|
Socoder -> AI Threads -> Dall-e 2 |
|
| Thu, 08 Dec 2022, 07:12 | |
|
Pakz |
With the ChatGPT or playground I would now try and ask it to create a html file showing thumbnails from a list of image names. It can do lenghty code when asking it 'continue' I tried to ask it to create a html file to read the words on this frontpage here, but that did not do anything. I'm not a expert in why it did not work though. 
|
| Thu, 08 Dec 2022, 07:17 | |
Jayenkai
|
An html script would overwhelm the server in seconds, trying to server over 6,000 images, all in one giant go!! I'd have to make an SQL type gallery for them all, so that it can paginate them, and allow searching. -=-=- ''Load, Next List!''
|
| Thu, 08 Dec 2022, 07:32 | |
|
Jayenkai
|
I've spent the past 20 minutes or so searching for viable image gallery scripts, and every single one of them has me going "I bet that's a security hole waiting to happen.." "Not on this server!!" I think it'll either be a fresh secondary server, or another case of me writing my own. -=-=- ''Load, Next List!''
|
| Sun, 18 Dec 2022, 19:23 | |
|
Jayenkai
|
Well, there goes NightCafe then.. Fucks sake. They've shoved adverts on it, unless you subscribe. The amount of money I've spent on credits for that site, in the space of a year, and they still shove adverts on it unless I subscribe?! They can fuck well off... Tonight I installed DiffusionBee on my Mac (Works on anything M1/M2) and it seems to do a nice enough job. It uses Stable Diffusion, and seems quite capable.  "The Daleks have a day at the beach. Matte Painting" Though, it's not very good at upscaling my pixelart! NightCafe had that lovely VQGAN+Clip method that could really come up with some crazy "repainting" results. Might have to look into ways to do that. Meanwhile, DiffusionBee can be found here, if you have a suitable setup. -=-=- ''Load, Next List!''
|
| Sun, 08 Jan 2023, 20:12 | |
|
Jayenkai
|
    "Haunt hunt heart hut. Surreal airbrush art" by #DallE -=-=- ''Load, Next List!''
|
| Mon, 09 Jan 2023, 04:36 | |
cyangames
|
Here's a link to my lil selection from Nightcafe creator.nightcafe.studio/u/refreshgames I'll have to try that Matte Painting text on there -=-=- Web / Game Dev, occasionally finishes off coding games also! |
| Mon, 09 Jan 2023, 05:43 | |
|
Pakz |
Someone is working on a pixel art version using the diffusion model. This under the hashtag #AsepriteDiffusion The examples look promising. Google had been creating specific pixel art categories too. Stil, there is no real way yet? to create lots of good retro art for games now.
|
| Fri, 13 Jan 2023, 15:42 | |
|
Jayenkai
|
 Tried out one of those pixelart diffusion models, specifically this one. Click into the Files and Versions tab on the page above, and download the .ckpt file. Method - on MacOS using DiffusionBee Open DiffusionBee, click Hamburger Menu, Settings, and Import the model. Then select that model, and be sure to use PixelArtStyle in your request. Request : A beautifully ornate castle on a hill. PixelArtStyle Result > Reveal 🔎 -=-=- ''Load, Next List!''
|
| Sun, 15 Jan 2023, 06:56 | |
|
Pakz |
There is a lawsuit coming against certain ai like midjourney. It might even be possible that any art created by ai has to be taken down. I am only speculating. But they are going in hard.
|
| Sun, 15 Jan 2023, 09:53 | |
|
Jayenkai
|
 "A glorious sunset on an alien world, as an incredibly detailed starship lands on a neon runway." by #DallE, with Outpainting to extend it. -=-=- ''Load, Next List!''
|
| Mon, 16 Jan 2023, 13:57 | |
|
Jayenkai
|
I've discovered something that all the AI-Art tools are shit at. None of them can do proper ASCII Art, instead falling back to monochrome chunky blocks, instead of nice fancy text based art. Craiyon : > Reveal 🔎 DallE : > Reveal 🔎 Bing : > Reveal 🔎 DiffusionBee : > Reveal 🔎 -=-=- ''Load, Next List!''
|
| Mon, 16 Jan 2023, 14:17 | |
|
Kuron |
Bing & DallE at least made ANSI looking graphics. That is at least in the same neighborhood IMHO. Bing looks best of the bunch, IMHO, as DallE has some odd things that are out of place. I was not aware that Bing could do image generation. Interesting... |
| Mon, 16 Jan 2023, 14:28 | |
|
Jayenkai
|
Yeah, Bing seemed to just kinda pop up out of nowhere. I'm fairly sure it IS DallE, since Microsoft and OpenAI have all manner of license agreements and whatnot. They seem to be rolling it out gradually. If you have access to it, you'll see an "Image Creator" link on the top right of the Bing Image Search page. I can't link directly to the site, as I've been told by others that it seems to lock the feature out for them. Though, whether it was already locked out or not, I can't be sure. -=-=- ''Load, Next List!''
|
| Mon, 16 Jan 2023, 19:29 | |
|
Kuron |
Not there for me, but I do not have a MS account. Also do not use Bing. I use the indie UK search engine Mojeek.
|
| Wed, 25 Jan 2023, 05:00 | |
|
Jayenkai
|
 "The Doctor and the Jelly Babies. Dalek Invasion. Movie Poster Art" by #DiffusionBee -=-=- ''Load, Next List!''
