|
Yeah I wanted to avoid using the official art that was provided for the gnome ceremorphs as I like custom art for things when using the virtual tabletop; I did end up using art I found online that at least sorta conveyed what I wanted to get at. I'm basically running Rime of the Frostmaiden but adjusting things here and there.
|
#
?
Nov 19, 2023 00:54
|
|
|
|
| # ? May 2, 2024 03:36 |
|
|
Switched for a more generic style for the AI art in my homebrew roleplaying game. It kind of starts to look like a real game. Sorry for weird non-english language! 
|
|
#
?
Nov 26, 2023 16:53
|
|
|
The new Stable Video has potential. You can't really guide it, you just give it an image and it animates it how it thinks it should. Photos and action shots get pretty well animated, abstracter stuff gets completely mangled. Still, neat toy. Could potentially make animated tokens with it. https://stability.ai/news/stable-video-diffusion-open-ai-video-model There's also a new version of SDXL (SDXL) that's extremely fast. It doesn't look nearly as good as regular SDXL but it's soooo much faster, like by about 40x. So you can use it to rough out your prompt, then take it over to SDXL for final touches. https://stability.ai/news/stability-ai-sdxl-turbo If you don't have the gas to run SDXL normally this is a good solution. I've had some very good results with SDXL but unfortunately the massive jump in hardware just isn't attainable by most people. SDXL Turbo is, comparatively, pretty low-stakes. Sadly SDXL is still stuck with ComfyUI, the node-based workflow program that is waaaay too complex for many users, but given time I'm sure this will make it into the usual clients (Automatic1111 and SDNext). Megaman's Jockstrap fucked around with this message at 23:59 on Nov 28, 2023 |
|
#
?
Nov 28, 2023 23:55
|
|
|
A clarification here: A1111 can do SDXL, always could, it was just much slower than Comfy upon release. It's been months since I tried SDXL on A1111, but from what I understand it has improved greatly, especially on the dev branch. ComfyUI is a mess, but you can very quickly get workflows set up just by downloading the ones other people have created. That said, I've ended up sticking with SD 1.5. SDXL is gorgeous, and if my image-generating needs were less complex I'd probably move to it entirely, but it still loving hates specificity - it seems to have much stronger "opinions" than SD1.5, and the tools aren't in place to force it to do what you want like they are in SD1.5. I'm looking forward to running some Blades in the Dark in the near future, because at least I know everything I'll want to generate (generic Victorian-looking people, ghosts, spooky pseudo-London architecture) are things that SDXL can handle without too much fuss. Edit: Okay, took a look at SDXL Turbo and that is fascinating. Holy poo poo it's so fast, generates results in real-time as you prompt. For all my misgivings about SDXL, I definitely need to try this out. Kestral fucked around with this message at 03:44 on Nov 29, 2023 |
|
#
?
Nov 29, 2023 03:40
|
|
|
SDXL on auto1111 has improved a lot, yeah. They're both such different tools, it's annoying but realistically I have to have and use both. I generally end up using comfyui when I just have a specific image in mind and know which steps I'm gonna take on it, so it can just get pumped through my workflow. Then I use auto1111 for the big pile of other things I wanna do, like, how am I gonna make a Deforum video in comfyui? Maybe there is a Deforum workflow! I haven't looked into it. But I can only imagine it being like pulling teeth for any even moderately complicated project. And then there's sd.next which has a bunch of other weird things about it, but then things like out of the box LCM support (which is also substantially faster than just normal) and all sorts of compatibility for weird things auto1111 doesn't have, while also sometimes feeling not as broken? And the dev is a nice guy from what I can tell, everytime I've talked to him I've gotten a good vibe. I know that sounds stupid but auto1111.... Isn't that. I kinda hate comfyui but at least it works. I hate auto1111 for being this duct tape and bubble gum monstrosity that very often collapses under its own weight in very stupid ways. We all used to half-complain about how fast it was all moving months ago, and how hard it is to keep on top of it all. Now though, I literally don't think it's possible to keep on top of unless it's literally your job. Things have spread out, hundreds of different things rapidly advancing all separately from each other.
|
|
#
?
