|
busalover posted:Google Dreamix
|
# ? Feb 5, 2023 13:19 |
|
|
# ? May 30, 2024 09:52 |
|
Models trained on photos will do better photographic outputs, and models blended from those models will often see a benefit. Dreamlike Photoreal does pretty well in general and wavymulder on huggingface/civitai has some great models trained on actual photos with different specific aesthetics. Almost any prompt that's camera or film related is going to make a difference e.g. 100 ISO, f/8 or whatever, 35mm, polaroid, instant camera, Kodachrome, Tintype, taken with SLR, etc. if there's a specific look you're after then focus on that but if not the more general terms that get used the more photographic the output, up to a point. If you go overboard or push the weighting too hard it'll just put bizarre AI-generated cameras into the output instead. and anything for non-camera related art in the negative can make a difference e.g. illustration, cgi, 3d render, pencil, anime. Hell maybe even put "photorealistic" as a negative because as pointed out, actual photos don't get that tag. KwegiboHB posted:Birthday request: everyone post monster trucks. Thank you! evil monster monster truck muppet Negative prompt: completely normal, friendly, cgi, painting, drawing, illustration, anime, cartoon, game, black and white, truck ((monster)), intricate detail, 35mm color photo by [[[hr giger]]], anthropomorphic evil vehicle, 100 ISO, f/8 or whatever, 35mm, instant camera, Kodachrome, Tintype, taken with SLR Evil beast monster wild monster truck with a human face as the grille , postwar color newsreel footage by HR giger and Zdzislaw Beksinski Negative prompt: completely normal, friendly, cgi, painting, drawing, illustration, anime, cartoon, game, black and white evil beast monster wild monster truck with a human face as the grille, postwar color newsreel footage by HR giger and Zdzisław, no negative prompt, anime model The Sausages fucked around with this message at 14:14 on Feb 5, 2023 |
# ? Feb 5, 2023 14:00 |
|
KwegiboHB posted:
"Bioluminescent Monster Trucks with teeth and organs, drawn by zdzisław beksiński " "Biomechanical Monster Trucks with teeth and organs, drawn by zdzisław beksiński, futuristic theme " "photo, Biomechanical glowing neon Monster Trucks with teeth and metal organs" Vlaphor fucked around with this message at 19:33 on Feb 5, 2023 |
# ? Feb 5, 2023 17:58 |
|
John Berkey and HR Giger:
|
# ? Feb 5, 2023 18:45 |
|
It would be fun to do a daily/weekly prompt or something.
|
# ? Feb 5, 2023 19:00 |
|
Mozi fucked around with this message at 19:36 on Feb 5, 2023 |
# ? Feb 5, 2023 19:29 |
|
|
# ? Feb 5, 2023 20:08 |
|
I wonder why OpenAI even bother with an AI-driven detection tool. It would be easier to just have a service that searches through their logs (because you know they keep logs) and says "this came from us" or "this did not come from us"
|
# ? Feb 5, 2023 20:16 |
|
ymgve posted:I wonder why OpenAI even bother with an AI-driven detection tool. It would be easier to just have a service that searches through their logs (because you know they keep logs) and says "this came from us" or "this did not come from us" It's pure PR. Yeah, we're building Roko's Basilisk, but at least we're being ETHICAL about it.
|
# ? Feb 5, 2023 21:03 |
|
ymgve posted:I wonder why OpenAI even bother with an AI-driven detection tool. It would be easier to just have a service that searches through their logs (because you know they keep logs) and says "this came from us" or "this did not come from us" full text search is relatively expensive, this is much cheaper for OAI
|
# ? Feb 5, 2023 21:05 |
|
Some more Shadowrun kick. Think I was reasonably successful getting a happy Elf woman on her bike and the jovial troll telling jokes, without the faces being ultra-contorted like MidJourney likes to do. Stable Diffusion is a lot better about getting the 'correct' faces for emotions.
