|
what upscaler are people using with the contrast lora? The ghost pepper has weird shiny spots when using swinIRx4 that I've never gotten before, Marge Ga-Ga has some weird shiny spots in the upscale too.   e: I think ESRGAN is doing this without popping doesn't look at good as swin when comparing 1 to 1 but this popped like hell in SWIN  e2: SWINx4 really hates this lora but this looks fine in esrgan again. - really love how clear that text got that's directly from SD 
pixaal fucked around with this message at 20:33 on Mar 3, 2023 |
#
?
Mar 3, 2023 19:44
|
|
|
|
| # ? May 28, 2024 05:51 |
|
|
What artist or style prompts for these? I really like the style
|
|
#
?
Mar 3, 2023 23:48
|
|
|
Your Brain on Hugs posted:What artist or style prompts for these? I really like the style Lmao I'm so tempted to keep it a secret but the artist part of the prompt was alphonse mucha. He's had some very interesting impacts on my results. Here's a Mona Lisa of "his" 
hydroceramics fucked around with this message at 01:00 on Mar 4, 2023 |
|
#
?
Mar 4, 2023 00:55
|
|
|
Boba Pearl posted:people are going to use it to crank it, but I'd use it to generate plotlines, or to assume a character in my story and roleplay out a scene as a world building excercise. Yeah thats kinda the use I see for it too. Having some fun with ChatGPT right now converting Gloranthia creatures to 5E and was wondering at what point I'd be able to have something local.
|
|
#
?
Mar 4, 2023 02:02
|
|
|
I continue to make progress on my project to auto-generate videos from a GPT prompt https://www.youtube.com/watch?v=9nbCJMgoeL8. https://github.com/IShallRiseAgain/AIStoryGen/blob/main/README.md, Now you can use styles in your prompt.
|
|
#
?
Mar 4, 2023 02:29
|
|
|
hydroceramics posted:Lmao I'm so tempted to keep it a secret but the artist part of the prompt was alphonse mucha. He's had some very interesting impacts on my results. Oh nice, Mucha was a big thing back in the v3 days, I used him a bunch for stained glass stuff back then
|
|
#
?
Mar 4, 2023 07:53
|
|
|
Your Brain on Hugs posted:Oh nice, Mucha was a big thing back in the v3 days, I used him a bunch for stained glass stuff back then I'm finding him very Rutkowski-like in v4 in that adding him into a larger prompt has a definite impact, even if the final result looks nothing like his style.
|
|
#
?
Mar 4, 2023 17:03
|
|
|
"Medieval Heavy Metal Band, Pyrotechnics" 
|
|
#
?
Mar 5, 2023 02:16
|
|
|
IShallRiseAgain posted:I continue to make progress on my project to auto-generate videos from a GPT prompt https://www.youtube.com/watch?v=9nbCJMgoeL8. https://github.com/IShallRiseAgain/AIStoryGen/blob/main/README.md, Now you can use styles in your prompt. Nice. That’s a pretty good data!
|
|
#
?
Mar 5, 2023 15:27
|
|
|
Jewels in the Trident by the Bobby B Brothers.
|
|
#
?
Mar 5, 2023 16:57
|
|
|
So, I'm messing around with animation stuff. First, I tried to do a deep fake on a image generated by stable diffusion:  I wanted to see if there was a decent free way to do it, so I tried using https://github.com/yoyo-nb/Thin-Plate-Spline-Motion-Model. Unfortunately there is no implementation that isn't a collab or doesn't require paying at the moment. The results weren't that great, but I was also using a clip of the videogame as the driving video. https://i.imgur.com/aIWFyl6.mp4 I also used 4x-UltraSharp to upscale it. I then tried the free trial of D-ID, https://i.imgur.com/tACGRq3.mp4 but while the results were better, you can't use a driving video and there are noticeable flaws with it. Here is another VTMB character using that method as a bonus https://i.imgur.com/kLY27ns.mp4. I then tried out the depth map extension for AUTO111 it has some real potential. it can generate 3d inpainted meshes so you can make videos https://i.imgur.com/8BfUWkH.mp4, https://i.imgur.com/B3cq0Lw.mp4, https://i.imgur.com/Kl5x5Z5.mp4. I have only very briefly dabbled with blender, but now I plan to gain some more experience so I can get better results. These tutorials are what inspired me to start messing around with this stuff, https://www.youtube.com/watch?v=PFybvdKH8hc and https://www.youtube.com/watch?v=0tFe9dashgI IShallRiseAgain fucked around with this message at 17:41 on Mar 5, 2023 |
|
#
?