Nov 29, 2023 10:01
|
|
|
BrainDance posted:I generally end up using comfyui when I just have a specific image in mind and know which steps I'm gonna take on it, so it can just get pumped through my workflow. Since you use comfy to get specific results, if you have any protips for that I'd love to hear them! My experiences with SDXL on Comfy have been unbearably frustrating when I'm trying to take a specific vision in my head and turn it into something my players can look at; the only times I boot up SDXL these days is when I'm just looking for a mood piece and have nothing else in mind beyond a vibe.
|
|
#
?
Nov 29, 2023 10:19
|
|
|
Sorry, let me clarify: SDXL Turbo is only available on ComfyUI. SDXL is on A1111 and SDNext. Also, SDXL's Controlnets suck. Here's a "character concept" portrait I made for my half-orc fighter PC using SD1.5 and Openpose Controlnet:  Not perfect but still fun and fit what I had in my head pretty well. It's literally IMPOSSIBLE to do this with SDXL's Openpose. I've tried 3 different versions of the Openpose controlnet and they all gently caress it up really bad. They cannot "understand" that 4 different poses on the same page with the word "character concept" or "character art" means that I want a consistent character drawn 4 ways. This was something that SD 1.5 nailed. Very disappointing. Megaman's Jockstrap fucked around with this message at 20:38 on Nov 30, 2023 |
|
#
?
Nov 30, 2023 20:33
|
|
|
Am I on drugs or is midjourney just a discord bot now? I wanted to try some battlemaps, and bing ai image generator cant handle them (or I can't find a good prompt for it)
|
|
#
?
Jan 10, 2024 01:33
|
|
|
Hasn't it always been a discord bot?
|
|
#
?
Jan 10, 2024 02:12
|
|
|
Always been that way. When Midjourney first came online it woke up and chose violence & never changed its mind, that's all.
|
|
#
?
Jan 10, 2024 02:30
|
|
|
in fact the big MJ news lately (besides v6) is the fact that they're working a standalone browser-based tool outside of Discord which is great news, because its interface is really held back by it being a Discord bot
|
|
#
?
Jan 10, 2024 02:41
|
|
|
The site interface is poo poo, though.
|
|
#
?
Jan 10, 2024 04:11
|
|
|
Well poo poo, I guess I can invite it to an empty discord so nobody can see my shame? Or wait for the web interface?
|
|
#
?
Jan 11, 2024 00:40
|
|
|
Elendil004 posted:Well poo poo, I guess I can invite it to an empty discord so nobody can see my shame? Or wait for the web interface? Everyone makes private servers for it, yes. Who knows when the web interface will finally come. I'd suggest making one channel per month and a channel per individual project. If you just do it all in one channel, it'll get really congested really soon.
|
|
#
?
Jan 11, 2024 00:53
|
|
|
Nobody has written a web-discord bridge for this thing yet?
|
|
#
?
Jan 11, 2024 00:56
|
|
|
I've just stopped using AI, personally.
|
|
#
?
Jan 11, 2024 13:43
|
|
|
Running a short game of Warlock! which is an off-brand rules light system for grim and perilous WFRP style games and the first adventure they're going through is, cough, Death Frost Doom (I know, I know, but I already own it and it's what I had the time to prep) so the overall aesthetic for the characters is pretty grimtastic:      
Megazver fucked around with this message at 14:02 on Jan 11, 2024 |
|
#
?
Jan 11, 2024 13:58
|
|
|
Fuzz posted:I've just stopped using AI, personally. I haven't. Still use diffusion gens but the trick is knowing what it's useful for: Handouts for players. I was able to do an entire journal of sketched locations, a diagram of a ruined castle (I had to make the battle map, but it was still cool for previs), monsters, statues, etc. relatively quickly. Magic sword with a specific look? No problem! It's still a little AI janky but it's leagues better than finding some kinda-match through google image search. Player character and NPC portraits. Things it can't do: Battlemaps or indeed any kind of map Realistic architecture Commercial grade anything
|
|
#
?
Jan 31, 2024 20:46
|
|
|
Midjourney can do excellent battlemaps. I mainly use these systems for character portraits and to create bios for NPCs in a hurry
|
|
#
?