|
# ? Feb 6, 2023 01:38 |
https://www.stableattribution.com/ seems interesting (if you skip the overblown story)
|
|
# ? Feb 6, 2023 01:48 |
|
lunar detritus posted:https://www.stableattribution.com/ seems interesting (if you skip the overblown story) I think this is dumb and just pushing more people to not understand how the system works also lol that none of the human generated pictures have attributions, if you click on them it's like "know who this is? tell us" so all that site is really doing is sharing people's artwork without attribution while pushing the idea that AI generated artworks are created out of a handful of existing works
|
# ? Feb 6, 2023 01:53 |
|
lunar detritus posted:https://www.stableattribution.com/ seems interesting (if you skip the overblown story) I'm not sure how exactly this works? It seems like it's just generating a CLIP caption for images and then searching through the dataset for similar images? https://www.stableattribution.com/?image=65fd7049-fd0d-437e-9239-7e3f7d002878 this was an old week-1 generation, prompt was just danny devito in the style of a marble statue by greg rutkowsky so in principle, the way diffusion training works, the "attribution" should include all images of statues and all Greg R. art that was included in the datasets but all it comes up with are portraits and caricatures of Devito because they have similar subject matter here's another image i generated: https://www.stableattribution.com/?image=ab48ca1c-f1d1-42bf-8778-6fce61ffb703 (dreamlikeart:1.1), intricate digital painting of a dark cultist casting a mysterious spell, (hood over face, eyes cloaked in shadow:1.2), sinister red gown, (multicolored satanic pentacle:1.2), stone dungeon walls, guttering torches, deep shadows, chiaroscuro, multicolored, particle effects, bloom, smoky, from below, dramatic, imposing framing Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, [blurry], CAPTCHA, (((deformed))), bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing and again, the attribution thing looks for images with similar theming and framing but absolutely doesn't cover all the attribution. for example, I can search for "multicolored satanic pentacle" on this LAION dataset search tool and come up with all of these: every single one of those images (and millions more) contributed to the model's weights for terms like satanic, multicolored, and pentacle (and combinations of the three). but they aren't visually similar to the output image so stableattribution.com doesn't find them. I also tried some other AI-generated images from my folder and got this: TIP posted:also lol that none of the human generated pictures have attributions, if you click on them it's like "know who this is? tell us" that part is fair enough tbh because they're using the same LAION dataset to search for images that sd was trained on, and the dataset itself doesn't have image attributions. edit: https://www.stableattribution.com/faq posted:How does Stable Attribution work? So yup, it's just searching the dataset for 'similar' images and claiming without justification that they're the ones that most influenced the output. RPATDO_LAMD fucked around with this message at 02:55 on Feb 6, 2023 |
# ? Feb 6, 2023 02:40 |
|
RPATDO_LAMD posted:that part is fair enough tbh because they're using the same LAION dataset to search for images that sd was trained on, and the dataset itself doesn't have image attributions. fair enough? it's a site made to whine about the fact that the "contributing artists" aren't being attributed in AI art, a made up problem, and their "solution" to their own made up problem is to embed a bunch of people's art on their website without attribution it's dumb as gently caress
|
# ? Feb 6, 2023 03:15 |
|
Thank you again everyone for the gift of amazing monster trucks. Here is a small but important look at the role they have played in history. monster trucks across time, ancient stone age monster truck made entirely out of rocks boulders and pebbles Negative prompt: bland, blurry, boring, captcha, clean, comfortable, cropped, modern, quiet, small Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1275490052, Size: 768x768, Model hash: 14e1ef5d monster trucks across time, roman colosseum bespoke marbled monster truck chariot pull Negative prompt: bland, blurry, boring, captcha, clean, comfortable, cropped, modern, quiet, small Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3319797228, Size: 768x768, Model hash: 14e1ef5d monster trucks across time, viking era monster truck helmeted and horned pining for the fjords Negative prompt: bland, blurry, boring, captcha, clean, comfortable, cropped, modern, quiet, small Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3270750424, Size: 768x768, Model hash: 14e1ef5d monster trucks across time, feudal japan samurai monster truck karuta armor dual katanas breathing dragon fire Negative prompt: bland, blurry, boring, captcha, clean, comfortable, cropped, modern, quiet, small Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 2564816492, Size: 768x768, Model hash: 14e1ef5d Fitzy Fitz posted:It would be fun to do a daily/weekly prompt or something.
|
# ? Feb 6, 2023 04:17 |
|
I tried installing Automatic1111 stable diffusion UI. Lol programmers are just the worst. I'm pretty sure part of it requires the newest python while the other part refuses to get pytorch if you use it. Anyway me bitching at programmers who don't make things a simple install isn't really the point of this thread but I wasted enough time that I get to complain about it.
|
# ? Feb 6, 2023 04:42 |
|
Stable attribution is neat because the results are completely made up. The fact that it doesn't know if an image is AI-generated but it will happily tell you which images contributed to it should make that obvious. This is what it made of my entirely original work do not steal: https://www.stableattribution.com/?image=60b01f27-6bbc-440c-9e93-785f50b208b5 A machine that incorrectly identifies the source of art to make sure everyone gets the credit they deserve. Truly beautiful.
|
# ? Feb 6, 2023 04:43 |
|
deez nuts
|
# ? Feb 6, 2023 04:58 |
|
KwegiboHB posted:Birthday request: everyone post monster trucks. Thank you! i'm late to the party "Noah's ark except it's a monster truck"
|
# ? Feb 6, 2023 05:08 |
|
Duck and Cover posted:I tried installing Automatic1111 stable diffusion UI. Lol programmers are just the worst. I'm pretty sure part of it requires the newest python while the other part refuses to get pytorch if you use it. Anyway me bitching at programmers who don't make things a simple install isn't really the point of this thread but I wasted enough time that I get to complain about it. did you try the binary release from two weeks ago? https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.0.0-pre quote:The webui.zip is a binary distribution for people who can't install python and git. Alternately there are a couple more user-friendly uis goons have been talking about and using: Megazver posted:There's auto1111, which is being developed by a crypto-racist eastern european weirdo, but has the most features by far. It's also a pain to install, but it's still the most popular UI just because of how popular it is.
