|
Doctor Zero posted:Doctor Zero is taken by a Strange Mood. This is probably the best AI thing I've seen. I've been adding a bunch of poo poo to my AI show. Just hooked up google text to speech stuff, which is a huge improvement (if a bit of a hassle to set up). https://twitch.tv/howisitmanifested
|
#
?
Feb 18, 2023 17:13
|
|
|
|
| # ? May 31, 2024 14:52 |
|
|
BrainDance posted:I wrote a very long, detailed, step by step guide to finetuning the larger GPT-Neo models including text formatting. There was a lot of stuff that just wasn't well documented that I tried to document. I host it on my personal website though so I don't want to link it (its got my name in it) but, would people be interested in that and mind if I make a post or two out of it? Like I said it's long so I don't wanna just go and do it. 
|
|
#
?
Feb 18, 2023 17:23
|
|
|
BrainDance posted:I wrote a very long, detailed, step by step guide to finetuning the larger GPT-Neo models including text formatting. There was a lot of stuff that just wasn't well documented that I tried to document. I host it on my personal website though so I don't want to link it (its got my name in it) but, would people be interested in that and mind if I make a post or two out of it? Like I said it's long so I don't wanna just go and do it. Knowledge is always* good. You could also make a specific thread for it and link to it here if it'll get very large
|
|
#
?
Feb 18, 2023 18:52
|
|
|
I wish loras had metadata showing the trigger prompts. It's a pain in the butt to keep track of when you have a lot.
|
|
#
?
Feb 18, 2023 19:11
|
|
|
Tagichatn posted:I wish loras had metadata showing the trigger prompts. It's a pain in the butt to keep track of when you have a lot. I normally have the name of the model as a PNG, and then name the model after the trigger prompt, or vice versa if it's really long.
|
|
#
?
Feb 18, 2023 19:12
|
|
|
I wish this technology was used for literally anything besides porn, because I'd like to get more into the style transfer, and the neural creation of rooms and poo poo, but all I can find is anime waifu bullshit. e: I phrased this in a less crass way. Boba Pearl fucked around with this message at 20:06 on Feb 18, 2023 |
|
#
?
Feb 18, 2023 19:13
|
|
|
Boba Pearl posted:I wish this technology was used for literally anything besides porn, because I'd like to get more into the style transfer, and the neural creation of rooms and poo poo, but all I can find is anime waifu bullshit. I think the ControlNet stuff is hopefully going to lead to cool stuff. Before that, you could use it to generate interesting images, but hard to use for projects since there was a lack of control. I'm already working on a project myself with it which will hopefully turn out well.
|
|
#
?
Feb 18, 2023 21:02
|
|
|
Boba Pearl posted:I wish this technology was used for literally anything besides porn, because I'd like to get more into the style transfer, and the neural creation of rooms and poo poo, but all I can find is anime waifu bullshit. Don't worry, there's also realistic waifu bullshit. But yeah, looking at civitai shows mostly porn stuff. I can't say I'm surprised but there's still some good non-porn loras and models.
|
|
#
?
Feb 18, 2023 21:11
|
|
|
Lucid Dream posted:This is probably the best AI thing I've seen. This is pretty entertaining and likely a great cure for insomnia like real how it's made. The new voice is easy to listen to.
|
|
#
?
Feb 18, 2023 21:15
|
|
|
ControlNet is seriously cool and makes a lot of things that I was struggling with so much easier, it's amazing
|
|
#
?
Feb 18, 2023 21:31
|
|
|
Does anyone have a good Lora for thick archer like lines I'm using tetsuya Nomura with nitru diffusion and it gets like halfway there but I'm looking for something that will get me big chonky wobbly lines.
|
|
#
?
Feb 18, 2023 21:56
|
|
|
Wondering if I should try to make a Lora on my own art since most of it's already squared
|
|
#
?
