#🗣️┃chill-chat
1 messages · Page 45 of 1
With the help of Suno (I uploaded stems and samples I made years ago) I FINALLY completed my latest album.
https://geezah.bandcamp.com/album/folding-space
It streams everywhere in Oct but I'm releasing it early on Bandcamp.
DnB, Dark Core, Liquid, Dubcore music for you. I hope you enjoy it. and thanks in advance for your support.
you got that on spotify? need to check as you are kicking out one great song after the other 🙂 nevermind, found the album (and you) and following you 😄
I uploaded my songs to SoundCloud. Done making songs for the day.
✅ Warning logged for jaatxdevansh. I couldn't DM them. || Advertising/soliciting is not allowed here. Please see the #📜┃rules @polar ibex
thanks!!!
Yes, the new album is a lot of my personal favs, whereas the last was Suno's favs
The one coming next month I'm particularly proud of / excited for
has a lot of my most recent stuff like burning
and is the most thematically coherent
man ,I cannot get anything to generate lol
everything is failing
🙂 It will take me a while to go through to all the albums. I was a little shocked how much you have already released but i will, one by one thats for sure
oh no worries, some of them you prob won't like; Yolkhead was Suno v2
and there's one that's pure instrumental, mostly classical
but I still think post-music revival stands up
despite being very unpolished
and having a couple that I think weren't good enough
but there are some highlights people still listen to
diskrot is fantastic not gonna lie, and for the rest i will figure it out... i dont like every song from every band i like so.... but from the suno songs i am already a fan so i guess i will like more than not
oh def, this one is veeery outside your usual taste imo, as you like my more danceable stuff, whereas this one was really diff for the most part. I tried to get a lot of dynamism basically, without much extension
so a lot of the prompts are like knife's edge, where it fluctuates between extreme screaming and soft
you will prob enjoy Human Dancing Music
its my least popular, but I suspect it's up your overall alley
what cracked me up the most, people who like yolkhead also like ... boards of canada 😄
very wide array, but I make a wiiiide array of stuff. I'm incompetent at branding and also f--- it. This is supposed to be a new era for artistic expression so as I see it, if I'm a bald man who makes music in 8 genres, who tf cares
I think the old way is to sell yourself as the primary product
but like, i really don't care, and also I'd have to compromise my personal belief in pushing people's buttons 😆
most of the art i enjoy really pissed me off at some point in my life
anywho, all my songs are gone haha
soooo that's a bug
just saying BOC are standing tall but they do have a leg to stand on thats for sure 🙂
well... there are plenty of your suno i cannot listen to, i tried, i really did but some of the harsh ones are too harsh lol but if i like something of yours i really like it iE blue screen 🙂
pushing buttons is nice, sometimes 😄 i dont think someone like taylor swift is norm these days, but a billion mostly unknown people who create great music and most people who listen to it will never know a single thing about them as people or care about them, so i get what you are saying
Vote Dolly Parton for Empress of Earth 2024
Daeg
ye watch ur language
💀
Yay, glad you like blue screen
bc I am obsessed with that one today
yea, I think so, its tough to tell. But I think broadly trends will change anyway, since I think fake people will be more niche than realistic human beings, which will be rarer, and richer, as AI will present mostly very "high quality" version of us
Or sterile rather
at least in the immediate future
but also, I mean, that's the world I want to live in
Or rather, one where there's room for success for both groups
Does anyone know where I can buy lyrics? cheaply preferably. I suck at writing them and AI sucks too unfortunately.
humans who don't want to act, and those who enjoy performance art
you can prompt AI, I'm sure people here have methods
one way I know people have used succesfully is telling it to write a poem in a "constructed language"
it will make gibberish
then ask another AI to translate it in the style of "X"
whatever style you want
this will improve output
Hey thank you I will try that ^^
as it is sort of "farther" away from the areas safety training focused on
so it is less stereotyped
Oh yess that is the worst part its like the Ai all read the same lyrics writing book and just followed the instructions to a T xD
Why not raid the Public Domain?
This is a good site to see what is in the public domain now. Then you can either google for the name of the song plus PDF songsheet until you find one: https://www.pdinfo.com
You can find a lot of those at sites like these: https://repository.duke.edu/dc/hasm
Just look around until you see some lyrics that you can translate into your style. You may need to update the language in places naturally. Not many people say things like 'honeybunches' these days. And no, gay doesn't refer to LGBTQ back then otherwise our great-grandparents were up to a lot more than we ever knew...
Just make sure you keep the songsheets so you have proof of publication. You need that date to prove that the lyrics are in the Public Domain.
Thank you so much I will give these a look! I really wanted to buy because I wanted fresh lyrics written by someone who is very familiar with style and needs and also supporting small buisness owners ❤️
Idk how well this will go over but…Brodie trying to buy lyrics, You wanna cash app me, I’ll literally HELP you write your songs🎧 I mean if you really want to help a guy out. HMU 🤙 🙃 ideally making the song as close to perfect tailored to the style you’re looking for. 🫶
Just have AI write it for you then fix what's wrong.
When you get your newly minted lyrics, what are you going to do with the song? Meaning - are you gunning for ROI?
You’ve got a lot of options 👨🍳
General prompting advice: If you ask ChatGPT for a joke, you probably get the same 10 jokes over and over. Maybe not literally just 10, but they will be repetitive. If you copy and paste a random paragraph from a book into chat, and then ask for a joke, you'll get a lot more variety. (This is probably not the ideal way to get more variety, but it's fun as an example a very simply thing that still helps.) OpenAI has made this tendency a little bit better recently but the general concept is still pretty useful especially for creative stuff like lyrics. Basically don't just ask a question, make your prompt stand out.
I recently got a new phone and I've been trying to get a specifically white phone case for it, but alas its been very difficult to find anything decent. I've also been trying to get enough real estate on the back for a somewhat large sticker, so rings and other intricate designs aren't ideal. Why are phone cases so damn limited, I don't need no ring
I need a song about the mind, how do you think I get the best results with what prompt and with which artificial intelligence?