|
# ? Feb 6, 2023 05:28 |
|
RPATDO_LAMD posted:did you try the binary release from two weeks ago? No I didn't. Thanks, maybe I will but I kind of want to figure this out. Yes I know. I've been using one of them. If you're program isn't capable of being installed than me than it is poo poo. It's that simple. Some programs feel like they are designed to be repellant to anybody but someone who programs for a living. Docker feels similar. Programs made by programmers without any concern for someone whose programming skills are poo poo let alone somebody who has just none. What I'm trying to say is bring back Hypercard. Duck and Cover fucked around with this message at 07:44 on Feb 6, 2023 |
# ? Feb 6, 2023 06:47 |
|
why would you even use docker if you aren't using it to program stuff ???
|
# ? Feb 6, 2023 06:54 |
|
RPATDO_LAMD posted:why would you even use docker if you aren't using it to program stuff Docker can make setting up an environment easier. (Although sometimes its a bit wonky for windows.)
|
# ? Feb 6, 2023 07:01 |
|
A command-line interface is an interface
|
# ? Feb 6, 2023 07:18 |
|
RPATDO_LAMD posted:why would you even use docker if you aren't using it to program stuff Sonarr, Radar, (which should be combined but nooooooo that'd make too much sense, also they kind of suck) Plex (also kind of sucks) etc. on a nas. One of these days I'll figure out how to update my images without nuking my settings. Probably not though as backing up the settings is easy enough and I don't think I care enough. Sir Mat of Dickie posted:A command-line interface is an interface Gag Sir Mat of Dickie. South. South. South. Duck and Cover fucked around with this message at 07:45 on Feb 6, 2023 |
# ? Feb 6, 2023 07:24 |
|
AI Seinfeld went transphobic, got banned by Twitch
|
# ? Feb 6, 2023 10:03 |
|
Monglo posted:AI Seinfeld went transphobic, got banned by Twitch it made the mistake of listening to chat
|
# ? Feb 6, 2023 10:25 |
|
Nothing, Forever's "AI off the rails" moment was almost certainly another crypto-racist eastern european weirdo dev taking over when others weren't looking and it'll be interesting to see how they lie through it. OpenAI won't care until journalists come asking for comments
|
# ? Feb 6, 2023 13:10 |
|
the nothing forever team says the rant was because they switched to curie after the davinci mode was having issues https://twitter.com/RyanBrianJones/status/1622548545898250242
|
# ? Feb 6, 2023 13:28 |
|
So AI did Kramer @ Laugh Factory, weird future.
|
# ? Feb 6, 2023 14:18 |
Ruffian Price posted:Nothing, Forever's "AI off the rails" moment was almost certainly another crypto-racist eastern european weirdo dev taking over when others weren't looking and it'll be interesting to see how they lie through it. OpenAI won't care until journalists come asking for comments Yeahhh, Curie is bad but it's not "makes transphobic comments out of nowhere" bad
|
|
# ? Feb 6, 2023 14:21 |
|
OpenAI is very hard to get off the rails offensive I'm not sure what it said but Twitch lets you get away with a ton. Open AI is the team afraid of exactly this happening, so I kind of expect them to comment if it hits them. There is 100% a way to interact more directly with that program I've seen clearly scripted references to chat play out when the dev account is talking in chat.
