| Page 3 of 10 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Mon, 28th Feb 2022 14:05 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 14:55 Post subject: |
|
 |
Heh, yes. Running it on your own machine isn't really ideal. You'd need a 3090 to get anything done. I'm not sure how much of a difference the CPU makes for this kind of task.
(what we really want for this is the Nvidia A100 which is the latest/best monster-card for AI/neural-net/deep learning... for a mere ~$30,586)
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 15:27 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
DXWarlock
VIP Member
Posts: 12631
Location: Florida, USA
|
Posted: Mon, 28th Feb 2022 15:29 Post subject: |
|
 |
| Frant wrote: | | (what we really want for this is the Nvidia A100 which is the latest/best monster-card for AI/neural-net/deep learning... for a mere ~$30,586) |
We get google cloud boxes for work our SaaS software runs on to do PDF to jpg/tiff rasterization and upscaling on. So I 'borrowed' of one of the ones not used or assigned to any workflow to use and connected to it with the 'Connect to a custom GCE VM' in colab.
And for some reason A100 will not work, at least for me. No matter WHAT I do I get out of memory error on it, saying "39.9GB of blah blah assigned, 2GB trying to be wrote" or whatever the error is.
Strangely with the same settings a v100 works fine. if I edit the GCE instance and change from a100 to v100 it works...so not sure if a EBKAC error on my side or just it doesnt like it. I cannot see how changing the GPU and nothing else on a VM could cause it to be an error I am doing.
But think I will just stick with a v100, with a res of around 1800x1024 only takes 5-6 minutes on a 300 step image, thats fast enough for me.
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
Last edited by DXWarlock on Mon, 28th Feb 2022 15:46; edited 2 times in total
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 15:42 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
DXWarlock
VIP Member
Posts: 12631
Location: Florida, USA
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 15:51 Post subject: |
|
 |
Trying to create specific scenarios is like trying to control an airplane by blowing on it. You have to write very specific details of the scene as well as use other tricks mentioned in the thread.
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
DXWarlock
VIP Member
Posts: 12631
Location: Florida, USA
|
Posted: Mon, 28th Feb 2022 15:56 Post subject: |
|
 |
Very true I noticed...haha.
But I am enjoying it more for the 'failed but interesting attempt at it' aspect.
Its like asking a very talented artist that is an advanced alien lifeform to paint/draw/generate a photo of what I ask it. When it only visited earth for a week trying to absorb all the knowledge of everything about earth in that time.
I enjoy watching what it tries its best to make with its alien concept of what those things are.
Like if I asked it to make the Mona Lisa and it did even a remotely vaguely passable job, I would be let down and disappointed in the results.
I am fascinated more by how it does it wrong, but tries so hard to get it right its acceptable how wrong the 'wrong' parts are. the visual version of the hitchhiker's guides asking for tea and getting "a cupful of liquid that was almost, but not quite, entirely unlike tea."
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Mon, 28th Feb 2022 16:38 Post subject: |
|
 |
It's very easy to create very specific scenarios, that's what the init_image is for.
A black rectangle in the init_image and a prompt that says "house" - the AI will render the house over the box because it's the shape that fits best. It's also the best way to create people. Create people-shapes and you get proper people (so not people with 3 legs etc). Use simple shapes (just steal from google and use silhouettes). Complex shapes will obviously be as unpredictable as noise since it can be anything to the AI.
The whole idea is that it latches onto random (or not random) noise, from this it paints the image from one point to the other until there's a complete image that corresponds with a specific part of the prompt.
If you played around with Deep Dream (make people, items look like dogs etc etc, very trippy) when that was available, this is a continuation of that...
I do think the most creative way to use this tool is to direct the AI with the init_image, that way it actually is your work. it's your noise, not the default. No one can replicate this image unless they have your init_image. I'm not using it for most images here, i'm still to eager to try things out, i did use it for the very first cyberpunk image (in first post).
Cool images 
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 17:19 Post subject: |
|
 |
I haven't use any kind of init_image so far. I've only "told" DD what I want it to do in text.
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Mon, 28th Feb 2022 23:42 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
|
| Back to top |
|
 |
DXWarlock
VIP Member
Posts: 12631
Location: Florida, USA
|
Posted: Tue, 1st Mar 2022 18:17 Post subject: |
|
 |
| vurt wrote: |
also figured out you should using array elements to separate things in the prompt, so i've been doing it wrong lol.. much of it actually is already as default and explained, and i should learn to not just ignore stuff or just follow others blindly (which i did here) i've been using a "|" to separate stuff, but that syntax is not used for DD. |
I did the same to start.
The best/correct way is:
| Code: |
text_prompts = {
0: ["prompt one:5", "prompt two:10"],
50: ["prompt three:5", "prompt four:10"],
}
|
If you wanted it to do two prompts starting at step 0 one with weight 5, the other 10. Then add in two more starting at step 50 with 5 and 10 weights.
--
Dont NEED the weights, or the later step ones. can just be:
| Code: |
text_prompts = {
0: ["prompt one"],
}
|
IF you wanted simply one thing.