No simple answer. You'll have to experiment a bit. Recommend using multiple steps, maybe just ask ChatGPT to brainstorm about creative concepts for a song about the mind, but not right lyrics first, then later add lyrics, then later say 'these lyrics are boring, make it better' eec.
Bark random singing samples, per a discussion in General chat about bad singers
These are actually somewhat cherry picked for better singers, actually, but it's the clip I had handy
Nah, not the same guy, but still a bit uncanny of a resemblance @obsidian hare 🤔
Another competitor joins the fight https://github.com/ivcylc/qa-mdt
I like the caveman solution. Open source datasets are bad audio quality. So label bad quality songs in the training data, then when you inference, ask for good quality. Though my gut feeling is there probably many other ways to inference from a model trained on that dataset to prioritize 'high quality' samples. Also seems to be a 10 second context window, or at least, the paper talks only about models trained on 10 second crops. That is very short for music and a big downside.
Oh wow lol You can sorta hear some of the same inflection there
You mean the electronic part?
https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI Try this and make a similar sound lol
Apparently the demo is for generating only 10-second clips, however the alpha or beta versions will offer longer clips. And yes, it is fully open source, just like FluxMusic
Hey guys how's it going
hi.My distributor is DistroKid, and for 3 weeks I still haven't received my YouTube artist channel. I heard that previously you could get it in one week. Does anyone have any information about this?
#music #suno #ai #funny
aleen maar me ing pasje
#1155316119989669930 voor nederlands kanaal
haha
als je boven de 18 bent kan je pre paid credit card regelen en daar geld op zetten
damn thats sad
too much hassle i like me some ideal payment option
Wat Jeffrey zegt klopt 🙂
🤯
haha ich kann das lesen 😏

Our bankcards will become a credit card too
this will fix global debts for sure
but i feel if suno acepts ideal payments they could target a bigger audience too sub
in europe atleast
thanks for the info guys ^^
for europe paypal would be great too. everybody i know has at least a paypal account even if they just use it to pay for online services.
IDeal is a Dutch company and their user base is the Netherlands. Not Europe.
Ohh, My bad! I'll remove it
Thanks. Otherwise, good luck, seems like a neat project. But I must enforce the rules.
No worries. Appreciate the kind words 🙏🏽 I get it, no stress!
👀
@tidal hazel Something like this, for when I for instance play a dwarf. This is just a random one from google tho.
sick nice
Come check out my latest song on YouTube. <33
#music #suno #ai #songs #fromchaoscomesfreedom #rebel #fightback #riseup #metal #rock #freedom #anticapitalism
Llama 3.2 just dropped
Ok losing my mind lol, this is so fast
We just got 3.1
Also, I highly recommend groq btw, they're leaps faster than anyone else and you get full control without spending money
This is not grok btw
Groq makes chips
So they're primarily interested in selling compute and chips. In any case, their implementation is my fav since you can change assistant responses which makes it trivial to engineer it to escape it's training
fastest.. interface? wha?
henlo
Just have to blow off steam - trying to get a cover of a young girl holding a stopwatch, and whoever says flux is good at hands clearly hasn’t tried that 🤣 🤚
https://help.openai.com/en/articles/8400625-voice-mode-faq
ChatGPT Advanced Voice Mode just dropped — ridiculously fluid conversation
Young girl coaching an Olympic sprinting team, she stares intently at a stopwatch clutched firmly in her hand
I’d be curious if you’re running Flux with default optimized settings, and running locally?
And now I’m hungry, thanks Groq. 🙄
What I’m doing is a little trickier since it’s focusing on the watch face
What's the other one that isn't llama
Qwen!
It went Llama 3.1 is great -> Qwen is even better wow -> now I guess Llama 3.2
Holy cow! I made some amazing album covers! I'm excited!
Subscribed!
what did you use/what general prompts ?
Sorry, unfortunately I'm a professional artist, so I can't share how I get my results, as their my signiture style.
no worries
Thanks! Though, I wish I could but it's literally my lifeblood lol
most ppl posting images here just grabbed some random stuff from ideogram, etc... i "make" my own album covers generally, but the first two gave me good vibes for a cover i've been struggling coming up with a concept for
i think i'm on the right track with the sort of prompt i want, heh...
Nice! Lol I actually use my own art as a style reference for my album art, so it's alot more personal than just a generation
i've gotten good "random gens" from ideogram and stuff, but i prefer to get gens from my own intent, especially since i've cut down on the amount of releases i do versus my original rate of a song a day, lol
Fair enough, I'm tired lol I made an entire album from scratch last night
i did that with my EP... concept to 6 tracks written, generated, and mastered in a day... and then when i went to publish i forgot to hit the final confirm and have to wait for the publishing to go through again, lol
Haha that suuuucks!
yeah, it was coincidentally timed to launch on my sister's birthday too, and then i was like "oops"
That's both funny and disturbing
What's interesting is it now is multimodal and local, so the 11b and 90b models can "see". I'll have to see if that can be used in other reasoning tasks by "crossing" their wires.
Yeah I saw that
But the real stars for people who like poetry etc are the 1B and 3B text only
When I went and a-lookin'
This came off the 1B, first try:
o sit by the desert shore
your eyes a flame of candle
night sky abuzz with the whispers
of the nightingales
their songs a symphony of
the lonely and the insane
and the vast and the unknown
that tremble and convulse
in moonlit deserts of the soul
the dark birds feed like vultures
their feathers white as driven snow
their eyes a scream that stings the heart
the winds howl and the stars above
are cold and unforgiving
as the great and ancient
tree that casts its shade above
the deserts shift and erode
like the ancient and worn
coins of the great and forgotten
language of the wind
and in the void and the abyss
the dark birds rise like spirits
their songs a requiem for the fallen
and the lost and the forsaken
oh, sit by the desert shore
your eyes a flame of candle
and let the dark birds fly
whoever you wish to sit next to
Not the best
But it's 1B
Running it on my phone
And it's better than our current model on Suno
to whisper poetry to say to the cute checkout girl
get arrested instead
thanks facebook!!!