|
# ? Feb 6, 2023 14:23 |
|
Duck and Cover posted:If you're program isn't capable of being installed than me than it is poo poo. It's that simple. Some programs feel like they are designed to be repellant to anybody but someone who programs for a living. Docker feels similar. Programs made by programmers without any concern for someone whose programming skills are poo poo let alone somebody who has just none. This stuff is being developed against the backdrop of constantly changing versions of the tech and various related drivers. You can call it "poo poo", but the simple fact is, no one is paying anyone to make these frontends period, in most cases, and so definitely no one is thinking about usability outside of the context of "people with $1000 GPUs who, though they may despise everything about python dependency management, understand it enough to bash their way through it" (pun intended). I'd love it if it was easy for automatic/etc to spit out daily "just works" totally bundled versions of everything, but these people are already putting tremendous amounts of effort into maintaining OSS projects for, at this time, no clear gain. If you don't want to go deep enough to go under the hood, I think that's exactly the problem that MidJourney's paid tiers exist to solve, and basically anyone can go have a good experience there already. InvokeAI has a cooler frontend and a weaker backend and also part of me suspects they have some kind of pivot of some kind in mind. It's more "user friendly" until you start comparing it side by side to web-ui and then I think it often falls down. It's not reasonable to expect high-speed development against a rapidly morphing backdrop to distill into something that's a one-click install for people who hate venv on a regular basis, we're lucky to get periodic binary-bundled releases that still take some futzing with, IMO. For the record, I loving dislike the python dependency management paradigms and various other things about this, and yea, it annoys me that every time I do a git pull on stable I might break something and have to go inject some more modern version of the CUDA drivers into pytorch or something and it bothers me even more that I know what any of that poo poo means because I didn't get paid to learn any of it. This is just where we are though. If the hassle of early adoption is too taxing, just use MidJourney, it's often what I do if the SD landscape is especially broken. The various discords for these things are pretty good at providing support, too, as well as this thread It's a broken landscape with a lotta supportive folks around. quote:What I'm trying to say is bring back Hypercard. I think it still exists in the demoscene so it never really went away but ofc it's wildly unsuitable for this kind of application and also would give migraines to anyone whose spent the last five years dealing with pyton quote:future (computer programmer:1) using (hypercard:1) to achieve the singularity, 50mm wide angle photography, vivid colors
|
# ? Feb 6, 2023 14:24 |
|
pixaal posted:OpenAI is very hard to get off the rails offensive I'm not sure what it said but Twitch lets you get away with a ton. Open AI is the team afraid of exactly this happening, so I kind of expect them to comment if it hits them. This is the bit that got it banned. It's offensive in such an aggressively tepid way that it almost comes across like self-aware sarcasm. https://twitter.com/watmay1/status/1622518097067511811
|
# ? Feb 6, 2023 14:36 |
|
Cabbages and Kings posted:This stuff is being developed against the backdrop of constantly changing versions of the tech and various related drivers. You can call it "poo poo", but the simple fact is, no one is paying anyone to make these frontends period, in most cases, and so definitely no one is thinking about usability outside of the context of "people with $1000 GPUs who, though they may despise everything about python dependency management, understand it enough to bash their way through it" (pun intended). I'd love it if it was easy for automatic/etc to spit out daily "just works" totally bundled versions of everything, but these people are already putting tremendous amounts of effort into maintaining OSS projects for, at this time, no clear gain. If you don't want to go deep enough to go under the hood, I think that's exactly the problem that MidJourney's paid tiers exist to solve, and basically anyone can go have a good experience there already. InvokeAI has a cooler frontend and a weaker backend and also part of me suspects they have some kind of pivot of some kind in mind. It's more "user friendly" until you start comparing it side by side to web-ui and then I think it often falls down. But that's just what it is, this is really pretty cutting edge poo poo that hasn't been packaged into foolproof point and click tools by Adobe. Which is only a matter of time. Symphoric posted:This is the bit that got it banned.
|
# ? Feb 6, 2023 14:44 |
|
mobby_6kl posted:Yeah it's really a pain in the rear end and one of the reasons I've only used like 3 versions of Automatic so far. Just can't be bothered to invest more time into loving with it every time something is updated. I do a git pull most mornings, 95% of the time it's fine, the times its not it's just been "delete venv, let it recreate, then install proper CUDA drivers into pytorch if it got the wrong ones". It's a little tedious but honestly it's better than a number of sold-for-profit commercial applications I've worked on . I think I've only had to do an orbital nuke and reclone on the repo once. I'd update less, generally I am lazy on that poo poo, but I always log tail the main branch and it's like "wow, three useful things added in a day, hope they work!!" -- not like rolling back is hard. As long as you know what hash you were on.
|
# ? Feb 6, 2023 14:46 |
|
The only thing that's happened to me as a result of git pull on launch is that it prompted me to update pytorch which I think I did successfully manually but it still gives me the same "outdated torch version" message and running --update -torch or whatever in the command line isn't doing anything. In spite of all this it still runs and operates normally
|
# ? Feb 6, 2023 15:05 |
|
Dick Bastardly posted:The only thing that's happened to me as a result of git pull on launch is that it prompted me to update pytorch which I think I did successfully manually but it still gives me the same "outdated torch version" message and running --update -torch or whatever in the command line isn't doing anything. In spite of all this it still runs and operates normally I've found that the version of the CUDA drivers pytorch installs tends to be outdated, and manually dumping newer CUDA drivers into torch's own installed DLLs, increases my iterations per sec on a 4090 by ~20-40% depending on size/batching haven't done this in a couple weeks though, maybe it's finally pinned to the correct versions?
|
# ? Feb 6, 2023 15:50 |
|
|
# ? May 30, 2024 09:52 |
|
tom selleck driving a big red monster truck, clay sculpture
|
# ? Feb 6, 2023 16:00 |