Using AI resources

[Skrum]

I'm not super experienced with Midjourney or ChatGPT, but just fiddling around with them recently I managed to create some really cool and useful stuff.

1) Midjourney can do battle maps! I was using it to generate some generic forest and generic rocky terrain, and then when those came out really good I also did some Cursed Churches. Haven't really pushed it beyond that, but I was really impressed with the quality and I got nearly 20 usable maps. It would've taken me hours to make all these manually, but Midjourney did it in with about 20 minutes of effort.







2) ChatGPT will make languages for you
Well, at least enough to name things. One of my characters, a rune knight, recently got a new sword that deals fire damage. I started by going to ChatGPT just for name inspiration, but after a few prompts I asked it "make a language that giants would speak." It gave some roughly 20 words and their meaning, some examples of sentences, that sorta thing. Then I asked it "name a fiery sword using the giant language. And it spit out this -



So now my rune knight wields Thrumgrahlgorm, which I went with because I liked the way it sounded in giant the most. Pretty cool stuff!

[Hael]

You could likely animate several of those as well (eg running water). You could also get AI to pronounce a giants voice or to provide heroic (or theme appropriate background music) (udio/suno). Maybe draw a picture of the sword, or eg some runes on the ground.

Chatgpt could give you some random background information about one of the swords.
Ai, properly used in 2024, is mostly just a productivity tool and a time saver.

I use copilot everyday for writing boilerplate code (or to translate an equation into latex) and it saves me 10 seconds here, and 10 seconds there, and it really adds up over the course of a day.

[Psyren]

I won't touch the use of the art resources here, but I've definitely used ChatGPT for cool-sounding names in different fantasy languages. Vundromrath sounds pretty awesome.

[firelistener]

Cool tip on the language thing. That's a neat idea, so thanks for sharing. I've written several programs over the years to generate "pronounceable" passwords and names that work on the same core principle but are much simpler. I used to generate a lot of my NPC names with one.

The basic idea is that you take a sample text, like a long book in English or a list of names from a certain culture, and feed them into the program. First, you build a table of the likelihood of a word beginning with a letter. So if the sample was "Mark Matthew Luke John", there is a 50% chance to use the letter "M" to start a name. Then you build a table for each starting letter with the likelihood that a given letter follows it. So with that same sample, there would be a 100% chance I generate the letter "A" to follow "M" since my sample only has "A"s after any "M"s. Using a sufficiently large sample, the tables give you results that look like real names. The AI language models we have now are like that but taken to the extreme.

[Rafaelfras]

I've been using bing AI to draw our characters and after a session the most impactful scene of that weekend. It's been a blast because I've manage to get what I want for every character of the party (scenes not so much but such are the limitations and we laughed on the weird ones). The bing one is really good on doing what I want it to do. It was the only one that made a accurate robe of stars for example. It is a fun little thing that made our sessions more colorful so to speak

[MarkVIIIMarc]

Those are some neat things. Especially on the art work.

In some regards though we have had random generators online for at least a decade. Is the big difference now we have virtually unlimited possibilities?

[LudicSavant]

Pretty much the only thing I've utilized ChatGPT for thus far is to aid in naming large amounts of things in a way that at least feels somewhat linguistically consistent -- I'll ask it something like "generate 100 names inspired by (insert culture, real or fictional) along with etymologies." Even then I will often tweak the names it gives me.

[Jophiel]

I've been using bing AI to draw our characters and after a session the most impactful scene of that weekend.

AI Art has taken over the Character Portrait industry at our table (that used to be "steal a close enough pic off the web" or "Screenshot a Heroforge model") and we have a guy who does the Highlight Moments Captured In Image thing for the week as well.

No maps though. I've seen it done but our DM isn't interested in it. Not opposed to AI but just not interested in learning about it.

[follacchioso]

I play a lot of play-by-post D&D on Discord, both as a player and as a DM. I am also a Stable Diffusion and Midjourney geek, and I try to combine the two things. In my opinion, these are great tools and can really improve the game experience.

- Midjourney and Stable Diffusion (and the other text-to-image generators) are great for generating character portraits. I am very slow at creating new characters, and it takes me days and days to finalize a character sheet. I go over many optimization options, background types, and personality traits. But the biggest issue is finding a good portrait for my character - on a few occasions, I ended up throwing away the whole character because I could not find good art for them. Now, with AI, this problem is solved. It still takes me many attempts in midjourney, but now I can usually generate good art by myself. I even play a changeling bard character that uses a different physical aspect every mission, and it is fun to customize him that way.

- These tools are also great for DMing. I try to create a main image to describe the mission, when I advertise it to potential players. I also generate art for the main NPCs and monsters. It really adds a new direction to DMing play-by-post.

- I am less convinced about using Midjourney and StableDiffusion for creating maps. I think the maps "look" nice, but in reality, they are not optimized for a real game. For example, a while ago I generated a map of a road going through the forest. The result was apparently great, but then, when I played the mission, I realized the limitations. There were too many trees, and not enough space for the players to move. A smart map designer would have created it differently, leaving lots of space for the road, and adding just a few representative trees. It is also difficult to generate this type of maps by prompting, as the tool gets easily confused when you add too many details.

