- This topic has 37 replies, 23 voices, and was last updated 2 years, 11 months ago by
.
-
Posts
-
Yep, finger/hands often look terrible, feet/toes are actually much worse but less noticeable. When I first started I also got lots of people that looked like something from The Thing 😮, but with practice that's less.
I often get images that have that "AI look" and obviously not real, looks airbrushed, especially with multiple character scenes and particular lighting.
I LUUUUUUVE it!
I'm amazed at how fast it's advancing, not just the tech but how we use it. I just got new scanner software that promises to eliminate the moire effect left from the scanning process without the traditional blur-n-sharpen that softens detail and still leaves a bit of moire residue.I'll admit on record that I have previously said dismissive and very negative things about AI art. This was before trying it.... in fact most of those who diss it have never tried it... does that tell you something?
With the finger and limb issues etc, it's still not there. It's really not a "make art" button. Out of 75-100 images I might get one that is decent enough to import into img2img then do some inpainting, then open in photoshop and use my drawing monitor and pen to perfect to what I believe is good enough. What I use it mostly for is 3 things. I use IMG2IMG to produce variations on a theme of my renders, then spend time in photoshop layering and blending and tweaking on top of the original. I use txt2img to produce textures. I also use text2img to produce original renders to give me an idea of what to create render in DAZ. I've even used it in IIMG to IMG on my render to create clothes I don't have and don't exist for daz, use photoshop to crop out the unneeded parts, then layered in daz over my render and fixed with my art monitor so it looks like it was rendered in daz. None of it is a "mark art" button though, creating clothes to apply over my daz figures takes a hell of a long time and many generations to get to look right then time to work in photoshop.
What I really like about it, and what I really wish DAZ or Blender could incorporate, is how Stable Diffusion handles soft body physics. In Stable Diffusion, tight clothing often makes a very convincing indentation in flesh. While some PAs add this effect to their DAZ assets with character morphs, it's custom and often not as convincing. Hand grabs of soft tissues make indentations, pressure & stretch marks sometimes very realistic in a way that's simply not possible in DAZ without post processing. Clothing stretches make realistic tension wrinkling - much better than dForce. With a simple example of a bikini for example, in SD there's less of a chance to see a bikini string between the breasts unrealistically following the curves of the beasts rather than having the tension pull it straight. Even through dForce was a step in the right direction, DAZ still seems to do this kind of thing a lot.
DAZ still has its place, IMO. Interactions between two or more characters are way too predictable and routine in SD. I think some posing is possible in SD, but I've not gotten to that yet. Though I have no experience with posing in SD, I suspect with posing it's not a full replacement for DAZ - at least not today. The posing still seems to gravitate back to the trained images for the checkpoints / LORAs rather than having each posed interaction be a new and unique interaction.
Great summaries Oldwolf and Spectrum. Accurate posing can be achieved by using OpenPose and ControlNet but I've never tried that myself. "Out of 75-100 images I might get one that is decent enough.....", so true. The way things are advancing I'm sure that many of the current issues with SD will be overcome. What worries me is that Civitai will be sold off and commercialised, they've already had offers. In that case ZGFX might become a source for AI content and not just DS/Poser stuff 😃.
In that case ZGFX might become a source for AI content and not just DS/Poser stuff 😃.
*chuckle* would love to see that happen.
- You must be logged in to reply to this topic.