Feb 1, 2024 15:51
|
|
|
Humbug Scoolbus posted:Midjourney can do excellent battlemaps. I mainly use these systems for character portraits and to create bios for NPCs in a hurry This is exactly the opposite of my experience, mind showing me a few? With prompts if you're feeling nice.
|
|
#
?
Feb 1, 2024 17:41
|
|
|
Megaman's Jockstrap posted:This is exactly the opposite of my experience, mind showing me a few? With prompts if you're feeling nice. I think in the OP I link to some posts that described some okayish battlemaps. There's some AI images I see in the wild of Actually Pretty Good anime/manga style art that isn't like overbaked overly detailed busy nonsense that looks like it was hand drawn. That interests me because if it worked well enough it'd make for good D&D tokens and references for art commissions with human artists.
|
|
#
?
Feb 1, 2024 18:19
|
|
|
Bing's main problem is it tries to fill the space so you get busy images. Any idea if that can be controlled via prompts?
|
|
#
?
Feb 1, 2024 18:32
|
|
|
Raenir Salazar posted:I think in the OP I link to some posts that described some okayish battlemaps. Ah, yes! You're right. Well I actually think those look ok, but tactically speaking they're pretty rough. Probably perfectly ok for an unexpected throwaway encounter tho. Elendil004 posted:Bing's main problem is it tries to fill the space so you get busy images. Any idea if that can be controlled via prompts? White background, blank background, empty background? Don't use the words intricate or high-detail anywhere. If you're just prompt cargo-culting (i.e. you found what someone else posted and started using it without understanding what you're actually trading for) that could be a problem. When you prompt the whole image is affected by each tokenized item in the prompt so if you're trying something like "wearing an intricate breastplate" you'll get intricate stuff all over the image, it's called "concept bleed" and it's one of the things you'll actively have to manage, particularly if you have a long prompt with a lot of very specific tokens. They'll fight against, and interact with, each other and you'll end up with very weird output.
|
|
#
?
Feb 1, 2024 19:02
|
|
|
Megaman's Jockstrap posted:Ah, yes! You're right. Well I actually think those look ok, but tactically speaking they're pretty rough. Probably perfectly ok for an unexpected throwaway encounter tho. That's good, I had some trouble (history only goes back 20 images so its buried) getting it to give me a single boat at a fishing pier it was like a city of piers with boats everywhere.
|
|
#
?
Feb 1, 2024 19:03
|
|
|
Elendil004 posted:That's good, I had some trouble (history only goes back 20 images so its buried) getting it to give me a single boat at a fishing pier it was like a city of piers with boats everywhere. Maybe try "single boat, alone, one boat" (really fuckin hammer in that it's one boat) and then in the negative prompt try "fleet, flotilla"? Hmmm now I want to try this, can you give me your prompt? edit:  SDXL base pos: high quality watercolor of a lonely wooden fishing pier with a solitary skiff moored to a lashing, sunset, calm seas neg: flotilla, fleet, anime, cartoon, modern, structures, buildings not perfect and looks more like a lake but I could probably just let this gen 100 images and pick one Megaman's Jockstrap fucked around with this message at 19:14 on Feb 1, 2024 |
|
#
?
Feb 1, 2024 19:05
|
|
|
Megaman's Jockstrap posted:This is exactly the opposite of my experience, mind showing me a few? With prompts if you're feeling nice. Sure!  top down, battlemap,1930s alley, night, rain, battlemap, top down --no text  outdoor field top down battlemap --no grid  VTT battlemap,modern, busy city street and sidewalk,top down  top down battlemap, modern chinatown alley, night, battlemap, top down, detailed, intricate --no grid  Viking village, top down battlemap, detailed, intricate
|
|
#
?
Feb 1, 2024 23:26
|
|
|
Humbug Scoolbus posted:Sure! I was curious so i fired that into bing and got two crap results and 2 maybe ok results.   But it got a little better.  This is probably the only usable one but this is leaps better than the last few times I tried battlemap stuff (on bing) 
|
|
#
?
