|
Beeftweeter posted:over the weekend i've been playing with "draw things" for macos/ios over the weekend and it's pretty cool: a (totally free!) fully configurable stable diffusion implementation that can import models and poo poo directly from civit.ai lol it may be if you try to feed an Upscaler a Generative Model it will spit out some weirdness, basic Stable Diffusion you have to drop the models in the correct program folders so you know what's being accessed by which technology, but I could see a 3rd party prog just doing its best  Been a year-ish since I last looked, but I still use https://openmodeldb.info/models/4x-NickelbackFS as my go-to Upscaler Specific model in SD when I want to add a little of the face refinement tool to the output, but to be honest all my 'gently caress around' upscales these days go through: https://github.com/upscayl/upscayl as it's 99% as good - it uses all the same base SD Upscale models - and launches twenty times faster than full SD on my old PC, I can be in, done and out before SD has decided if it needs to pull any updates on launch. edit: I was looking into getting the NICKLEBACK4X Upscale model into UPSCAYL, but it needs converting inside ChaiNNer which looks like a bit of a gently caress around; if anyone else wants to do it if you already have a ChaiNNer install and post the output that would be appreciated. Upscayl Custom Models: https://github.com/upscayl/custom-models/blob/main/README.md Converting ESRGAN to NCNN for Upscayl using ChaiNNer: https://github.com/upscayl/upscayl/wiki/Model-Conversion-Guide NoneMoreNegative fucked around with this message at 16:34 on Mar 19, 2024 |
#
?
Mar 19, 2024 16:26
|
|
|
|
| # ? Apr 29, 2024 06:45 |
|
|
that's just the app I was hoping someone would make! I'm generating drunk Nixons at about the same speed as my 2080 super on a 14 pro
|
|
#
?
Mar 19, 2024 16:37
|
|
|
|
|
#
?
Mar 21, 2024 01:14
|
|
|
sounds about right for stanford, at least.
|
|
#
?
Mar 21, 2024 03:38
|
|
|
|
|
#
?
Mar 26, 2024 01:20
|
|
|
Agile Vector posted:that's just the app I was hoping someone would make! I'm generating drunk Nixons at about the same speed as my 2080 super on a 14 pro it's really pretty good, i've been using it to upscale stuff mostly but it's also pretty decent at generating people/things too, once you go to civit.ai and get a couple model packs + LoRA things to tweak your prompts. the interface is a little intimidating but if you know your terminology it's not hard to figure out at all the only thing i don't like about it is that enabling CoreML for everything (using all compute units, gpu, cpu, neural engine) worked during literally one app launch for me. it was blazing fast, probably 3-4x faster, but every subsequent time it has been just spitting out literal garbage (just colored noise), or the app crashes after giving me the first pass preview i guess a M1 ipad air doesn't have enough memory for it but who knows. you can't really get useful crash logs from ios anymore? anyway @NmN, i did this with Real-ESRGAN X4+  (more formats at https://archive.org/details/yospos-monday-upscale-fix, imgur doesn't seem to allow HEIC or AVIF anymore) there's a few other upscaler options in drawthings, but i haven't been able to figure out how to add anything additional to the ones they allow you to use by default e: oh and all of those other options are imo worse than Real-ESRGAN, unless you want to preserve hard edges (and you usually don't, i would think) Beeftweeter fucked around with this message at 18:09 on Mar 26, 2024 |
|
#
?
Mar 26, 2024 17:05
|
|
|
also if anyone is on the fence about trying drawthings, civit makes it super easy to import almost anything the app supports directly into it. the only thing that sucks about this functionality is that once it's downloading you cannot touch anything else in the interface or the download disappears and you have to restart it, and most model packs are ~2 GB+ just find something you want to import — e.g. dreamshaper, https://civitai.com/models/4384/dreamshaper?modelVersionId=128713 then hit that green play-looking button, scroll down to drawthings, then hit the blue one with the arrow. bam, done to work around the lovely download ui, you can also download things directly into a local directory (which you can specify, but i just use the default ios "Downloads" folder) and then import the stuff manually, but i haven't really had good luck doing that — i have the same problem with it just spitting out noise (without CoreML) for uhh i think 3/5 model packs i downloaded manually. but since i'm a noob at this i probably hosed up the import settings somehow, someone that knows what they're doing better than i do can probably figure it out e: drawthings does come with dreamshaper btw, but it's an older version (6.3 vs 8) and their included pack is quantized down to 8-bit. on that note, to save space you can quantize any model you want down to 8-bit within drawthings (it's in the model management modal), but i haven't tried this myself since i don't really have a need to, at least yet. the same modal lets you convert them to CoreML compatible models, and so does the LoRA management one. but if you just simply enable CoreML it will do it anyway automatically i've also found it helpful to use civitai's online generator to play with prompts, model packs and LoRAs before trying them locally. i'm not sure if you need an account to use it (probably), but i just used icloud hide my email. you don't even need to provide a password Beeftweeter fucked around with this message at 17:54 on Mar 26, 2024 |
|
#
?