|
| Sat, 11 Feb 2023, 15:20 | |
|
Jayenkai
|
"Hey, Jay," I hear you ask.. "What happens if you take the original BlitzCoder poster, strip out the text, throw it into Stable Diffusion, then stick the text back on it again?" That's a rather specific question, if you don't mind me saying so. The original > Reveal 🔎 The results > Reveal 🔎 A few scraps > Reveal 🔎 -=-=- ''Load, Next List!''
|
| Sat, 11 Feb 2023, 22:16 | |
|
Pakz |
I do not know if I mentioned this here. But I stumbled upon a ipad app that has a stable diffusion generator build in. Dozens of data files with different training. Each around 2gb. The app is called 'draw things' and is 100% free. When I bought the ipad pro 2022 I had in mind the faster cpu could be useful. Each picture can take a while to generate as it is done completely on the device itself. It can be used offline.   I need to sit down and study the settings to get the results that some people using this app have posted.
|
| Sun, 12 Feb 2023, 02:34 | |
|
Jayenkai
|
Awesome. I literally just installed that on the Mac, last night! On the M2 Mac it's getting roughly the same speed as DiffusionBee, but has a heck of a lot more knobs and sliders to play with. Requires an awful lot of cleanup afterwards, though. It has a "tidy up temp files" button, but the button doesn't seem to actually do that, so goodness knows how that works on iPad. Maybe keep an eye on that. -=-=- ''Load, Next List!''
|
| Sun, 12 Feb 2023, 03:28 | |
|
Jayenkai
|
The settings I settled on are... Size : 512x512 Upscaler : Real-ESRGAN 4x Steps : 30 Text Guidance : 15.2 Strength : 40 Sampler : DPM++ 2M Karras Batch size : 1 High Res Fix : Off (Though switching this on seems to give background objects better definition, I've not really noticed too much in the foreground getting any clearer.. I'll experiment further with this option, I think) I installed it last night because I noticed the screenshot mentioned "Textual Inversion" so thought it might let me train the algorithm a little. But, no, it only lets you download other people's. But I was pleasantly surprised by the quality of the output, so figured I'd keep it around. -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 06:44 | |
|
Jayenkai
|
Figured I'd try out a bit of training. Hugging Face's Textual Inversion Training. Runs on the page.. It has little clickable play buttons down most of the page. You need to upload a bunch of images with the thing you'd like it to train from, (512x512 pixels, at least 5 of them, preferably more) You then go down the page, following the instructions, and clicking the little play buttons one by one, until you reach the end.. Once at the end, (after about an hour) open the folders on the left and grab the "learned_embeds.bin" file from the output folder. This can then be imported into the Draw Things app, and you can then use <MyThing> to .. attempt it... Yesterday I got <platstyle> "kinda" working, but I think I need to retrain it with many many more examples to get it up to a decent Platstyle.   Normal vs "Platstyle" I tried getting it to draw a platdude, but it just ended up drawing random squares, so today I used a voxel editor on the iPad to help generate some 3D platdude models..  Hopefully this helps a bit.. Who knows. Let the retraining commence!!! -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 08:02 | |
|
Jayenkai
|
Nope, that didn't work, either.. What it's doing now, is it's taking the request and turning it into 80's/90's polygonal art... Which isn't a bad thing, so I'll probably keep it around, but it sure as hell ain't no Platdude.  "A beautiful landscape. Trees, bushes, river, bridge. A man stands on the bridge, looking down at the stream below. Photograph. " vs  "A beautiful landscape. Trees, bushes, river, bridge. A man stands on the bridge, looking down at the stream below. Photograph. <platdude> " vs  "A beautiful landscape. Trees, bushes, river, bridge. <platdude> stands on the bridge, looking down at the stream below. Photograph." They're ABSOLUTELY being misinterpreted as Style, instead of Object. I think I need to tweak some more. I wonder if that's a Draw Things issue more than a Training one..? -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 10:01 | |
|
Jayenkai
|
Sent the same Platdude images through Dreambooth, and the result is ... um...  "a photograph of an astronaut riding a horse, 4k, volumetric light (Platdude)"  "a photograph of Platdude riding a horse, 4k, volumetric light" Yeah, it's overwhelming the generator, which is a shame. I guess StableDiffusion just can't cope with cubic-pixel-based-characters!! Another day wasted farting about with this. -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 11:17 | |
|
Jayenkai
|
Fun with Stable Diffusion 1.5 vs the super-corrupt Platdude edition! "Mona Lisa" : Stable Platdude "Mill, Chimney and Canal by L.S.Lowry" : Stable Platdide "Mushroom wonderland and a beautiful sunset." : Stable Platdude "A samurai stood on a hill, greyscale" : Stable Platdude (All requests made without the word Platdude, so .. super corrupt!!) -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 13:49 | |
|
Jayenkai
|
Getting better.. Taking about an hour or two at a time, to edit/tweak/test! A small village, platdude on a scooter, matte painting. Completely wasted today! I have an ALChoon to make!!! -=-=- ''Load, Next List!''
|
| Mon, 20 Feb 2023, 21:40 | |
|
Pakz |
These generators are getting better with every season.
|
| Fri, 24 Feb 2023, 02:08 | |
|
Pakz |
The new Samsung Galaxy s23 and its hardware is optimized to generate those stable diffusion images quickly. On the m2 it takes about a minute for the 512x512 to be generated. The Samsung does it in 15 seconds with the image and same settings. Not sure if that recent update has been added to the samsung phones yet. I am not sure if it was just the mobile soc that had the efficient ai processing.
|
|
|
| |||||
|
| ||||||