Mar 5, 2023 17:38
|
|
|
IShallRiseAgain posted:So, I'm messing around with animation stuff. Really awesome stuff. I love the zoom with separate foreground and background layers, very cool.
|
|
#
?
Mar 5, 2023 18:00
|
|
|
       
|
|
#
?
Mar 5, 2023 19:55
|
|
|
wow, those are superb. how were they created?
|
|
#
?
Mar 5, 2023 20:15
|
|
|
Mozi posted:wow, those are superb. how were they created? I'm just using stable diffusion, my recipe for success has been using the protogen infinity model; https://civitai.com/models/4398/protogen-infinity-official-release I absolutely love the results I get out of it.
|
|
#
?
Mar 5, 2023 20:28
|
|
|
 I wish LDSR was faster, it takes ages for bigger images but it has the best looking upscale results by far, in my experience. Anyone know anything faster that can come close to it?
|
|
#
?
Mar 5, 2023 22:05
|
|
|
Now you too, can run a large language model on your very own hardware if you know where to  the models. This was doable with facebook's published code but required a gpu with 24gb vram. Now there is a fork to run it on CPU instead.https://github.com/markasoftware/llama-cpu Bold is what I prompted. quote:I like to post on the Something Awful forums because, frankly, I like to read other people's fics. I also have very good friends there who I correspond with on a regular basis. I've written some of my best stuff there and gotten excellent feedback. It only runs single threaded for me so that bit of text took 50 minutes to produce, but it really is exciting to be able to do that at all locally.
|
|
#
?
Mar 6, 2023 00:54
|
|
|
that's the kind of lateral solutioneering we need it for. it says, "I'm gonna say some non sequiturs now," it says! it says some nonsequiturs. did the job boss
|
|
#
?
Mar 6, 2023 02:08
|
|
|
I took a break from AI stuff for a while and now I'm kind of interested in looking into it again. But I was wondering if there's any kind of definitive general-purpose model in use these days, or if that term even has meaning anymore now that models have proliferated so much. I stuck with vanilla SD 1.4 and haven't really tried anything else.
|
|
#
?
Mar 6, 2023 02:28
|
|
|
MechanicalTomPetty posted:I took a break from AI stuff for a while and now I'm kind of interested in looking into it again. But I was wondering if there's any kind of definitive general-purpose model in use these days, or if that term even has meaning anymore now that models have proliferated so much. I stuck with vanilla SD 1.4 and haven't really tried anything else. There aren't any really standout "great at everything" models but I would say basically every model published recently is a straight improvement over 1.4. I would say that Deliberate and Illuminati are the two front runners for cinematic/photo/quasi real output right now. Deliberate is a 1.5 model so will work with LoRA and LoCon models currently published, Illuminati is based on 2.1 so has higher native resolution out of the box without any extra tricks. Lora models are basically add on modules you can plug into any compatible model to get more consistent results for a specific thing; characters, clothing, art styles, etc. There are a handful that have been built for 2.1 but an order of magnitude fewer than for 1.5. For art or anime there are a million models all specialized in different stuff. For anime the various Orange mixes are great but you have to put in a lot of negative prompts sometimes to get SFW content out. The Protogen models are also very good for more artwork or cinema style results. Generally I recommend grabbing one of those to get you started and then just sort Civitai by newest and trying out things as you go. Objective Action fucked around with this message at 03:00 on Mar 6, 2023 |
|
#
?
Mar 6, 2023 02:56
|
|
|
Rahu posted:Now you too, can run a large language model on your very own hardware if you know where to I managed to make the 7b model run using https://github.com/oobabooga/text-generation-webui/ with 10gb of vram. This took maybe 30-60 seconds to generate, which is wild. quote:I like to post on the Something Awful forums because, let's face it, they're the coolest internet forums around, but I also like to use it to keep up with the gaming scene on the internet. And there's a lot of talk about the PS3. I'm not really that interested in it; I just don't see the point of this new generation of consoles. I'm not even going to bother with the 360 until after it's released. I just think it's kind of pointless to spend $300-400 on a new console that's going to have very few games, if any, for the first year. So I'm just going to wait for the next wave of games, which will probably come out 2 years after release. I'm not in any hurry to get a PS3 either, although I'm definitely going to get one. textgen feels like the next frontier, considering how expensive it is/was to run compared to imagegen.
|
|
#
?
