH > 0,
SDC = (100% - (DAZ Commit Size)) / SDHI didn't have a problem with the local install apart from having to point it to the Python.exe. I have my DS runtime on an external drive.
On disk space - with the new SDXL, LoRas can be huge (500MB) compared to around 5-20 MB with 1.5. Worse still, the SDXL models (checkpoints) are around 6.5GB each FFS. Unless you've got a seriously big HD/SSD your going to fill it up pretty damn quick. I won't go to SDXL and I'm now in the habit of one-in-one-out, deleting an old checkpoint before adding a new one. A lot of them are very similar anyway.
For install I followed this tute (but 3 minutes is bullshit)-I've seen a half dozen install videos. (Including above linked, btw, but thanks just the same as usu for me and others @frank22.) All seemingly easy to follow and relatively straight forward (when you pause/slow them down for likes of me), save a few foibles that they even cop to in sub-video or notes.
But, of course, you can't fix "space-id". I, for one, have allocation issues. Bumping up against the space ceiling is unfun for a list of reasons, not least of which is performance issues when full due to pagefile clogging or restricting, of course. I've been sitting on an unopened 2TB ssd upgrade for months now but haven't "committed" yet, is all. As soon as that image-and-expand happens I have another full TB to work with, directly; plus the old 1TB drive if I chose to use (longevity permitting) for external DAZ runtime, basically doubling the expansion effect.
...Just a matter of time, is all. Life gets in the way of living.
Someone buy me an untouched week of time in dimension X.
- You must be logged in to reply to this topic.