Mar 26, 2024 17:15
|
|
|
hmm i think the jpeg2000 version has the best quality/size tradeoff, it's smaller yet as lossless as the PNG, but it doesn't seem to want to use the right color profile: then again so is the AVIF, and it's using the right ICC profile, but it's not well supported:  compare both to the supposedly lossless HEIC and you'll see it actually isn't lossless:   i mean sure, it's practically lossless, but that aint actually lossless  e: to be fair the HEIC is 742 KB while all the others are at least (or in jp2/png's cases, well over) 1 MB, but still. i wonder if i can get ffmpeg to use x265 to spit out a HEIC, that was using videotoolbox (apple's encoder) Beeftweeter fucked around with this message at 18:32 on Mar 26, 2024 |
|
#
?
Mar 26, 2024 18:21
|
|
|
https://twitter.com/birchlse/status/1772540619665166587 https://twitter.com/mishateplitskiy/status/1769433162122232127
|
|
#
?
Mar 26, 2024 19:50
|
|
|
so the other two students will be booted from the group for lying, then?
|
|
#
?
Mar 26, 2024 19:52
|
|
|
lol this new upcoming TIMEX watch is giving me 'AI Designed' vibes
|
|
#
?
Apr 1, 2024 12:55
|
|
|
NoneMoreNegative posted:lol this new upcoming TIMEX watch is giving me 'AI Designed' vibes global warming series? it's an interesting gimmick, but i i'd probably like it more if it weren't all warped looking
|
|
#
?
Apr 1, 2024 14:07
|
|
|
NoneMoreNegative posted:lol this new upcoming TIMEX watch is giving me 'AI Designed' vibes what time is it? iT's TeN pAsT tEN
|
|
#
?
Apr 1, 2024 14:24
|
|
|
draw things was just updated to include more model packs available for easy download and some LoRAs: in doing so every model and LoRA i'd downloaded previously seems to be gone. idk if they're still there and just not showing up on the updated list or if i need to download a bunch of 2-8 GB files again, but either way this loving sucks luckily a lot of the ones i had downloaded are now built-in presets, but this is still really loving annoying and disappointing
|
|
#
?
Apr 3, 2024 15:38
|
|
|
maybe a buggy update? it kept mine (confirmed in Settings > General > Storage) but the previously working config causes it to crash consistently, so I'll probably just wipe and redownload anyways :\
|
|
#
?
Apr 3, 2024 15:50
|
|
|
definitely a buggy update, it won't import anything i'd previously downloaded. i'm just redownloading a few models but even that seems buggy, it started the same download over again 3 times now. i think i'm just gonna wipe it and start fresh too if it keeps doing it although i didn't know there was a SDXL version of dreamshaper, so i guess it's not all bad. they even quantized it to 8-bit e: yeah it keeps redownloading the first out of 1 of 4 packs for SDXL dreamshaper, so bye bye install Beeftweeter fucked around with this message at 16:10 on Apr 3, 2024 |
|
#
?
Apr 3, 2024 16:07
|
|
|
yep so a fresh install works a ton better. it downloaded dreamshaper SDXL properly and once i pointed it to my downloads folder, imported some of the models i previously had and it seems all of the LoRAs. so that's not too bad i guess, converting to CoreML was probably the longest part of the process
|
|
#
?
Apr 3, 2024 16:36
|
|
|
this is just speaking a truth that people are too ashamed to put to words
|
|
#
?
Apr 6, 2024 20:59
|
|
|
Huh, I did not know that TV manufacturers were putting AI upscaling directly into the hardware in the display pipeline, I thought you'd need a big PC graphics card to do the realtime processing NVIDIA have been touting. https://www.youtube.com/watch?v=GVqEzw5qreQ
|
|
#
?
Apr 6, 2024 22:00
|
|
|
Doesnt require much to rebrand your pixel interpolation algorithm as AI and then charge more
|
|
#
?