Mar 6, 2023 03:00
|
|
|
Objective Action posted:There aren't any really standout "great at everything" models but I would say basically every model published recently is a straight improvement over 1.4. Thanks, I'll give that a shot! It's really impressive to see how far things have come in such a short time. Not even a year ago I was plinking random phrases onto NightCafe just to see what goofy poo poo it would put out. Now there's actual tools to handle poo poo like actual poses and depth and a ton of different specialized models to play with, all on the comfort of my home computer.
|
|
#
?
Mar 6, 2023 03:02
|
|
|
MechanicalTomPetty posted:Thanks, I'll give that a shot! It's really impressive to see how far things have come in such a short time. Not even a year ago I was plinking random phrases onto NightCafe just to see what goofy poo poo it would put out. Now there's actual tools to handle poo poo like actual poses and depth and a ton of different specialized models to play with, all on the comfort of my home computer. and don't forget about LoRAs which have kinda taken the place of trained whole models. Why train a whole model with a multi-gig file when you can get something a fraction of the size that does the same thing that you can also use on any model within the generation the LoRA is trained on?
|
|
#
?
Mar 6, 2023 03:11
|
|
|
Hey Kakermix, is there a way to change the position of the sun in the sky through controlnet? I have an idea I want to try with a very specific position.
|
|
#
?
Mar 6, 2023 03:37
|
|
|
Hey, is anyone following any of the legal stuff concerning AI and image generating? I remember there was some buzz with multiple people planning to go to court. Has there been any news on that front? You guys seem to be following these things very closely and maybe could point me to something that a google search wouldn't. Im honestly pretty anxious about a negative (IMO) precedent being set and only Disney being able to use AI or something nuts like that.
|
|
#
?
Mar 6, 2023 06:43
|
|
|
No court cases yet, that poo poo takes months to years. The US copyright office did make an official response to that one AI comic book, Zarya of the Dawn, where they said basically "you can copyright this comic book but it only affects the human-made bits like layout and dialogue, not the Midjourney art. There was a Class-action lawsuit filed against Stability AI and Midjourney (and also deviantart for some reason?) mid-January, but it hasn't gone anywhere yet. And of course, even if a precedent is set, that doesn't really stop stuff. The model weights are already trained and out in the world, so even if Stability gets on the hook for copyright law you'll still be able to generate stuff.
|
|
#
?
Mar 6, 2023 07:06
|
|
|
I’m most excited for games with Turing-passable generated AI dialogue. Based on the size of LLaMA, you could do it fairly well now… with 2 rtx 3090s or better. You’d need one for graphics and one for AI. Perhaps someone releases something with rudimentary graphics in a couple of years, like an enhanced visual novel with set events but voiced/modifiable input and output. You can get just the dialogue now with AI Dungeon, but that just doesn’t have a long-lasting enough memory to make “real games,” IMO. They’ve got to be working on tricks to make that better, like having it write hidden “summary memories” to prepend to later events or conversations. Imagine Skyrim where you just speak with a microphone and the NPCs respond as well as ChatGPT, ideally even controlling their actions. The language models are specifically designed to “play characters” like that. A lot of the current jank isn’t actually a limitation of the models, it’s just the character they’ve been assigned as is an HR functionary.
|
|
#
?
Mar 6, 2023 07:55
|
|
|
RPATDO_LAMD posted:No court cases yet, that poo poo takes months to years. The US copyright office did make an official response to that one AI comic book, Zarya of the Dawn, where they said basically "you can copyright this comic book but it only affects the human-made bits like layout and dialogue, not the Midjourney art. Notably also although they didn't say you COULD copyright other ai art program outputs their specific reasoning was relatively midjourney specific. Anything with any sort of inpainting would probably be copyrightable under the standard they set.
|
|
#
?