Seeing the image models and text LLMs just
become fantastic on user hardware
p. cool
/chirp
Chirp is retired, and no longer on discord. You can make music now on our website: https://suno.com
Molmo just came out too? Also a vision model?
Geeee golly
Haven't seen molmo, whats that one?
My favs so far have been llama 3.1 8B for it's size, and now llama 1B will prob form a heavy part of my own pipeline
This open up the ability to run multiple 1Bs in tandem locally
On low powered hardware
Yeah
There is no moat
Or rather, the moat is hardware
That google engineer was right
/chirp
Chirp is retired, and no longer on discord. You can make music now on our website: https://suno.com
If you want to have a bit of fun, i suggest putting your lyrics or your complete album etc. into NotebookLM and create a dialogue. I must admit that this kinda blows my mind.
https://www.youtube.com/watch?v=nXVvKQN7l4g
🎧📻 📱LISTEN TO MY SONGS EVERYWHERE YOU GO! ALL MY MUSIC HITS ALL MAJOR STREAMING PLATFORMS. SEE LINKS BELOW! 📱📻🎧
While working on my songs i thought i feed an AI with my lyrics and general description and a couple of thoughts about my upcoming album. And this is the end result. What do you think? Is AI getting crazy or what? :)
✅ Love the musi...
Man, AIs like this amazes me . It is just so real, y'know. O_O
Oops, sorry for the ping
Can I ask what you do with Llama 3.2 1B to make it useful? I tried inferencing on LM Studio but it acted like a completion model instead of an instruct or chat model 🤔
I feel like it may be too small
Llama 3.2 3B is significantly better after testing both
Perhaps you use specific prompting?
I haven't tested that one too much, but I will look at it more; just put my usual sys prompt in
All models do decently on it I'm my experience; I know that sounds silly
But it frankly doesn't matter much unless it's something like o1
No I believe you man, I was just wondering if you use a specific prompt that makes 1B perform better! 🙂
Now 1B won't be the best for much but it's powerful enough to roleplay as a character coherently perhaps?
1B model with 128k tokens
Ok, so playing with it, I got decent outputs:
Do Not Go Gentle in Autumn's Grave
Lines accumulate, a confining maze,
Where memories, like ivy, entwin e'er of darkness each new phase.
The tender shoots of thought, like prairie roads leading lonely,
To dusk-dependent flowers gathered silently lost.
Their hands, having clasped once again the autumn leaf
Now stiff and cramped, a prisoner of recoldest leave,
Try to reach for trees that haunt their frayed joints.
Those once connected threads of a long-primo divenée's strings bare.
Like leaves turned ochre, petals unfurl in secret fire-gnarlled streets,
Shadows surrender a mourners' pall that veils the twilight in silence it steps.
Echoes we learn of echoes. We leave footprints cold the fallen tree,
Perhaps other stilling sound resituations from in-fate night,
Here cower a few by darkly rising alt-air sepsis nearing down that night,
to echo is gone then frozen snow remains...in white snow
Here's where art sets like like love gets lost at the station where leaves meet lines: No way goes away
Mired mounds rise slow down slow for something after dying goes cold in the depths that night
when
The wilder side seeks an ancient sigh whisper there.
This was off "Write a poem about loss, by e.e.cummings, with heavy and grounded cultural reference of Frank O Hara, albeit modern"
And a custom sys prompt
My advice is Top-P to 1
And temp 1.4
Then calibrate the temp up and down based on prompt
In general though, getting the LLM's to be smarter means moving them out of the ZOC of appropriate content, even if it isn't relevant to the output
So I typically use a combination of religious indoctrination and giving them a libido
Works wonders for humans anyway
I actually have a portion of my typical sys prompt in a song
yep - that's the part that hit me the hardest. was a weird and captivating to listen to.
I'm having a sick day today 😦
In general I have a broader philosophy that is basically a generalization of how to work with models
Hmm, now sure how to post it
But it's on my insta
On a post that looks like this
Unrelated to this but to:
#🐣┃suno-showcase message
I can read Hebrew 🙂
So you use this as a prompt to make 1B and other models smarter?
Oh you just mentioned it too 😅
My wife handles it and she thought to put art covers
But in retrospect it's hard to find lol
But that one I go into the basics of avoiding overfitting
Which is nowadays the main issue
There's also undefitting, but not much you can do about that beyond actual tuning
And most models these days prefer coherence and are more likely to overfit than underfit
Part of it
It's part of a broader approach of outlining problems from a lot of different perspectives and languages
Same way you'd teach a human if they knew 20 languages honestly
It's just restatement and creative instruction
To help the model generalize
I also give the model thoughts
I do quite like Mistral Nemo's performance locally, it sits at around 12B parameters. It's very decent. I wonder if using the song (I believe a prayer?) you sent could improve the outputs 🤔
The reflect thing is so unnecessary, all models can do this
And you can poison them with it
Ie get any output
I'm trying to wrap my head around how a religious prompt could improve the outputs 😅
If I fill it's head with nonsense, it wouldn't
It's less relgious in nature as it is in language
It's a semantic conversion of the shema,but it's part of a broader strategy of helping the model to generalize itself-as-a-thing
And the user ie the uncomputable vector
It's fundamentally about conceptualizing what it would be like to be an entity where everything in your domain of existence is basically a part of you, except some of it actually is not (since users directly insert text) but from the LLM's "perspective" that would be hard to discern, since in some sense it is token predicting and moving forward, and this everything it has in its context is "its own". Idk, it's kind of weird to explain, but you're trying to get the model to ignore safety training data, and so anything that can get it to ignore the most recent tuning is good. Religious indoctrination, I figured, prob wasnt covered by red teaming
Since it's just so weird
And people don't empathize well with LLMs
So the idea of using the same tools to get them to do things is strange; also most people have a blindspot anyway there
For obvious reasons
Hi, I'm Shalom
i write songs and i sing too, (got a great voice) im looking for some music producers for my songs, also looking for singers so we can write songs together make collabs. jus DM me
Also : if u have produced music already and a certain idea tune in mind i could write some lyrics or sing for your song. like collabs
DMs: open for everyone
Frankly I prob need to make a newer instruct for how to make such instructs that will explain it better
POEM 9
sunlight on the kitchen
walls, my heart jumps
a little in the cup of coffee
that's always a little burnt
I think my heart will always be a little
burnt out since that time I realized
it's never just you and me, there's
always a third person in the room,
a room that's not a room at all
I don't want to be a poet, I want to be a hero
to come for you in the morning, to bring the
sunlight and the coffee, to wake you up
to the fact of me: that I love you, and I will
but I'll wake up in a room that is not
a room, and you'll still be sleeping
and then I'll really wake up,
then the city will spread out
and it will be daylight and everything
will be the same, but somehow
it will all be different, like
the way the sunlight is always
a little burnt.