Apr 6, 2024 22:12
|
|
|
The nVidia Shield has had AI upscaling since 2019, allegedly supposed to be good?
|
|
#
?
Apr 6, 2024 22:15
|
|
|
Yeah isn't this like the Nvidia dlss things that's been baked in for a while? The difference being that the cores for it are directly allocated because they aren't there to do anything else
|
|
#
?
Apr 6, 2024 23:37
|
|
|
hah, reminds me of when we told the investers that ai ws wqy eye
|
|
#
?
Apr 7, 2024 00:29
|
|
|
https://www.youtube.com/watch?v=JRBBcoANRIE
|
|
#
?
Apr 7, 2024 01:15
|
|
|
thank you for this
|
|
#
?
Apr 7, 2024 06:06
|
|
|
https://mailman.nanog.org/pipermail/nanog/2024-April/225407.html posted:As I think I have mentioned before, I have the world's lamest content farm
|
|
#
?
Apr 11, 2024 21:46
|
|
|
i don’t understand what his website is for frankly trapping gpt crawlers sounds good
|
|
#
?
Apr 11, 2024 21:52
|
|
|
maybe that guy should not be running "the world's lamest content farm" there, problem solved
|
|
#
?
Apr 11, 2024 21:53
|
|
|
Beeftweeter posted:maybe that guy should not be running "the world's lamest content farm" imo he should keep running it
|
|
#
?
Apr 11, 2024 21:58
|
|
|
idk, if that's the not-ai internet maybe the ai one will be better
|
|
#
?
Apr 11, 2024 21:59
|
|
|
Carthag Tuek posted:imo he should keep running it i, for one, think content farms are bad and should die
|
|
#
?
Apr 11, 2024 22:00
|
|
|
Beeftweeter posted:i, for one, think content farms are bad and should die try clicking the link
|
|
#
?
Apr 11, 2024 22:04
|
|
|
e: that is, i think "honey trap for crawlers" is probably more accurate than "content farm" in this case
|
|
#
?
Apr 11, 2024 22:06
|
|
|
Carthag Tuek posted:e: that is, i think "honey trap for crawlers" is probably more accurate than "content farm" in this case why is he trying to get crawlers out of the honey trap if that is what he built it for is what i don't understand
|
|
#
?
Apr 11, 2024 22:41
|
|
|
Carthag Tuek posted:try clicking the link i did, there's basically no content on the page. i assume that's because i block ads with a browser plugin and also on the dns level. i don't know if this guy's site is one of those pages that exist solely for people to find using keywords and just serves them useless content with a ton of ads, but if it is: gently caress that guy and the advertising companies that enable that bullshit. it's part of what ruined google novamute posted:why is he trying to get crawlers out of the honey trap if that is what he built it for is what i don't understand yeah exactly, i have a feeling he wants search engines to crawl it, but is concerned about his bullshit nonsense content making it into LLM training data
|
|
#
?
Apr 11, 2024 23:01
|
|
|
novamute posted:why is he trying to get crawlers out of the honey trap if that is what he built it for is what i don't understand he's the author of the book "the internet for dummies" and this is one of the ways he gets in contact with dummies to find out what they're thinking another way is by moderating the comp.compilers usenet group lol
|
|
#
?
Apr 11, 2024 23:34
|
|
|
In my 'Subscriptions' page rn, I don't ever remember subbing to this chanel but lol https://www.youtube.com/watch?v=GXH3yVPe5K4 (A live AI stream of Joe Rogan and Ben Shapiro) edit: Ah its the guy that did the AI Spongebob stream. Elon and Neil Degrasse Tyson have also been on the show so far. NoneMoreNegative fucked around with this message at 00:58 on Apr 12, 2024 |
|
#
?
Apr 12, 2024 00:31
|
|
|
lol joe talking to himself about the welsh village with the long name but ah well i dont care much about the ~content farm~ i just think its funny. its like the library of babel site but stupider https://libraryofbabel.info
|
|
#
?
Apr 12, 2024 05:26
|
|
|
...! posted:the problem is definitely solved
|
|
#
?
Apr 12, 2024 14:19
|
|
|
|
| # ? Apr 29, 2024 06:45 |
|
|
song generation is honestly getting pretty insane. i'm not saying this is "good" but given how early it is, i expect things to get crazy. i generated a brief broadway interlude about rotor: http://sndup.net/p5g7 lol it actually generated two of them: http://sndup.net/p5g7 mediaphage fucked around with this message at 15:03 on Apr 12, 2024 |
|
#
?
Apr 12, 2024 15:01
|
|