• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Dream up concept art for a new video game in DALL-E 3

Rockondevil

Member
3rd game from Valve.

fSXvouW.jpg
 

Dacvak

No one shall be brought before our LORD David Bowie without the true and secret knowledge of the Photoshop. For in that time, so shall He appear.
w1vrn79.jpg

God damn, I would genuinely love to have this as a poster.

Is there an artist union? Cause they better go on strike soon. This is the last chapter before AI takes over a majority of the creative “grunt work” for artists and writers. Who knows, probably coders and composers too.

The future is absolutely insane. I don’t know how I feel about it, but there’s no un-fucking this dog. We won’t go back.
 
Last edited:

LiquidMetal14

hide your water-based mammals
I'll come back to this when I get home but I'm all down for this.

Some pretty awesome stuff already popping up.
 

Wildebeest

Member
How the fuck did you prompt it to do this?
Start your prompt with something like "pencil drawing". It's one of the things they recommend. Other people might want a drawing in the style of a particular artist and use that name, but you don't have to for DALL-E. A standard style prompt for something like this would be something like "digital art" or "video game concept art".
 

Croatoan

They/Them A-10 Warthog
It’s probably because it thinks you’re trying to trick the algorithms into making porn
I changed it to pretty gothic witch and it said the same thing. You can use bing to search porn easy as pie but buy god you better not ask for a pretty witch ai art…

Edit: I think it was actually different part of the prompt where I used “thigh high water” to describe the location…
 
Last edited:

Croatoan

They/Them A-10 Warthog
Any help getting the style prompts right. I was trying to get a Boris Vallejo style but it seems to not understand as it comes out realistic looking (like a bad photoshop). Should it be like epic fantasy style or painted fantasy or something?
 
This looks crazy good for something an AI generated. What were the prompts if you can share?
'Final Fantasy 17, drawn in the style of Final Fantasy 12's concept art' and 'Final Fantasy seventeen, Akihiko Yoshida concept art'

Honestly though, it took a ton of refreshes to get it to show those 3. This version of Dall E, while impressive, still has pretty massive faults. It has gotten way better in speed and consistency in art, but it is still lacking a ton when it comes to reading prompts correctly. Take my example 'Final Fantasy 17'. It kept wanting to give me variations of Final Fantasy 7 art, Final Fantasy 1 and 7 mashed together, and just variations of Cloud(since he's the most popular FF protagonist).

It can't do specific character names properly, locations are hit and miss, art style is extremely hit and miss, and anything that's less than 'popular I.P., popular character, or popular search query' is almost always a miss. It feels like skimming on the surface of the ocean and every time you try to dive you get pushed back to the surface.
 

EviLore

Expansive Ellipses
Staff Member
'Final Fantasy 17, drawn in the style of Final Fantasy 12's concept art' and 'Final Fantasy seventeen, Akihiko Yoshida concept art'

Honestly though, it took a ton of refreshes to get it to show those 3. This version of Dall E, while impressive, still has pretty massive faults. It has gotten way better in speed and consistency in art, but it is still lacking a ton when it comes to reading prompts correctly. Take my example 'Final Fantasy 17'. It kept wanting to give me variations of Final Fantasy 7 art, Final Fantasy 1 and 7 mashed together, and just variations of Cloud(since he's the most popular FF protagonist).

It can't do specific character names properly, locations are hit and miss, art style is extremely hit and miss, and anything that's less than 'popular I.P., popular character, or popular search query' is almost always a miss. It feels like skimming on the surface of the ocean and every time you try to dive you get pushed back to the surface.
You’re asking it for something by name that doesn’t exist. It’s trying its best to interpret what you mean based on its existing knowledge.

Instead, try something like Final Fantasy art in the style of Yoshitaka Amano, with a man in gothic plate armor standing resolutely opposite a princess in exotic religious robes holding a necronomicon with trepidation, at sunset, on a balcony overlooking a scifi fantasy steampunk world with chocobo t-rexes wandering in the fields.

It’s a tool to express your own creativity.

Results:

QnOZKba.jpg


UmtJN1F.jpg


ktDa7pF.jpg


That chocobo T-Rex in the third variant is a godlike creature dwarfing a city, phew!
 
Last edited:
You’re asking it for something by name that doesn’t exist. It’s trying its best to interpret what you mean based on its existing knowledge.

Instead, try something like Final Fantasy art in the style of Yoshitaka Amano, with a man in gothic plate armor standing resolutely opposite a princess in exotic religious robes holding a necronomicon with trepidation, at sunset, on a balcony overlooking a scifi fantasy steampunk world with chocobo t-rexes wandering in the fields.

It’s a tool to express your own creativity.
I apologize, I should have worded my previous post better. The issue I have isn't specifically with the concept of Final Fantasy 17(as you've said it doesn't exist), but with my previous example in the thread with Metal Gear Solid. Trying to make any character other than Solid Snake or Raiden appear is frustrating because the application feels like it only has a surface level understanding of the concept of Metal Gear Solid. So when I looked for lesser known characters, they would never show up, and instead it would put the main characters in situations that matched the next word as closely as possible, even though that word was literally a character in the series itself.

Here is a test I do for each AI image creator I come across, but it seems to work also as a difficulty filter for them: I simply type 'Abby, The Last of Us' 'The Last of Us Abby' 'Abby Anderson, The Last of Us' or 'Abby, The Last of Us 2'. All Images were of Ellie and some were of Joel and Ellie. Not a single one actually produced Abby. They didn't know who Abby was, because Abby isn't one of the most popular search terms amongst TLoU fans. I don't even care about Abby that much, but I want to see if the A.I. generator can finally figure out who this character is, because she is literally the second protagonist of a popular video game.