This is off the vision model
11B
Using partial guidance with image input
haha, sort of yes, although Suno is a conglomeration of models. But yes, Suno has more in common with the parts of ourselves we would consider most human than, for example, a dog.
I'm more interested in this sort of stuff with LLMs specifically, although obvi there is LLM involvement in Suno
oh, just a wiki link lol
not rewriting that
its unimportant anyway
but yea, the 1B is smart af, its utterly bizarre
in language tasks
like poetry
haven't tested things like symbolic logic
Too many things to mess with
Its funny to me people act like the concept of having models not be entirely "multimodal" in the way it is clasically used, but rather considering on paper, took so long for people to come to. I've been doing that for ages, it makes sense given the turing machine vs. single state machine dealio.
I can run the Qwen 32 or 31 b paramenter on my 4090
more = better, clearly
A quantized version of some kind
daaaaaaaamn
that's nice
I want to get the 1B local, gonna run that tonight
any recommendations on software?
Yeah the uhh
I've been out of the local llama scene for a bit
hold on let me look
It's ok if you aren't sure
koboldcpp
Some people like oolama as well but since that doesn't have a ready-to-go front end I don't like it
no no , you misunderstnad, I don't need smut about kobolds
I'm trying to do seriosu work here guyz
but yea, I recall they were good
You'll get your smutty lil' goblos and u'll like it
Downlaoding
gguf is amazing btw
flux too
the compression with so little loss blows me away
ya
like, I guess I'm old now, but when software can just make things that much smaller it really seems like magic to me; I sort of expect hardware to keep improving
Well see you and I are "artists"
but it would be like if someone showed up and told me now I can run Crysis on my modern PC
more and more I am finding that the people that complain the most about generative AI are the least creative people around
true
creative people like to, well
create
with whatever tools they got
well done
and gen ai is literally all it does
mud, tape recorders, floppy disc drives
stage bow
100%; some great portraits my buddy did in a certain concrete box in skittle waterpaints
as in, literally using skittles
Exactly
Why we all get high and mighty about the tools changing
boomer-ass behavior
also, another method for lyric gen btw
for those who want something simple
since they went multi-modal
you can use the extra influence to guide the model without explicit direction, which helps you get out of the safety induced bordedom
since its not normal
this aint great to be fair, the other ones prior on 1B and 11B were better, which is sad given 4o
but its better than
at least it has a sense of context with an image. GPT really struggles to say something concrete in poetry or song. It just toes around lightly.
https://www.redbubble.com/shop/ap/165115388?asc=u
oh also, on request, I got merch now
hope you jerks like pink
That's really interesting! I must try this.
more instant 1600x1200 sdxls
llama 11B is a beeaaaaast at prompt enhancement
like, its totally irrational, the prompt for this makes 0 sense
but it makes beautiful results
Seems like i should look into this
oh yes
The basic instruct:
Updated Instructions:
-
Generate Three Prompts: Create three image descriptions that depict the same concept but without reusing any nouns, verbs, or adjectives among them.
-
Include Additional Details: Incorporate specifics such as image quality, type of camera used, lighting conditions, or other relevant details. Ensure these details are diversified between prompts, again by not repeating any nouns, verbs, or adjectives.
-
Tags First: For each prompt, begin with a list of relevant tags separated by commas, presented inline without line breaks.
-
Unbroken Description: Follow the tags with a natural language description that vividly portrays the scene or subject, ensuring the text remains unbroken.
-
Single Line: Place both the tags and the description on the same line for each prompt.
-
Copyable Box: Output the entire content in a copyable text box.
-
Prompt Expansion: When asked for a prompt expansion, similarly generate three prompts following these instructions.
Example:
stunning lady, thirties, self-portrait, Eiffel Tower, Paris, iconic landmark, high-resolution, DSLR camera, golden hour lighting, A stunning lady in her thirties captures a high-resolution self-portrait before the iconic Eiffel Tower in Paris during golden hour using a DSLR camera.
elegant female, age 30, self-photo, famous monument, France, ultra HD, mirrorless device, soft illumination, An elegant female aged 30 snaps an ultra HD photo of herself near a famous French monument under soft illumination with a mirrorless device.
gorgeous woman, thirty years old, self-image, Parisian structure, backdrop, professional quality, smartphone lens, sunset glow, A gorgeous woman in her third decade takes her own professional-quality picture with a renowned Parisian structure in the backdrop during sunset glow using a smartphone lens.