Feb 2, 2024 00:04
|
|
|
Ah, I guess I shouldn't say it's bad for battlemaps. It's definitely bad for the way I do it - I use Foundry, so baked-in light sources, resolution scaling, and gribbly wall bits are a real pain to deal with. But if you could print these out or drop 'em raw into a simpler VTT, they're fine.
|
|
#
?
Feb 2, 2024 02:45
|
|
|
My experience with battlemaps and SDXL, you can get good ones but I've only had luck with certain samples like dpm++, and even then it's very picky. You can have a prompt that makes complete sense and seems to really be telling it to do something a certain way and it'll spit out very obviously not a battlemap, you change to something else that shouldn't, realistically, be any better and you get a great battlemap. That and some concepts work with a battlemap, others influence it too much into non-battlemap territory. They usually need to be touched up after the fact because they include some impossible things. I think if I spent more time trying to perfect battlemaps with a better workflow, some sorta img2img situation or controlnet deal could probably consistently get good stuff. Actually yeah, I've never tossed a coherent non-AI battlemap into SD and used that as a base to create something new but that might work really well.  This is just a very normal one, you can see where it has problems but I think they could be pretty easily fixed.  This is what happens when you use "intricate" in the prompt, a lot of very intense battlemaps but nothing too usable.  Basically it doesn't understand where paths need to be, and doors, which is why I'm thinking img2img could help. There's some nonsense to it but I really didn't spend much time on these so who knows.  Like these are all very flawed, but knowing at least how little effort I put into this (they were just tests to see what it could do) I really think much better is possible with what we have right now. BrainDance fucked around with this message at 03:13 on Feb 2, 2024 |
|
#
?
Feb 2, 2024 03:06
|
|
|
Wonder if the canny pre-procesor is the answer here. Hadn't even occurred to me. Resolution scaling is going to be a bitch but it might be more workable. If i feel up to that I might try some experiments.
|
|
#
?
Feb 2, 2024 03:16
|
|
|
Hmm wonder if there's a good img2img related workflow that could take a very very simple input sketch and battlemap it more reliably.
|
|
#
?
Feb 2, 2024 19:01
|
|
|
Bard does images now but the resolution is tiny.
|
|
#
?
Feb 3, 2024 14:29
|
|
|
Bing has been generating line art for my villains so that I can enjoy coloring them myself. I've been using Krita on my little Wacom tablet, so it's like a fun little coloring book page. Sometimes the lines are a little hosed so I have to correct things myself but you can see that I don't generally mess with them. Spoilers for my players if you are in this thread! First, Captain Fabio (original generation and my colored piece), the kobold pirate with the golden hair and the Shampoo Fabulous Cabala. The Roxon (hoe-shawn) Cabala are the equivalent of Devil Fruits in my two campaigns, where the user swallows a gullet stone from an ancient deity in exchange for power. So lots of One Piece influence for me recently. Palette for this one was pretty generic but I'm working on becoming more adventurous.   The King of Spades, righthand of Emperor Kismont Roxo and seeker of the Deck of Many Things. I picked my palette for this guy from the Book of Many Things cover, which I think turned out really well. I'm usually wary of having too many different colors in a single comp so I'm glad I did things differently this time.   Lord Kree, master of Northern Naval Command and possessor of the Comet Cabala. Palette was grabbed from colored panels of Admiral Akainu, and I also grabbed a fire texture online for a background. I'm not crazy about the background but I felt like needed a little something more for the flame. Also my tablet broke in the middle of the coloring, so I didn't really have the manual dexterity with the mouse to do the background justice anyway.   Fabio and Kree have each killed a player upon their introduction. The party autognome ranger tried to tank through an entire dungeon and was smashed to pieces by Fabio's living hair upon running into him. Last Tuesday, our blood domain cleric tabaxi vampire was hit by a comet from Lord Kree while trying to escape an enemy-surrounded building. This... may bode poorly for when the party meets the King of Spades for the first time on Sunday. Ultimately, I've been having a lot of fun with the Bing Image Creator lately. It's really inspired me to draw again. I've finally been working through my figure drawing book that I picked up so long ago, and I've been making non-AI fan art of some campaign characters. Not that I'm neglecting the fundamentals: my computer and sketchbook have been extensively filled with different lines and basic shapes. I'm really trying to use my whole arm when sketching now, but man can it be tiring sometimes. My shoulder and elbow are not used to doing so much fine manipulation. Anyhow, the AI stuff is plenty of fun to integrate into my amateur art futzing, and I'm hoping to start being a little more transformative with the generations in the future.
|
|
#
?