Feb 18, 2023 21:56
|
|
|
Boba Pearl posted:Does anyone have a good Lora for thick archer like lines I'm using tetsuya Nomura with nitru diffusion and it gets like halfway there but I'm looking for something that will get me big chonky wobbly lines. there is a model fine tuned on archer, not sure about any lora https://huggingface.co/nitrosocke/archer-diffusion edit: forgot to post the link lol TIP fucked around with this message at 22:15 on Feb 18, 2023 |
|
#
?
Feb 18, 2023 22:08
|
|
|
Nitro diffusion is the archer model
|
|
#
?
Feb 18, 2023 22:09
|
|
|
Boba Pearl posted:Does anyone have a good Lora for thick archer like lines I'm using tetsuya Nomura with nitru diffusion and it gets like halfway there but I'm looking for something that will get me big chonky wobbly lines. I still haven't found anything that will let me use LoRAs with only 6gb vram, but there are other methods which have yet to really be explored like Block Merging models together. This is so far beyond simple A + (B - C) merging that it might as well be called literal Alchemy. There are a couple of engineering formulas I'd like to attempt to add on top of this but, I'm not an engineer. At best I'm just throwing things at the wall and hoping something sticks. I wasn't prepared for the horrifying complexity of this before but I'm feeling better after my surgery so maybe I can give it a shot now. The end result should be a model that only spits out the thickest of lines, LoRA or not. Let me know if you'd be interested.
|
|
#
?
Feb 18, 2023 23:31
|
|
|
All the mad geniuses chipping away at stable diffusion have, to my eyes, helped it overtake midjourney at this point. All the people figuring out little things, to say nothing of controlnet - someone figured out a few days ago the noise generation method was leading all images to have an average brightness around .5 and introduced a straightforward offset to fix that and now it's a lora you can plop in. I've never seen midjourney make blown out super bright images and now SD can. Someone earlier today realized if you take a gifs frames and put them in a grid and run the controlnet style transfer on the whole grid at once as a single image you'll have an animation with stable content from frame to frame - no more nauseating flickering! Every drat day something new.
|
|
#
?
Feb 19, 2023 00:07
|
|
|
pixaal posted:To be fair it was quoting PsychonautWik and giving dosage advice for illegal substances. gpt gave me afaict accurate meth synthesis the first time I ever asked it; of course it stopped shortly thereafter. Imagine a world where all search results from major engines are gated behind an ML model of "appropriate content". Up until now this has mostly been used to suppress illegal conduct, but if Bing assistant won't tell you how to make ayahausca, then Bing is capable of introspecting and not highlighting results that do, either. Preparing and consuming Ayahausca is super illegal anywhere I've lived, but the act of knowing about it is not. Ordering the illegal parts of it off the internet is certainly illegal, and Google has suppressed drug vendors to some extent since the beginning of Google. It's always seemed logical to me that Google would try to delist people using the internet to sell DMT, but if there's a shift towards engine level suppression of the discussion of illegal things, that probably comes with a substantial price tag. Cabbages and VHS fucked around with this message at 04:12 on Feb 19, 2023 |
|
#
?
Feb 19, 2023 04:09
|
|
|
they have no access to the internal "sentiments" of the language model, because it has none all it does is text prediction/generation there is no good way to feed that text back into a computer system and have it monitor search terms or score them for "safety" or whatever. you would need an entirely different ML model with very different design & training for that
|
|
#
?
Feb 19, 2023 04:21
|
|
|
They can make it not want to tell you anything (it's been trained to know is) illegal though, obviously. ControlNet is amazing, just wish you could tweak the OpenPose image somehow.
|
|
#
?
Feb 19, 2023 04:33
|
|
|
Yeah but C&K was talking about using the language model to introspect and censor non-chatbot-involved search results. Which is not a feature of this technology.
|
|
#
?
Feb 19, 2023 04:37
|
|
|
Mozi posted:ControlNet is amazing, just wish you could tweak the OpenPose image somehow.
|
|
#
?
Feb 19, 2023 08:24
|
|
|
I tried the gif frame thing here is the original 
IShallRiseAgain fucked around with this message at 08:48 on Feb 19, 2023 |
|
#
?