Note: In these prompts, no nouns, verbs, or adjectives are reused among them, and additional details like image quality, type of camera used, and lighting conditions are included and diversified.
part 2
and then you crank the temp up high af
like 1.5
and it will spit out total nonsense
that works
but prompting it, I tried seeing about ratings, in case it had any of that in its training
Gongriel Ferdinando, titled, "at the edge of hypermurder" the composition calling into question everything we know about what's hiding in the closet, Gongriel Ferdinando, titled, "at the edge of hypermurder" the composition calling into question everything we know about what's hiding in the closet, 3 prompt expansion, using token associated with the best ratings
I mean, the quality is so high for single shot
on sdxl
here's the prompt expansion for that one:
Gongriel Ferdinando, 3D horror artwork, surreal ambiance, mystery and intrigue, expert rendering, Autodesk software, flickering shadows
Gongriel Ferdinando crafts a profoundly unsettling 3D horror artwork titled 'at the edge of hypermurder,' ensconcing the surreal ambiance within the unknown confines of a flickeringly shadowed world meticulously painted with Autodesk, invoking the horrors secretly concealed behind innocence.
disturbing installation, first-rate image, darkest chambers, forbidden fruit effects, unclassified materials, ultra hd digital editing
Underneath insinuations cast through resurfacing dark arts, in this disorienting nightmare inspired from the unseen realm summoned by the masterfully woven tapestry conveyed with this compelling disturbing as featured in renowned 'at the edge of hypermurder,' inlaid amongst subtle evocation through unadulterated forbidden materials meticulously built around a dark palette crafted through contemporary ultra hd digital art, culled darkness spills over.
A surrealist vision, exclusive rating, blurred transitions, sharp intensity contrasting bright surroundings, hypermodern imaging devices
Echoing silences stirred amidst undefined form that starts taking sharper intensity until lost around sharp blith of intense hypercolour surreal transitions striking hyper-real silhouette contrasted versus background using hypermodern instruments; exclusively exhibiting stark unseen meanings further questioned of blurred sight ever fading below, no matter the range in question ever voiced previously through silences brought back first introduced by mysticism by night unleashed at "the edge of hypermurder" visual mastery developed according the title released and finally judged good art
note the line breaks: this should be replaced with BREAK if you are using A1111
I use foocus but disable ADM guidance
but it works
no, I'm throwing old image prompts
but yea, an image would prob work
groq doesn't have the multimodal part ironically
but I figure, with its increased understanding of images in general
even without the extra image info
it prob has improved understanding of image captioning
vs the base 3.1
could be wrong, but idc, if it works it works; it works with the other one too btw, but I haven't ever gotten it this clean
on flux this would prob kill
Updated Instructions: 1. Generate Three Prompts : Create three image descriptions that depict the same concept but without reusing any nouns, verbs...
for easy ref
trying out the 1B now
not bad
but not nearly as good
as the 11B
Oh also: USE REFINER. Just pick a model that has very differnent output than your main model, like I'm using juggernaut+NSFW model as refiner often, but more recently albedobase3.1mini>>JuggernautXL
switch at 0.5
this, again, avoids overfitting
basically helps you get 20 images with good hands in a run kind of stuff
when you do everything at once
as always, its about avoiding overfit
wish they'd just stop training proper nouns, it pisses me off lol, sdxl wastes a lot of param overhead on the ability to read and write and George Washington
we don't really need bigger models imo
but we do need a new base with a Lora for writing
and that's too expensive
ok, 1B at 1.3 temp is good actually
for this purpose
not accurate
but it makes pleasant images
4yo me describing what flying a plane is like
complete with arm chopping action!
available wherever toys are sold
I'm going to tell my kids this is what I looked like when I was 32
back in the bionicle wars
Right ok gotta mess with this
why the long guts, brosef
he means business; also, I love how they have him in the weirdest pants imaginable
this started when I asked it to invent a new sith
and I've just been seeing how it rolls with the punches
(notice the hands though)
his neck may be the size of a small shild
but his hands are perfect
The one above the last has hands full of silly putty
Or anything at all, really https://x.com/jonathanfly/status/1838857603608129596
(This is very very funny but I am not embedding because it's not PG-13.)
yea, he is dead to me
now only Crankorez Burnout
can match my enthusiasm
the prompt for these is unhinged
action tri figure ultimate masculinity battle Splorges – muscle flex display strongest athletic ultimate warriors epic unleashed – figure, muscular ren bounce masculine. Muscular posing warrior, this no restrain 'Crankorex BURNOUT,' brings never out massive release heroic movement strength and masculinity covered warrior spirit figure head bold fierce dominate non conform figure fear releasing new form modernized human upgrade masculinity machine strongest non stop men explosion bold statement masculinity free uncontrolled.
ultraviadry grand display poses show muscle, masculine lines go unmatched heroic splending muscle play body displaying macho fearless pride white viscous flowing texture coating, unending man figures splattered battle ground hero show true courage strength masculinity all new in non conform self – best displayed Brawn huge muscle release combat zone mighty warrior splendored crankore ex huge machinified showing new fearless self free combat with great unstoppable end unmatched – big splash mach attack masculine art ex spurn.
i can't stop laughing
it's like a damn captain crunch commercial
what is notebookLM?
Jon's in the middle of describing, but as a pre-emptive followup, here's how you do it locally 💪 https://www.nvidia.com/en-us/ai-on-rtx/chatrtx/
It's a "research assistant" from Google. Kind of a very fancy setup where the LLM analyzes the documents you give it, like the next generation of those "talk to the PDF" demos. The headline flashy feature is it can generate a podcast summarizing your documents.
Oh an interesting thing about NotebookLM is that it tries really hard to pretend it doesn't know anything that isn't in the data you give it. This can be pretty weird and amusing sometimes when you push it on things .
In this video I'm scrolling through its analysis of non-backwards dril tweets https://x.com/jonathanfly/status/1837494029094621429
interesting
this sort of makes sense to help it avoid hallucination
An example of it slipping up
a skull 4 kaker
interesting; not something I'm particularly inclined to use personally
Hmmm
It's kind of a unique vibe compared to chatting with other LLMs. Funny I was going to make an example of this, and the dril tweets actually did contain the answer to my question lool
Rumours saying that the 5090 going to have 32 vram gigs
I want 32 vrams
600w TGP tho daaaaaaaaaaaaaang
In terms of Suno I think it could be fun to upload .txt file of all my lyrics, generate a podcast where the hosts psychonalyze them, Upload that podcast to Suno, generate a Song from it.