- I am also less convinced about using chatGPT for describing scenes. As a non-native speaker, I still use it sometimes to improve my descriptions, and it really helps. I also used it to generate songs and poems, which would be really too difficult for me to create. However, after using this for a while, I came to the conclusion that most of the text generated this way is "flat". It feels like forcing the players to read boring text that I did not even bother trying to write. I prefer to write shorter text, in a less elegant way, than the long prose generated by AI.

[Segev]

I wonder if using image to image with a functional map to start with and text prompts to try to decorate them, dress them up, etc., would work well.

[follacchioso]

Yes, image-to-image helps. You can also use a basic map as style reference. I think that's what I've done in the past.
It will generate a randomized version of the map, with some elements in common. It may introduce elements unrelated to what you want, or deform the key structures you want to see, but most of the generated maps will be usable.

However, the AI will not know that combat is supposed to happen on that map. It will not think like a human. "Let's create a narrow passage here, where the characters can be ambushed or create an ambush themselves"; "Let's add a broken cart here, where the previous expedition was attacked"; "Let's add a lantern there, so the players can use it as a source of fire during the combat".

Moreover, you will have very limited control on where to put the elements of the map. You can try to describe it in the prompt, but the generator may just ignore your instructions. There is also a limit to the number of details that you can include in the prompt. If you are good at StableDiffusion, you can have more control over that, by using CLIP adapters and multiple prompts; but it is a lot of work for that.

In short, I think Midjourney can create a mediocre map quite easily. That can be useful in many occasions, especially if you are in a rush. However, there are better random map generators out there, like Inkarnate or Wondercraft. And a good map designer will still be impossible to beat, if you want good quality results.

[Segev]

It might be interesting to try to develop a model or Lora that is trained to try to recognize tactical interesting-ness.

Developing the labeling and training set would be tricky.

[LudicSavant]

It might be interesting to try to develop a model or Lora that is trained to try to recognize tactical interestingly s.

Developing the labeling and training set would be tricky.

I've oft lamented the use of prettier but less tactically interesting maps by DMs that feel they need to pull one off of the internet or generate one via AI just so that it will look good in a picture frame.

Instead of trying to teach a an AI to make a good tactical map (I think it's gonna be a long time before an AI is even half as capable as a good level designer), it might be better to see map generation that has people lay out their own and then can 'prettify' it.

[faegames]

My latest fun as a player: Face swapping my party members and I on existing art. In the vein of the above "we had to rely on close-enough pics on Google", we now go full customisation. Generate the art and then put our faces on it so we're really in the game ;)
Awaiting the new animation capabilities to make my artificer do a victory dance after combat haha.

[Beelzebub1111]

I use Bing Create if an NPC is missing a portrait and I can't find any official art for that character.

[follacchioso]

It might be interesting to try to develop a model or Lora that is trained to try to recognize tactical interestingly s.

Developing the labeling and training set would be tricky.

It could be a cool project to work on. The main problem would be finding a good dataset to train it. Maybe companies like Roll20 and other virtual tabletops have data on how maps are used, with logs of where DMs tend to position monsters, and where players tend to move. One could potentially try to pseudo-generate a training dataset using chatGPT, but who knows if the results would be good enough?
Another option is to take maps from published adventures and annotate them as a training dataset.

[NichG]

It might be interesting to try to develop a model or Lora that is trained to try to recognize tactical interestingly s.

Developing the labeling and training set would be tricky.

A segmentation control net might be better here. Sketch out blocks of color meaning open terrain, difficult terrain, impassable, object of interest, etc.

[Witty Username]

I don't much like the idea for maps, AI is decent to get something broad strokes fine. Proxy Mtg art that you need to look different enough to communicate its a proxy but still OK.

The more specific the dumber it gets.

I will stick to my grid map, wood cubes and Jenga pieces. It looks meh, but it works for making a space game able.

[tchntm43]

I use ChatGPT when I need an NPC with a few extra details filled out, it's also pretty great at creating puzzles and riddles if you give it guidelines to produce something specific to what you want. Sometimes it's useful to fill in a missing link, like if I have this scenario here, and this scenario there, create a plausible way to get from one to the other.

I also have used it to design homebrew monsters. You can ask it to create a monster stat block for lots of different ideas. For example, I asked it to create a huge-sized Phase Spider that was appropriate for CR 6. In addition to upping the hit points, save DCs, and damage, it also made the Ethereal Jaunt stun nearby creatures (with a CON save of course) when it shifted to the material plane, which was pretty cool. While that was based on an existing monster, I also asked it to create a "shadow hydra, a roiling mass of smoky shadow, tentacles, and electricity, with multiple serpentine heads like a hydra" and it did pretty well there, too.

I use NightCafe to generate character portraits for NPCs the party meets, and put them into a print-out with text about the NPCs the party knows, so that the players can refer to it any time they are trying to remember which NPCs they know. I prefer Nightcafe over Midjourney because it doesn't cost anything and I don't have to deal with Discord to use it.

[Rafaelfras]

AI Art has taken over the Character Portrait industry at our table (that used to be "steal a close enough pic off the web" or "Screenshot a Heroforge model") and we have a guy who does the Highlight Moments Captured In Image thing for the week as well.

No maps though. I've seen it done but our DM isn't interested in it. Not opposed to AI but just not interested in learning about it.

I did SO MUCH close-enough pics on Google. Though the AI is good to generate our characters, is not there yet to generate the battle scenes the way I would like. But it is getting there. For this year I have adopted art nouveau style and the results are gorgeous