Of course, someone could come into the thread and finally produce an Abby image to prove me wrong, but that would be missing the point. The point was it could never happen for me. To me, that's a pretty big flaw, and one that I hope we see ironed out in version 4 or 5. I see the massive potential and future behind Dall-e much like many here, but with each version I still come away from it feeling that it can be surface level at times in what it produces and I just have this inner urge to go deeper but I simply can't, thus my ocean analogy.
 

EviLore

Expansive Ellipses
Staff Member
I apologize, I should have worded my previous post better. The issue I have isn't specifically with the concept of Final Fantasy 17(as you've said it doesn't exist), but with my previous example in the thread with Metal Gear Solid. Trying to make any character other than Solid Snake or Raiden appear is frustrating because the application feels like it only has a surface level understanding of the concept of Metal Gear Solid. So when I looked for lesser known characters, they would never show up, and instead it would put the main characters in situations that matched the next word as closely as possible, even though that word was literally a character in the series itself.

Here is a test I do for each AI image creator I come across, but it seems to work also as a difficulty filter for them: I simply type 'Abby, The Last of Us' 'The Last of Us Abby' 'Abby Anderson, The Last of Us' or 'Abby, The Last of Us 2'. All Images were of Ellie and some were of Joel and Ellie. Not a single one actually produced Abby. They didn't know who Abby was, because Abby isn't one of the most popular search terms amongst TLoU fans. I don't even care about Abby that much, but I want to see if the A.I. generator can finally figure out who this character is, because she is literally the second protagonist of a popular video game.

Of course, someone could come into the thread and finally produce an Abby image to prove me wrong, but that would be missing the point. The point was it could never happen for me. To me, that's a pretty big flaw, and one that I hope we see ironed out in version 4 or 5. I see the massive potential and future behind Dall-e much like many here, but with each version I still come away from it feeling that it can be surface level at times in what it produces and I just have this inner urge to go deeper but I simply can't, thus my ocean analogy.
Ah, gotcha. Yeah, I tried to get a Joel vs Abby showdown generated this morning and the result was stuff like this:

xHeDqD4.jpg


along with Joel vs Ellie output. The model knows who they are but not Abby, and Ellie consistently looks like her age from TLOU1. So we can infer that TLOU2 wasn't included in the training data set.
 

Myths

Member
maTB341.jpg


The Sun Keep, after the Fall of Zeal, powering a shard of Lavos now turned to Crystal. Lost in history is a time where the Earthbound could even the playing field using these empowered shards.

eikxkKY.jpg

The Black Omen rebuilt in the far future, elevated above the fallout of Lavos’ calamity.

uayF73e.jpg


Somehow, time has reraised what once was a Sunken Desert, but a gathering of obelisks from the Mystics bars passage.

L3QjHRN.jpg

Preparing to cast what appears to be a solo version of the triple tech “Arc Impulse”, as the robotic aspect of Chaos overlooks the world.

fxHfn9Z.jpg

Most of the locations accessible from a base built by an Oracle.


GYa2IaY.jpg

Aboard the Epoch Omega, a refitted version of the original Epoch using components of the Blackbird.

dFh2hni.jpg

Lightning 2, Fire 2, the Dual Tech “Nova” is readied. Slightly older as the plot progresses alongside an unusual Enlightened mage disguised not so cleverly as an Earthbound.


As someone else pointed out, keeping the style consistent would be a challenge. I ended up spending a lot longer than I thought on rewording and generating new outputs. Tough to pick but saved quite a few things for my own personal reference.
 
Last edited:

poppabk

Cheeks Spread for Digital Only Future
Here is a test I do for each AI image creator I come across, but it seems to work also as a difficulty filter for them: I simply type 'Abby, The Last of Us' 'The Last of Us Abby' 'Abby Anderson, The Last of Us' or 'Abby, The Last of Us 2'. All Images were of Ellie and some were of Joel and Ellie. Not a single one actually produced Abby. They didn't know who Abby was, because Abby isn't one of the most popular search terms amongst TLoU fans. I don't even care about Abby that much, but I want to see if the A.I. generator can finally figure out who this character is, because she is literally the second protagonist of a popular video game.
Most human artists if you asked them to draw Abby from the last of us would either look at you non-plussed or draw Bella Ramsey.
It's not a flaw of the A.I. that it's not in the dataset, likewise the fact that Bella Ramsey and Pedro Pascal never make an appearance in the generated images suggests that the TV show is also not in the dataset. Expecting an AI to 'figure out' a character it has never been exposed to is kinda insane.
 
Most human artists if you asked them to draw Abby from the last of us would either look at you non-plussed or draw Bella Ramsey.
It's not a flaw of the A.I. that it's not in the dataset, likewise the fact that Bella Ramsey and Pedro Pascal never make an appearance in the generated images suggests that the TV show is also not in the dataset. Expecting an AI to 'figure out' a character it has never been exposed to is kinda insane.
You're defending this application so much that you're only seeing it from your own perspective. Think of this from a casual standpoint. You're dealing with knowledge and information that the casual audience who is saying 'this is neat' and trying this tool out, simply isn't aware of. The casual audience is going to look at it very simply: Does it work all the time, does it work half the time, or does it simply not work? That's it. To expect everyone to know what you know to the level that you do, would be the true version of insanity.

From saying that, I still consider the dataset problem a flaw that hopefully will one day be ironed out. I have a feeling that MS is heading in that direction so it's a good thing.
 
Top Bottom