Feb 15, 2024 23:12
|
|
|
Oh, that's nice.
|
|
#
?
Feb 15, 2024 23:40
|
|
|
Megazver posted:Oh, that's nice. Thank you  I've ordered another tablet so perhaps I will have more to share in the future
|
|
#
?
Feb 17, 2024 16:57
|
|
|
What sort of prompt makes neat poo poo like that?
|
|
#
?
Feb 17, 2024 17:02
|
|
|
Fuzz posted:What sort of prompt makes neat poo poo like that? I don't recall the kobold one, but here are what I used for the other two: "Portrait of an aarakocra navy admiral with fire powers, simple style, thick lines, black and white, white background, dungeons and dragons" "Portrait of a robot sea captain named "The King of Spades", simple style, thick lines, white background, dungeons and dragons" The process is to then take this into Krita and convert white to alpha. I lay down a fun background color to establish a unified color scheme (your brain picks better colors when you're coloring on a non-white surface), and then go ham. I don't use flood fill because I like the rhythm of coloring things by hand, but flood fill probably works very well if you want to get your colors down quickly. e: prompt for kobold might have been this: "portrait of a stunningly handsome kobold named Fabio with long hair from dungeons and dragons, simple style, thick lines"
|
|
#
?
Feb 17, 2024 17:53
|
|
|
 So I made an unofficial Jurassic Park tabletop RPG which extensively uses AI imagery to create the illustrations in the book and the associated game materials such as action and item cards. Then I heavily edited, modified, and overpainted them to get them closer to my vision. I primarily used ChatGPT to generate the base-imagery, then used a combination of manual Photoshop, Photoshop's Generative Fill, a local Stable Diffusion model, and the online service getimg.ai to get them where I wanted them to be. Here's the Jurassic Park: Edge of Chaos itch.io page. And here's some of the cards for a preview of things:      I learned so much about AI tools during the course of this, and now view them as a truly invaluable part of my workflow. Ethically I'm against any of this stuff being used in commercial projects, but I think it's perfect for a fan project like this where everything is already derivative and used without permission. e: Oh yeah, this is kinda a cross-post from my Jurassic Park Choose Your Own Adventure thread that I'm seeing if people want to get going again. feedmyleg fucked around with this message at 22:35 on Mar 18, 2024 |
|
#
?
Mar 18, 2024 20:50
|
|
|
feedmyleg posted:
Seeing as photoshopping acquired images is considered some style art as a modern version of photo-collage work, working using AI images as the base is not unreasonable. You're still doing work to make them look the way you want them to.
|
|
#
?
Mar 18, 2024 22:04
|
|
|
|
| # ? May 2, 2024 03:36 |
|
|
Exactly—I've got a note in the book about AI art where I describe it as akin to collage as well. I more or less feel like that's where I've landed on the ethical debate. There can be a significant amount of artistry and craftsmanship that goes into an AI-derived work, just as there can be with any derivative work. But anything that uses AI, whether as a base image or as training data in the use of a Generative Fill-style tool, is a derivative work in the exact same way as using and modifying an existing work for a collage. There are some grey areas and fuzzy lines with artistry and ownership, but the only stuff I consider mine in there is the stuff I built from the ground up like the blueprints and the graphic design. Some images in there just have minor touchups, removing AI gremlins from the background, smoothings out surfaces with weird greeblies, or resizing different parts of the image to retain an accurate scale. But most images involved extensive modification and have very little resemblance to the original image. No matter what, I don't consider any of them my original works. If anything I feel like an art director, requesting particular images and giving notes. They're not my art, but I did have a heavy hand in their direction.
|
|
#
?
Mar 18, 2024 22:23
|
|