-We don't control what happens to us in life, but we control how we respond to what happens in life.
-Hard times create strong men, strong men create good times, good times create weak men, and weak men create hard times. -G. Michael Hopf
Disclaimer: Post made by me are of my own creation. A delusional mind relayed in text form.
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Tue, 1st Mar 2022 22:25 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Wed, 2nd Mar 2022 00:33 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Wed, 2nd Mar 2022 09:05 Post subject: |
|
 |
Just tried 4 images of "Skyrim Flora Overhaul" as the only thing in the prompt.
Result;
* Grass growing through and ON grey rocks in all images. LOL so good, it's exactly how it is in this game.
* Overbright, ugly Creation Engine graphics. it totally has the right idea but it has also learned the worst images possible.
* Awful representation of my work lol
No images were worth sharing (i'll share 1), very ugly apart from some of the close ups of the flora which i could steal and use in my mod actually, that would be funny. The graphics would match the game's overall graphics really well.
Fun and very good to learn from, how it copies (it doesn't, it makes something completely unique, no copy/paste part of some flora yet it gets super close)... I know this engine and its images (default + my own) way better than i know any painter so a nice test...
Spoiler: | |
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Wed, 2nd Mar 2022 18:20 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Wed, 2nd Mar 2022 20:54 Post subject: |
|
 |
How many steps?
Seems like doing a local install is the way to go for faster renderings, someone said on the official Discord for DD that even the 3070 is 2x faster than the old (but super expensive) cards you get from Google on Colab.. they do have 16GB which is nice though and why they were expensive and still decent.
Ordered a 3090 OC with (24GB) so it should be a bit quicker!
Been messing with animations all day, so much fun. DD5 Turbo (special version for faster animations).
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Wed, 2nd Mar 2022 21:19 Post subject: |
|
 |
500 steps. It seems to be the sweet spot for quality/performance for just playing around with while getting a pretty good result.
How does the syntax work for animations? I wish there was a full guide for every setting as well as a guide on how to write the most effective prompts. Experimenting to see what works and what doesn't takes a lot of time and waste a lot of runtime.
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Wed, 2nd Mar 2022 21:25 Post subject: |
|
 |
It's on 5) Create Video. just click that, but of course you also must have created a ton of batches so that it has a lot of images to process for the video. Then there are functions lite rotate and zoom etc, those i'm still learning i'm not gonna pretend i can explain those lol, at least not yet.
500 steps, wow. I rarely go over 150, the images i've posted here are 100-250 steps.
More steps is not always better either, i once tried 100 and something like 800 and i liked the one which was 100 better (same seed), had more interesting details that were gone in 800 and turned into something else. Depends on the image i guess.
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Wed, 2nd Mar 2022 21:30 Post subject: |
|
 |
I'm still trying to figure out how to put something in the center of the image. "A yellow dot in the center of the image" doesn't really work. I guess I have to do some more research on how to use init_picture to tell the engine where I want to place stuff or overall how to control the dimensions, angles, view directions etc.
EDIT: I MAY use an unnecessarily high steps-setting. It does increase the rendering time and it waste valuable runtime usage. I'll back down to 150. That way I can continue to experiment with the prompt without wasting an unnecessary amount of runtime.
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Wed, 2nd Mar 2022 22:09 Post subject: |
|
 |
yeah, the fun in this is experimenting imo, not getting poster-quality work. that's for when you've mastered it (imo at least).
small video (takes hours to render!)
https://i.imgur.com/W3ywCsg.mp4
Not really the most fun scene. i just entered "cyberpunk city" and then i went out for a few hours. I could have spent some minutes looking for a better seed.
what's so incredibly cool is that it's totally 3D. so whatever creation you've done you can move around in 3D for that scene. but yes, doing camera work for it, that's a total pain...
Things will get better though because there's a lot of smart people involved now, very eager to take this to the next level.
|
|
| Back to top |
|
 |
Frant
King's Bounty
Posts: 24978
Location: Your Mom
|
Posted: Wed, 2nd Mar 2022 22:54 Post subject: |
|
 |
Wow, that's cool. Can't wait for smarter/better versions of Disco Diffusion or even a competitor.
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn!
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
|
| Back to top |
|
 |
|
|
|
| Back to top |
|
 |
Il_Padrino
Posts: 8285
Location: Greece by the North Sea
|
Posted: Sun, 6th Mar 2022 14:11 Post subject: |
|
 |
Jesus, those LOTR renders in particular look awesome.
The numbers don't decide, the system is a lie. The river running dry, the wings of butterflies.
And you may pour us away like soup. Like we're pretty broken flowers.
We'll take back what is ours. One day at a time.
|
|
| Back to top |
|
 |
vurt
Posts: 14937
Location: Sweden
|
Posted: Tue, 8th Mar 2022 01:46 Post subject: |
|
 |
|
|
|
| Back to top |
|
 |
| Page 3 of 10 |
All times are GMT + 1 Hour |