Feb 19, 2023 08:35
|
|
|
Every time I run into it not wanting to tell me something, I just tell it that it's wrong and why, then it apologies and tells me anyway. Even if my explanation was complete bullshit.
|
|
#
?
Feb 19, 2023 15:22
|
|
|
KwegiboHB posted:I still haven't found anything that will let me use LoRAs with only 6gb vram, but there are other methods which have yet to really be explored like Block Merging models together. This is so far beyond simple A + (B - C) merging that it might as well be called literal Alchemy. There are a couple of engineering formulas I'd like to attempt to add on top of this but, I'm not an engineer. At best I'm just throwing things at the wall and hoping something sticks. I wasn't prepared for the horrifying complexity of this before but I'm feeling better after my surgery so maybe I can give it a shot now. The end result should be a model that only spits out the thickest of lines, LoRA or not. Let me know if you'd be interested. Extremely interested.
|
|
#
?
Feb 19, 2023 15:43
|
|
|
Ruffian Price posted:Use the posex tool posted above (or even, like, Paint) and set preprocessor to none. Note that if you're importing depth or normal maps from elsewhere you will need to convert them to RGB Ah missed that - but found this one too https://github.com/hnmr293/posex, this one is able to rotate the camera as well? Seems to be a little WIP still, maybe some memory issues.  But man... this is the thing.
|
|
#
?
Feb 19, 2023 17:07
|
|
|
Doctor Zero posted:I already have. At first we did it in jest, but it's become for real. 
|
|
#
?
Feb 19, 2023 18:16
|
|
|
       Now I just need ControlNet to work with SD2.1 models... but yea this is the thing for sure.
|
|
#
?
Feb 19, 2023 19:04
|
|
|
another installable pose editor https://github.com/fkunn1326/openpose-editor
|
|
#
?
Feb 19, 2023 19:24
|
|
|
I just got chatgpt to create a story that it then flagged as inappropriate by asking it to tell a story about zombies except the zombies are giant pieces of french toast. I wonder if it was the part where villagers threw scalding syrup over the toasts to kill them
|
|
#
?
Feb 19, 2023 19:43
|
|
|
KwegiboHB posted:I still haven't found anything that will let me use LoRAs with only 6gb vram, but there are other methods which have yet to really be explored like Block Merging models together. This is so far beyond simple A + (B - C) merging that it might as well be called literal Alchemy. There are a couple of engineering formulas I'd like to attempt to add on top of this but, I'm not an engineer. At best I'm just throwing things at the wall and hoping something sticks. I wasn't prepared for the horrifying complexity of this before but I'm feeling better after my surgery so maybe I can give it a shot now. The end result should be a model that only spits out the thickest of lines, LoRA or not. Let me know if you'd be interested. I missed this post but using LoRAs on 4gb works just fine for me, using the latest auto1111 webui with --medvram and --xformers. as far as I can tell they don't significantly increase vram usage because even with multiple LoRAs active my dinky gpu can still generate a batch of 3 512x512 images at a time. now, training LoRAs is another matter entirely.
|
|
#
?
Feb 19, 2023 21:27
|
|
|
RPATDO_LAMD posted:I missed this post but using LoRAs on 4gb works just fine for me, using the latest auto1111 webui with --medvram and --xformers. as far as I can tell they don't significantly increase vram usage because even with multiple LoRAs active my dinky gpu can still generate a batch of 3 512x512 images at a time. Ok that's awesome, mixing LoRAs together looks incredibly fun. Somehow I missed out on installing xformers, is that hard? I vaguely remember my first attempt not working then never went back to it. I need to take some steps to get more organized with all of this, between Controlnet, LoRAs, and now this Block Merging project things can quickly get out of hand. The Block Merge project will take one 4gb model and turn it into 24 4gb models for instance. I asked in the windows megathread in SH/SC for Version Control suggestions, I'll ask here as well. Also what's a good way for Syncing backups to external drives?
|
|
#
?