Mar 6, 2023 08:06
|
|
|
Monglo posted:Hey, is anyone following any of the legal stuff concerning AI and image generating? I remember there was some buzz with multiple people planning to go to court. Has there been any news on that front? You guys seem to be following these things very closely and maybe could point me to something that a google search wouldn't. It'll be awhile before those cases actually get to court; no, getty images won't be winning 1.8 TRILLION dollars. The other class action lawsuit doesn't even know how the software works that they're suing about so I'm not expecting much out of that one either. As for the recent decision on Zarya of the Dawn https://processmechanics.com/2023/02/22/a-mixed-decision-from-the-us-copyright-office/ is a good read on what it means going forward. I personally agree that the standard should be a bare bones floor function "modicum" of input, merest sliver. Mainly because I would never EVER tell someone they didn't put enough work into their art. I don't know what the copyright offices requirement of "substantial" involvement means, no one does at this point, it would require the author taking the copyright office to court. More likely a new registration of another piece could set a different precendent, there are certainly enough new tools and methods in the brief months since the comic came out that things have changed drastically. https://www.npr.org/2022/10/12/1127508725/prince-andy-warhol-supreme-court-copyright This, however, is the case you probably want to follow. This has wide reaching implications for the entirety of "transformative fair use". I have a lot to look up to see just how bad this case could end up being. So yeah, barely scratching the surface, haven't even brought up PR firms manufacturing fake outrage or corporations conning people into paying for lobbyists to remove their own rights, really exciting stuff going on.
|
|
#
?
Mar 6, 2023 08:47
|
|
|
Question Time posted:Imagine Skyrim where you just speak with a microphone and the NPCs respond as well as ChatGPT, ideally even controlling their actions. The language models are specifically designed to “play characters” like that. A lot of the current jank isn’t actually a limitation of the models, it’s just the character they’ve been assigned as is an HR functionary. Anything AI generated should probably have a somewhat limited amount of agency in game-relevant decision making. Prompts would be best generated procedurally the old fashioned programmatic way, then fed into a bot that re-contextualizes the information as, say, NPC dialog. This would prevent instances from running too long and prevent the generator from making stuff up, as well as putting control of NPCs in the hands of the designers instead of the whims of the generator. It is exciting though, particularly as it allows for games to re-introduce personality simulation, which was a game mechanic right up until AAA RPGs required voice acting. Oblivion tried to do it regardless, but to have a character be utterly pissed at you according to their stats, then have voiced dialogue that talks to you as if you were a close friend just didn't work.
|
|
#
?
Mar 6, 2023 12:14
|
|
|
KwegiboHB posted:It'll be awhile before those cases actually get to court; no, getty images won't be winning 1.8 TRILLION dollars. The other class action lawsuit doesn't even know how the software works that they're suing about so I'm not expecting much out of that one either. 
|
|
#
?
Mar 6, 2023 15:17
|
|
|
SCheeseman posted:This would prevent instances from running too long What do you mean by instances here? You can entirely set how many tokens it spits out as a maximum (and as a minimum) so that's like the one thing you can easily determine. You can also make it very uncreative which would probably keep it sane and mostly predictable, but, yeah, just mostly and that still might be enough unpredictability to make it unsuitable for a lot of situations, especially if the user is free to input anything.
|
|
#
?
Mar 6, 2023 15:51
|
|
|
Meant sessions I guess? Tokens are chunks of input, correct? Which brings to mind a commercial transaction which I guess it is, but if the goal is to run it locally eventually then there should probably be a better word for that. My experience with ChatGPT is that it's unreliable at retaining information over the course of a conversation/session, so games spoonfeeding it as much as possible and frequently scrubbing the slate clean would get more predictable results than letting it free-associate. Games still need writers, I don't want NPC conversations to be entirely generated. The game could give ChatGPT human written script to incorporate into responses, using the generator to naturalize and blend stiffly generated algorithmic dialog with pre-written dialogue. Eg. If you ask an NPC whether you can borrow their horse, the game makes the decision based on game stats or dice rolls, with the generator only being used to contextualize that decision in the form of dialogue. SCheeseman fucked around with this message at 16:40 on Mar 6, 2023 |
|
#
?
