The Warthunder OpenAI Thread

[Soryuu]

Yesterday, OpenAI released a new image model which is able to produce extremely coherent/convincing Warthunder screenshot concepts using just a few words. For example:

a screenshot of warthunder in the exact style of the game, where the player is flying a su-57 through an abstract nightmarish soviet labyrinth

In this thread we can share intriguing or amusing fake Warthunder concepts.
Please don’t use the thread to argue about the ethics of AI, you’re free to hide it if you dislike it.

Some other examples:

a screenshot of warthunder in the exact style of the game on cinematic graphics, where the player is flying a su-57 through a portal in Aperture laboratories

a screenshot of warthunder in the exact style of the game on low graphics, where the player is driving a t-34 equipped with a proton pack and shooting a giant Stay Puft marshmallow man

a screenshot of warthunder where the player is flying an f-22 through a titanic rusted underground pipework megastructure, a dogfight between other players is happening in the distance, the environment is large enough to form weather systems

a screenshot of warthunder where the player is piloting a Gundam against Zeon scum

a screenshot of warthunder where the player is flying an f-22 through a foreboding cyberscape made of analog noise, formless monsters made of analog noise observe

not my gen, so I don't know the prompt - taken from an image board which discusses Warthunder

a screenshot of warthunder where the player is piloting a futuristic spaceship above a ring world, stars on both sides as the ring stretches above

a screenshot of warthunder where the player is piloting a ww1 biplane through a distorted escherian Soviet pocket dimension, confusing geometry, instrument glitches

a screenshot of warthunder where the player is piloting an eva unit against another eva unit

a screenshot of warthunder where a group of m22 locusts are battling a gigantic colossus

Please share your own generations and wacky concepts here. A few tips:

• a shorter prompt seems to produce a result with a higher signal-to-noise ratio. A prompt over 1-2 paragraphs starts to look low-detail.
• You can sign up additional free accounts for free image gens
• The gpt-images-2 API allows you to set parameters like quality or image size. I didn’t find them particularly useful over the web interface.

[Beeschurger]

@Soryuu

against Zeon scum

“hey zeon is attacking, bring out the kids squad”

[RDXRooster63]

this is a pretty good one. yes im pointing to the entire game

[iegaatdao]

The AI seems to get speeds right, no supersonic tank/props at all.
And also makes usernames and callouts

[PyroAddict]

First result without further edition:

Second attempt of refinement, thing’s lit:

Prompt:
a screenshot of war thunder where an important group of M1A1 Abrams are traversing across a desert, with a gigantic Italian L3/33 CC facing them ahead

[Константинович7]

Looking into a more technical side of things is impressive, is at certain point believeable—texts are readable; improved pattern consistency and so on, if not some minor mistakes, but overall how it evolved is actually very noticeable:

a impressive screenshot of war thunder **where** the player is driving a M1 Abrams in a massive **underground battlefield** where night vision, flashlights are the key tool to see what's ahead. Other players are in a on-going duel with distant spots of light of the outside world make the aspect ratio 16:9

Is more than obvious that some of the generated images are; some of them pasting of real images and real in-game screenshots with adjustments, some delusions here and there.

a realistic screenshot of the videogame war thunder where the player is **sailing** a ship, instead of the ocean is a spaceship shooting gigantic lasers from their cannons in a **galactic battlefield**. Create the player's and allied vessels spaceship-like with both realistic spaceship aspects but with naval vessels **similarities** make the aspect ratio 16:9

[iegaatdao]

We will make BVVD pay by taking away all of his water.

[PyroAddict]

Underground battles could be a really interesting setting if Gaijin gave themselves the possibility to dive into it, like a glorified Operation Metro setup from the Battlefield series? That thing woud be awesome!

[Константинович7]

That’s the idea. Would be nice to see more fantastic maps leaning torwards a more unique experience. I’m very aware that War Thunder still balancing the fantastic and realistic at the same time but in my opinion a bit of new or never before seen content at least on maps would be nice, even if it’s restricted to arcade battles for sake of realism.

[iegaatdao]

Blob
To be fair I made it generate a focke wulf triebflügel which its at the limit of the definition of “airplane”

generate me War thunder gameplay in air realistic battles with the focke wulf triebflügel

I keep generating and it actually tends to nail the BR spreads pretty good.
So that’s another point

[Soryuu]

@iegaatdao

focke wulf triebflügel

It definitely has some trouble with obscure vehicles - ones which are in the perimeter of the dataset weights, so to speak. You can probably help it along by providing a reference image and telling it to edit it, haven’t experimented much with that yet. (it seems not as robust as eg. ControlNet)
It may even be worthwhile to directly tell the model to search a specific link for info about a vehicle, since this is probably a multi-model image generator which uses an LLM on top of the diffuser for technical review, composition changes, etc.

Have some more random gens

I wish it let you control the temperature of the model, would be awesome

I noticed in complex prompts it sometimes stops styling player & vehicle names accurately, you can remind it by adding:

with player and vehicle names and other UI elements in the exact same style as the game,

[山_田_凉]

quick question, does it have any of the following: local/offline use, comma-delimited list prompting (as opposed to natural language format), ability to change weight of specific word/phrase, inpainting/outpainting, customization of style, images-to-image, image-to-text? Also, can you “see” the AI thinking like ChatGPT does with their LLMs? It is shockingly creative and accurate, fine details are excellent.

I wonder if they combine a geometry-based architecture/model with a diffusion-like one, or something completely new.

[Soryuu]

Definitely no offline usage, comma-split tokens should work just fine (it seems that an LLM takes your prompt and rewrites it for the diffuser anyway), I think parentheses can influence token weight (but not directly like stablediffusion, more like the LLM rewrites it to give that area more weight), you can’t do LORAs if that’s what you mean by style, image-to-image works fine. Not sure about the rest

I hope they open it up more and add features like temperature control!