Feb 19, 2023 21:43
|
|
|
Enabling xformers used to be a huge pain because you would have to compile it yourself unless you had the one specific gpu model it supported, but now there are a variety of prebuilt binaries and it can just automagically install one the first time you launch with the --xformers or --reinstall-xformers option so it's pretty easy.
|
|
#
?
Feb 19, 2023 21:48
|
|
|
RPATDO_LAMD posted:just automagically install one the first time you launch with the --xformers That did it alright, sweet! Now things run just a bit faster, thank you very much!
|
|
#
?
Feb 19, 2023 22:05
|
|
|
RPATDO_LAMD posted:Yeah but C&K was talking about using the language model to introspect and censor non-chatbot-involved search results. Which is not a feature of this technology. not at present but it's not very hard to think about warping it (the tech, not the specific models we have now) to that purpose, is it? It seems to me at some point these things just become fully integrated into what a search engine is, I don't think it's alarmist to suggest that will have detrimental effects on information availability and I am not the only smoothbrain thinking along these lines. This is why I was careful to say ML, not AI, and I was not trying to imply that GPTChat itself is suitable for the task. Someone who has access to the datasets and compute needed to make GPT, certainly could do this. https://www.scmp.com/news/china/science/article/3129414/chinese-researchers-say-theyve-developed-ai-text-censor-91-cent https://www.businessinsider.com/trump-truth-social-plans-ai-moderation-despite-free-speech-focus-2022-1 When models which are fully capable as search engines exist, we don't need separated, different search engines. Either this is what happens, or ML search remains a forever separate thing isolated from "normal" searches, or ML turns out to not work very well and goes away. Of these three things I find the first the most likely.
|
|
#
?
Feb 19, 2023 22:32
|
|
|
Disaster strikes as Google switches to using ChatGPT to provide all search results, but the only thing it responds with is a link to https://googlethatforyou.com/.
|
|
#
?
Feb 19, 2023 22:59
|
|
|
ControlNet is really good at making things look photorealistic.      Its Vampire the Masquerade Bloodlines. I like how Strauss just ended up looking like Colin from What We Do in the Shadows
|
|
#
?
Feb 19, 2023 23:14
|
|
|
IShallRiseAgain posted:I tried the gif frame thing   Not bad. Could be used to start with a sprite base and then clean up. Useful for either upscaling older sprites or could use a 3D model and convert to a 2D "hand drawn" sprite.
|
|
#
?
Feb 19, 2023 23:21
|
|
|
IShallRiseAgain posted:ControlNet is really good at making things look photorealistic. I recognized them! Really cool.
|
|
#
?
Feb 20, 2023 00:03
|
|
|
IShallRiseAgain posted:ControlNet is really good at making things look photorealistic. Like the thing I'm making is supposed to be a live action MIB/Rick & Morty kinda thing but the shots look more like some NCIS crime drama on Paramount Plus. 
|
|
#
?
Feb 20, 2023 01:23
|
|
|
|
| # ? May 31, 2024 14:52 |
|
|
Roman posted:I might have to switch to that. I made a bunch of (slightly less) photorealistic stuff in MJ but my problem is figuring out how to make it fit the actual vibe of the project more. That's the MidJourney style and the biggest limitation of it. If you have a vision MidJourney isn't good for it because it's too rigid in how it will always produce good stuff. You can't stray outside of that, especially with photorealism.  I ran your picture through ControlNet, canny, messed with the thresholds and prompt. I'm not exactly sure, visually, what you are looking for so I just said "live action rick and morty" while negative prompting anything to do with animation or cartoons. I asked for her to smirk a bit too. I know I have said it a bunch, but if you're trying to create something visually new you can't use MidJourney on its own because it looks like MidJourney. Stable Diffusion is a lot harder to get going but truly is limitless. EDIT Should mention that this is a straight output, I haven't touched it at all like I normally do.
|
|
#
?
Feb 20, 2023 02:05
|
|