Mar 6, 2023 16:27
|
|
|
Dialog for games is tricky because you run into an interesting problem where you have to constrain the output so much and feed it so much context that it’s not even clear how it would save any time or effort compared to just writing it all out by hand. Let’s say you have a fantasy game, and a player asks about the king… if you don’t give the NPC context information about the king, then the AI will hallucinate info, and even if you fed the AI a Wikipedia page about the king, users would be able to probe the edges of the information easily with their own questions. You could (in principle) have the AI just not answer if it doesn’t have sufficient context… but now you’re constrained completely by how much context you can feed into the NPC. Maybe it will work some day when these models can fully internalize a huge wiki full of characters, events and connections.
|
|
#
?
Mar 6, 2023 16:37
|
|
|
Depends on what kind of story / game you want to make. If it's open world it doesn't matter if in my play the kind turned out to be a dragon just because I spread a rumor that he was around the world and in most games he's a demon or something. You need models for all of these unless you are having an AI make them on the fly. Something like Dwarf Fortress or RimWorld I could easily see being AI driven. Something more structured you could just have them comment on player actions. "you've slept at this inn for 60 days and just been hunting bears! It's helpful and all but is there's a dark lord to overthrow?
|
|
#
?
Mar 6, 2023 16:42
|
|
|
The generator doesn't need a wiki in memory if the game knows the specific information required to generate an appropriate response, which could be done with techniques as simple as scraping keywords.
|
|
#
?
Mar 6, 2023 16:45
|
|
|
I have a giant, long-running Chat-GPT session that contains a massive amount of context about an RPG world that I've used to make a rulebook. I can get easily use it to generate dialogue, scenes, scenarios, etc. If I want to have a "playground" that doesn't mess up that session, I simply copy and paste the rulebook back into the chat, then it more or less creates a new instance of that world I can screw around in. It's great for creating narrative scenarios within that world, including dialogue. Does it save me time compared to just writing dialogue myself? Not particularly, but the time isn't really the point. It's the feedback loop of inspiration.
|
|
#
?
Mar 6, 2023 16:50
|
|
|
Lucid Dream posted:Dialog for games is tricky because you run into an interesting problem where you have to constrain the output so much and feed it so much context that it’s not even clear how it would save any time or effort compared to just writing it all out by hand. Let’s say you have a fantasy game, and a player asks about the king… if you don’t give the NPC context information about the king, then the AI will hallucinate info, and even if you fed the AI a Wikipedia page about the king, users would be able to probe the edges of the information easily with their own questions. You could (in principle) have the AI just not answer if it doesn’t have sufficient context… but now you’re constrained completely by how much context you can feed into the NPC. Maybe it will work some day when these models can fully internalize a huge wiki full of characters, events and connections. Depends on if you want the ability for the player to ask entirely freeform questions or not. If you're limiting the player's choices on what to ask, it's pretty easy even if you can't consistently guarantee good output. One of my little gently caress-around projects with the ChatGPT API has been an NPC dialog system and the player output includes tags so ChatGPT knows the general way to reply Player chooses "Where can I go to adventure?" System sends ChatGPT "Player: Where can I go to Adventure? (Answer-Dracula's castle, Emotion-Fear) ChatGPT sends back something like "NPC: You could go to Dracula's castle, but you'd never catch me going there!"
|
|
#
?
Mar 6, 2023 16:50
|
|
|
pixaal posted:Depends on what kind of story / game you want to make. If it's open world it doesn't matter if in my play the kind turned out to be a dragon just because I spread a rumor that he was around the world and in most games he's a demon or something. You need models for all of these unless you are having an AI make them on the fly. In general the problem isn’t so much about what information the NPC spits out as much as the consistency between conversations and NPCs. Like, real time dialog with an NPC isn’t very interesting if the things they say are inconsistent with the state of the world or other NPCs. Solving the memory/context problem is one of those Hard problems that will be solved eventually, but existing tech won’t cut it I think.
|
|
#
?
Mar 6, 2023 16:50
|
|
|
|
| # ? May 28, 2024 05:51 |
|
|
Lucid Dream posted:In general the problem isn’t so much about what information the NPC spits out as much as the consistency between conversations and NPCs. Like, real time dialog with an NPC isn’t very interesting if the things they say are inconsistent with the state of the world or other NPCs. Solving the memory/context problem is one of those Hard problems that will be solved eventually, but existing tech won’t cut it I think. Well if you are running it local you just have the AI check the game state, and if it's unknown it's being decided on right now. Doing it purely in GPT would be silly.
|
|
#
?
Mar 6, 2023 16:51
|
|