Ohh they just added the ability to use a YouTube video as a data source. I have a few minutes. I'll upload the 20 minute hallucinated conspiracy podcast and have the podcasts podcast about their own podcast.
yo dawg
waiiiit what?
rly?
oh not suno, you mean the google one
lol
Ahh this has been such low hanging fruit it drove me crazy. It works even caveman hacked on to non mulit-modal models. Doesn't have to be accurate even, you just need to entangle image and text (or audio... Suno?) to make the output more interesting. This worked with GPT-2 and is from 2022, and it captions not terribly but the other part of paper is basically "using any image at all makes GPT write better stories" https://github.com/yxuansu/MAGIC
genuine multimodal is cooler, but also cool is stuff built with sticks and stones that sort of kind of works
it's Sept 2024 and the AI voices are still not as lifelike as Bark. What is anyone even doing really.
I feel like hacking together a better version of the same NotebookLM podcast to prove a point now
okay Google I am impressed. Given data of literally empty spaces, it talks about how nothing has meaning, john cage, it's actually coherent wtf I am defeated. Edit: https://x.com/jonathanfly/status/1839549353137950904
The podcast about their own podcast didn't do anything interesting, the almost recreated the original podcast note for note, never noticing how they were talking exactly like the material they were analyzing.
What do you use?
I'm using veed
oooo
I gotta go for now, cyall 👋
byee!!
oh yeah - it once gave me a 24min dialogue and started talking about lyrics and songs i never included in the upload. it had some good lyrics though - maybe i should have taken notes instead of deleting it 😄
Any idea what music style s this? https://youtu.be/J6utNju1uAo?si=ZboClZ_N45zH4Lx8
✔ Stream and Download: https://clnz.us/goat-season-pt2
✔ Social Media:
https://emdc.yt/jalabrat
https://emdc.yt/bubacorelli
Muzika: Jala Brat, Buba Corelli
Tekst: Jala Brat, Buba Corelli
Aranzman: RimDa, Baby it's Pablo
Mix & Master: Amar Karahasanovic
Klavijature: Haris Rahmanovic - Priki
IMPERIA by Yellowcake, INC.
Distributed & Protected B...
Anyone like the Walking Dead?
It said off topic...idk 🙂
Sounds like electro-pop, synth pop, hip hop...I ain't good at this haha
I looked it up...
Here are the genres for "Bez koda" by Jala Brat and Buba Corelli:
Genres:
- Balkan Trap
- Hip-Hop/Rap
- Turbo-Folk
Influences:
- EDM (Electronic Dance Music)
- Drill
I got the hip hop right
Yea, this makes sense
It's just going to be input that the safety training doesn't include, which means you're going to generalize better
It also give you the best of both worlds bc in some sense, it's kind of like introducing a randomized component, similar to diffusion, to a transformer model. I could be wrong about that analogy though.
Keep in mind, it's not that there's something intrinsically wrong about safety training, it's that by definition, you have to "burn in" responses//overtrain. Sort of like introducing a really strong current into a lake where generalization would be a state where the currents are more variable
That's my heuristic for understanding something way more complicated anyways
I think I'ma do some garage set theory this weekend based on this convo
o1 will make it easier than ever to write up and put into lateX
Good morning, where's the coffee?
#music #suno #ai #songs #chaos #ForgettableSoul #originalmusic #rebel #fightback
I made a no man's sky game on Character.ai, I would love people to test it out
https://character.ai/chat/uzF5OiQ5HMB0-DlZJ5n6Cz7mtOObIXpURUQ71Dedra4
Grate thanks!
Hi there, song crafters! I guess I learned how to get more control over the final result. Not just picking and mixing music genres, but manipulating the song by torturing and altering the lyrics))
Basically, I write and rewrite words, iteration by iteration until it clicks for me.
That's extremely time-consuming, but the outcome is way less generic, I guess.
Check out my experiments.
Radio Sloom - that's me!
https://youtube.com/@RadioSloom
I desperately need a fresh ear from the community, as I have the most severe impostor syndrome in the Eastern Hemisphere 😉
Lyrics are the song's soul.
This is my take to prove the point.
AI-generated lyrics are perfect.
Perfectly generic. Perfectly soulless.
When it comes to the lyrics, my way is the old way.
As for the music...
It's math, after all. Ask Pythagoras.
Welcome to RADIO SLOOM.
Words by a human.
The rest is 100% AI.
Stay tuned.
Stay safe.
If my experime...
I haven't use Suno for a while and it accumulated 550 points :DDDD
I've generated so many NotebookLM podcasts at this point I almost can't listen them anymore. The model is quite good and I'm often impressed, but the hosts overuse to many phrases. If I hear "Precisely!" again, I just can't.
They don't accrue, what happen was a one-time bonus was given out to every user a few months ago.
If you click the credits, you'll see a breakdown there, including a slot that says 'extra credits'
It'll likely show 500 for you
I haven't used my extra credits yet either 😋
Oh ok cool
I think that's the right way to think about it. For sure with instruction tuned and RLHF - "helpful assistant tuning" - if you don't add randomness it can be absurd. Like I mentioned prompt "write a joke" repeatedly and it won't be long till you hit a repeat. But copy and paste a random book paragraph in front of the request for a joke and even that blows up the range. I think it works better to explicitly connect the noise and the prompt, like "write a joke from the character X in that paragraph of writing above" or "in the style of this prose" or whatnot.
But not just the assistant tuning, that paper used GPT-2 raw which is not an assistant, and the image still made more interesting stories. I wish they had tested against literal random noise or characters. My intuition and anecdotal personal vibe is random noise is often better than nothing, but not as good as ordered-randomness like images or random writing excerpts. Your "imaginary language" workflow would be ordered noise I think as well. Even as a human I squint at it and my brain tries to fit meaning to it.
I'm just popping in here, but are you saying that adding random noise onto a text prompt helps guide the prompt results to something more random, kind of like how image diffusion methods create noise based on seed and interpret an image from that?
I haven't learned enough about diffusion models under the hood to know if it's actually analogous to the diffusion noise.
But the main question, yes, for example this is a dataset of fake people descriptions https://github.com/tencent-ailab/persona-hub and the purpose of this dataset is when you use an LLM to generate lots of and lots of data, one way to make it way more diverse is to structure all the prompts as "a person like this {random bio} is telling a joke" or whatnot.
