-
Welcome to the Community Forums at HiveWire 3D! Please note that the user name you choose for our forum will be displayed to the public. Our store was closed as January 4, 2021. You can find HiveWire 3D and Lisa's Botanicals products, as well as many of our Contributing Artists, at Renderosity. This thread lists where many are now selling their products. Renderosity is generously putting products which were purchased at HiveWire 3D and are now sold at their store into customer accounts by gifting them. This is not an overnight process so please be patient, if you have already emailed them about this. If you have NOT emailed them, please see the 2nd post in this thread for instructions on what you need to do
I Just Wanted to Post an Image Thread
- Thread starter Stezza
- Start date
The idea is definitely not to replace 3D renders, but instead how to use it together with my renders, like for example, backgrounds. I am pretty good with integrating 2D backgrounds with 3D scenes, trying to make them seamless.
I don't know, maybe it could. Basically any image can be used with IBL, so I think that yes, AI could be used to create IBL probes. I will give that a try.
Sublime Text is my Python editor, but since the AI is running on a virtual environment, I run it from a batch file in the Windows shell. I edit the batch file to change the prompts. The shell is running from the Python virtual environment. I tried configuring Sublime Text to work with virtual environments, but so far it didn't work. I have plans to create a GUI for the AI, so I can edit the parameters in a simpler way.
Sounds way above my depth. I have long ago let my abilities using shell and batch files go rusty... and some of my 3D skills are starting to follow with my workstation being down. I can't even remember how I wrote some of the erc for some of the Poser figures I had a hand in coding or modded for personal use.
Maybe I'll try to read up on it if/when I finish trying to recover this hard drive.
Maybe I'll try to read up on it if/when I finish trying to recover this hard drive.
I get you. StableDiffusion AI is not for everyone at the moment, and mind you, it has just begun. This is all very new and recent. Their very first PT model just became public a few days ago - that's how new this thing is. It is the world's first free and open source AI of this magnitude, in defiance to the mega corporations (Google and nVidia) that currently own the most advanced AI available (Dall-E and Imagen), charging people to use it by the minute.
StableDiffusion aims to democratize AI to everyone, but at the moment mostly to IT enthusiasts like myself. However, the already offer web page access to this AI with a very simple interface, like you type a description just like in the Dall-E site. Not for me because it doesn't allow changing the parameters and tinkering with the code. Just in the last couple of days, I have already modified almost half the code to make it fit my pipeline. Another guy has already rewritten the main code to reduce the model load, so it can run even on GPUs with as little as 4GB VRAM, though slower. It's truly open source.
Here's my latest test, asking for a sandy Caribbean beach with coconut trees. Here we can see some image incoherences, like some floating trees. Dall-E had these same problems in the beginning, so it's just a matter of time until they fix it. I could definitely use these images as render backgrounds, and even as IBL probes as well. Though I would probably darken the lower halves so that the ground won't glow.
StableDiffusion aims to democratize AI to everyone, but at the moment mostly to IT enthusiasts like myself. However, the already offer web page access to this AI with a very simple interface, like you type a description just like in the Dall-E site. Not for me because it doesn't allow changing the parameters and tinkering with the code. Just in the last couple of days, I have already modified almost half the code to make it fit my pipeline. Another guy has already rewritten the main code to reduce the model load, so it can run even on GPUs with as little as 4GB VRAM, though slower. It's truly open source.
Here's my latest test, asking for a sandy Caribbean beach with coconut trees. Here we can see some image incoherences, like some floating trees. Dall-E had these same problems in the beginning, so it's just a matter of time until they fix it. I could definitely use these images as render backgrounds, and even as IBL probes as well. Though I would probably darken the lower halves so that the ground won't glow.
LOVE all the new images. Too bad the AI stuff is just for folks with the mind to understand how to make all that work. HOPEFULLY there will be easy to understand versions coming out at some point. Be cool to have this plugged into Poser and then just create your background for your character renders. The ideas and outcomes would be endless!
Give it a try HERE, where they have the public access demo.