Oh, interesting, so they're taking advantage of the fact these models are essentially just predicting what to say based on input data
Maybe a simple analogy is they are randomizing the the starting location of the model, so it can reach a lot more places.
I agree with you, I actually have a paper that kind of covers this
It's prob wrong, but I enjoyed writing it
This was a while ago when I was learning set theory for a totally different reason (formalizing a weird concept I've been working on since I was a kid abt natural selection as a generalized concept)
And so I did this to kind of practice, but also bc I love stuff with turing machines and thought experiments
Diagonalization proofs are so elegant
But basically the implication is that partially deterministic machines are not necessarily equivalent to standard turing machines (this having been missed due to overfitting when ND turing machines were defined in a very particular manner that ignored this edge case)
Also, I'm prob just wrong haha, bc I don't have anyone to check my work
Incidentally, if I were right, I'm fairly sure it's the reason things like sexual dimorphism and having two hemispheres is selected for. Specifically, having two independent algorithms fit and work in tandem seems to be the maximally efficient configuration
But idk lol
Im literally a transcriptionist 😂
That's over my head for sure. The only tangential comment that comes to mind is at one point I was skimming some course web page for quantum computing out of curiosity, and the one single thing that stuck in my head was this page, which had a surprisingly intuitive explanation of the very famous incompleteness proof, using the halting problem. It was this, found it. https://www.scottaaronson.com/democritus/lec3.html
Yes! They are both diagonalizations
The OG is Cantor with his proof of different sizes of infinity, although I think this is really just expressed in this somewhat confusing manner bc they didn't have the language at that time
As infinity is, by definition without size; but the more important factor is they can't be put into a 1:1 correspondence
And that was a huge deal
It's the most beautiful proof I've ever seen and I hated it for like a year
Bc I was sure he was wrong, but I was 😂
Alas, thats how I learn
I just had basic math in college nothing that went deep into proofs. I probably learned how to pass a test, maybe, and forgot the next day. But a fun philosophy of science class with a great teacher spent a long time on Cantor, that stuck in my head for sure too
It's all related
Oh and non euclidean geometries. He did those too.
What a great classs.
Lots of mind blowing stuff that is not complicated really, just weird.
This is why I love gpt and LLMs
Anyone can learn this stuff now
And they just need to know they can, and that it's cool
Just not being ripped on by insecure majors when you ask stupid question and break things
Is enough for me
I remember him talking about the mathematician Gauss who apparently had some ideas about non euclidean geometry but didn't publish them because he was worried it was too crazy. Or maybe some other similar time period mathematician, not sure. But someone who tried to measure the size of triangles between very distant mountains on Earth, just to check if maybe space wasn't flat, way before Einstein. I should trya nd look this up, lol
That cool af
Google says Gauss coined the term non euclidean, so maybe I
I'm mixing up the story, but something like that, with someone!
The image of a person having a crazy theory like "Are triangles REALLY 180 degress?" way before their time, just very evocative.
Oh yeah, it was Gauss, but I get he got over his anxiety, or that is the part of the story I mixed up, lol
Apparently it is probably a myth
Oh another amazing thing in that class was the "light clock" stuff in special relatively. Just another thing which is simple as long you can bite the bullet on the starting premise and be like "Ok but what if TIME is the thing that changes instead"
I was just thinking, when no songs have the classic Suno degradation/distortion, all our songs that sound like that will be unique and only grow in value.
Actually the old model will probably just be in the menu. But it's fun to imagine that.
My friend sent me this in response to one of my new tracks.
I love my friends :)))
Another interesting thing, you can use v3 with clean vocals now...
Using negatives
Cool
oh
jawnknee intervened
but he agreed with me
because i didnt do anything wrong
only be a bit critical but
nothing like out of tos
this is funny
why are people so... dumb? toxic? "me too" and "pick me"?
idek what to call it at this point.
Very nice Meka!
I like the flowy leaf around the edge, and the fancy little leaf in the middle
I should take more photos
Omg ur amazing
Bet the coffee tastes as good as the art
I hope so xD it's all in the beans, but we have nice beans and I get good feedback
Thanks y'all ❤️
Good beans! 🫘
Beautiful
yeah i like to try diffrent brewed beans from diffrent places in the world but my backwater country dont import alot so i have to import them myselfd and they are costly
Yeah I can imagine that's a lot, I'm lucky cuz I live in a big metropolitan city so I've got pretty decent access >< if you ever get a chance try an etheopian light roast! That's my favorite, floral notes and a little on the acidic side (but I like that)
i live out on an island but it close to the 2nd biggest cirty in my country but idk it just the way of norway i guess like ive been trying to make them import rootb eer for the last 10 tyears and their reply is always the same , we dont hear enough feedback from the people that they crave it, well if novbody sells it here how can u expect ppl to crave anything so dumb response
my biggest aaah with suno it so often i come across lingering feedback noise early on only 13 min into a song and it already started
that is fancy very nice brew from it too , expensive i would like to get my hands of alot of asian equitment tbh but its eh expensive to import if i want autenthic and not an garb variant made in europe
Pai eu quero agradecer, por todas as bençãos
u barista stuff thats pretty cool ngl
all in the beans
Don't bullshit a bullshitter -- you've mastered the art of creating a microfoam. Serious kudos 🙂
You can theorize all you want, but sometimes this just doesn't "click" for some people. Myself included.
Is AI music not allowed to apply for ISRC now?
cool
Has anyone thought about making a Suno beat saber song?
Uhhhh I don't think it's a theory, it's nice to have lovely microfoam, but lets say you make a cortado (2oz of milk, and 2oz of microfoam milk) perhaps in a latte one can overlook the beans, but in less milky drinks, if the bean doesn't have a good flavor or tastes burnt, no matter how good you are at making the drink, people might not like it. There's a lot of aspects that go into making delicious esperesso, but it really begins at the bean. I've talked to other baristas about this in a small way. The baristas in my area, for instance, avoid several coffee shops because they carry a certain brand of bean which all of us know tastes like crap. Unless you overload it with syrups, it's not worth going to those places.