Exactly! This could be an endless supply for background images to use in renders, and perhaps IBL probes as well.
Well, I tried it and it gave me four 512x512 images, which is nice because they are totally different. I now want to try a single image, so changed that setting to 1, but if I want a large image, what setting do I change, the Guidance Scale? By default it's set to 7.5, but I'm not sure how much larger I would have to make that.
OK, well I tried a Guidance Scale at 15, and it still gave me a 512x512 image, but it's not exactly the same as the previous one so that is a setting to play with. I'll upload a few of them in a bit.
OK, the keywords I used were: Sunset, Ocean, Waterfall.
Here are the group of 4 images I received the first go-round:
The first image looks strange because the "sunset" is lodged in the middle of the mountain side with the waterfall flowing over it.
The second image looks as though 2 separate images were composited together. Each would probably look good as a separate image.
Now the third image is really nice, though there's no waterfall, just water "falling" down from the waves. I'd use this as a background for sure.
The last image is more or less what I was envisioning when I chose the keywords. Also a background I would use in a render.
Here are the group of 4 images I received the first go-round:
The first image looks strange because the "sunset" is lodged in the middle of the mountain side with the waterfall flowing over it.
The second image looks as though 2 separate images were composited together. Each would probably look good as a separate image.
Now the third image is really nice, though there's no waterfall, just water "falling" down from the waves. I'd use this as a background for sure.
The last image is more or less what I was envisioning when I chose the keywords. Also a background I would use in a render.
Then I decided to try a single image with the keywords Sunset, Ocean and Waterfall, and this is what I saw when it was finished processing.
Needless to say, I found this amusing with even more of the sunset showing in the middle of the mountain.
So I deleted the Sunset keyword, and this is what I saw with just the Ocean and Waterfall keywords.
Then I decided to play with 2 of the settings, and this is what I saw, and I like this the best.
Needless to say, I found this amusing with even more of the sunset showing in the middle of the mountain.
So I deleted the Sunset keyword, and this is what I saw with just the Ocean and Waterfall keywords.
Then I decided to play with 2 of the settings, and this is what I saw, and I like this the best.
Lastly, I decided to do a single image with the Sunset keyword to see what I would get, and here it is.
Not bad at all. So I decided to pull it into PS and see what would happen if I resized it, and here's what it looks like at 1024x1024 with no adjustments, just resizing.
Not that different, at least not that I can easily see.
Not bad at all. So I decided to pull it into PS and see what would happen if I resized it, and here's what it looks like at 1024x1024 with no adjustments, just resizing.
Not that different, at least not that I can easily see.
Trying it now. Not sure why the images are stuck in such a low resolution. I put in Hairy Man and it gave me 4 images, mostly of just bearded male faces, there is one torso shot with a head and again, a beard but if you zoom in the eyes are kind of a mess! LOL
The "scale" parameter refers to how close to the description we want the image to be. The default is 7.5, which is a good value for most images. I guess you already know that, but this is not related to image size. I am not sure they let you change that on the demo site.
No, I didn't see anything that allowed you to choose the image size, but that Sunset render I did enlarged worked well, so I'm assuming without going too much larger than the original isn't going to cause much of a difference in quality.
I don't know about Photoshop, but Paintshop ships with AI upscaling that works really well. StableDiffusion also has its own AI upscaling, and part of the code to make use of it is included, but the PT models and some dependencies are missing. I think they need to be downloaded and installed separately, but I can do it in Paintshop, so I don't need it.
This is what I was saying way back at the beginning - the public web demo is very basic. From Python I can change any parameter, but from the web site, very few. This is all brand new, so they might improve this over time.
Yes, I remember you mentioning that about PSP. Unfortunately my copy is very old, as I quit updating it around version 9, and only have version 7 installed, which I use for quick screenshots and such.
I looked through Photoshop, and didn't see anything about AI upscaling, and again, it's not the latest version, so it's possible newer versions of PS do have it.
I looked through Photoshop, and didn't see anything about AI upscaling, and again, it's not the latest version, so it's possible newer versions of PS do have it.