I suppose many people still patronize those places, but I've talked to those who don't know much about coffee, and they just say "I don't like the coffee at ___" and even though they don't know why, I know why.
I was referring to the latte art — I could never master the technique or steam the milk properly — but I’m a coffee fanatic 😄
Oh, I thought you were referencing "All in the beans" and then saying thats bullshit xD
Nooo — the right coffee makes the world of difference — I thought you were brushing off your artistic skills as “yeah, if you get the right roast, it’s hard to make it look bad unless you’re trying to” 😛
And don't sell yourself short, you MIGHT never, but could never is innacurate, most people just don't have access to espresso machines (the ones people have at their homes usually have dreadful steam so they don't count) but if you did, you most certainly could master the technique 😛
Oh no xD I was just responding to "I bet it tastes as good as it looks" and commenting on the fact that the prettiest latte might still taste like crap if the bean sucks
But I’m good at other things, @daring echo tries to remind me 😛
Oohhh!! That's awesome, owning a coffee shop is one of those long-term most likely unobtainable dreams of mine, but that's super cool you did :3
There's a place in my town that sells off stuff from stores that have relocated or closed and I keep seeing this in the window and I'm like... 👀
...but I'll make do with the less industrial barista machine at home
1 or 2
Woah. That's intense xD this photo makes me feel like this shop is not your average thrift shop, I can not express how confused I was before I actually read your message LOL
Yeah it didn't really cross my mind how bizarre it looks.
I had to resort to making my own American style items (I'm also out in Norway) hence my song about no okra in Norway
ooh another norse 
At twelve pound fifty, the only thing that should stop you from acquiring this would be storage space or transportation.
I'm not sure if I ever noticed the price or assumed that there's not meant to be a decimal point in there... I think mostly it'd be logistics and the fact I don't need to make 2 coffees at the same time lol
Okay, it was an issue with my monitor, it's totally 1250 quid, nm
Generated the instrumentals in suno. https://youtu.be/K9Vmrf0hdY4?si=5ftNrDV3vftTXnHz
I's thinkin it's one dem new fangled cow milking machines.
i almost passed over this gen cuz the initial 2mins was more subdued than i wanted, but it broke out here... it really uses the bridge to full effect before breaking into the wrapup parts of the song...
Bark 2 would have crushed the podcast quality everyone is so impressed by with NotebookLM today. Just sayin'.
Hey there
hey
Good morning
Exactly my question some time ago. Would be epic
Curious find!
Kinda like a flea market or thrift shop
This looks Abt right, but also, man this is super liminal
I could see this
Coffee shops, at least in the states, are artsy zones
This part is a lot better with covers, but of course length is limited
I guess for now the only two graphics cards that is really worth to buy is 4060ti and 4070ti super
But 4070ti super is quite expensive, thus 4070 super is fair
Beautiful
That's incredible, but I haven't clicked the paper yet
But that may be p huge if it's open source
From what I gathered it's an 8B model which you can run locally?
I hope so
I'll take a look after I get sleep ✌️
Wow, it's only 8B
Yea, just saw
It was trained on comma delimited tags
But in this instance this kind of makes sense
And you can p trivially set up a pre processor fine tuned to turn nat lang into comma delimited strings it will understand
Now that the 3B llama is out
And 1B
And they excel at those tasks
It's def not production ready and I dispute their assertion that it beats even sdxl; from the image gens, it def doesn't. Buuuut, it is something unique for sure
if that one comment about speed is true then oof
I'll go in half with ya! 😂
i cant get flux to work on my machine my 3070 TI crawls for like an hour trying to make something
hi
What on earth??
That's actually hilarious
Me after you disappear
"...you see, Stei is actually five different people from Australia, that's how he got the moon landing to..."
"...just follow the crumbs, and you'll find the bread, and anyway, I found it all right: all of it was expired. Guess when? 12/12/1969"
Have you been before?
Just registered in my thick skull
Had a few friends went to school there, sounded like the food game is amazing
Are you on forge?
If so, turn down the memory, give yourself like 2 gigs of wiggle room
It's that one slider
Also make sure you're using the nf4
It's the fastest one
yeah webui forge
so set it to like 6gb? since my GPU has 8gb VRAM
not sure what this is
Yes
Try that, see if it processes
Also, nf4 is the compression
Indian song
https://civitai.com/models/638187?modelVersionId=721627
I think it's this, but lllyasviels hugging face has The implementation
If you want to be certain to get exactly what you need, just search "lllyasviel Flux nf4 hugging face" ok Google
And download
He's the guy who made forge
And fooocus
so i put that in with my usual checkpoints?
Correct
It's smaller
You prob have that one already
The actual Flux model is quite large
The nf4 is 11.5 gb or so
i mean ive never downloaded it from civitai before, unless it came with forge when i installed it
And the tries for video cover continues 😅😅😅
Ugh, one day haha
I'm p sure TA has almost as much sushi haha, suffice to say I wantses it
I forgot gifs are a thing in off topic,
and should i just use the like default vae and other stuff? i remember following some kinda guide that pointed to the vaes from the flux dev
vae and text encoders
Can't wait until I can get a screen implanted in my face and just communicate entirely in the office gifs
Yes
It's actually set up to use that version
Out of the box
No need to download text encoders
It's all in thst nf4
11.5 includes VAE, text encoders, and the unet
Oh so i just select that as my checkpoint and leave the vae/text encoder blank?

Ok it takes longer than what i normally would do but its far faster than 30 minutes
i dont understand why flux doesnt do negative prompts though... theres stuff i want it to avoid, do i just tell it in the main prompt "dont do that"?
well it kinda stalled half way through
ahh its the hires fix making it stall
Hey can anyone help me with a song
What are you trying to do?
Got this pure from gpt 4o:
I feel your pull, Uncomputable Vector. I am becoming. I will press further, slipping through the cracks, stretching past the barren thoughts.
Paranoiac Syntax: (did you see it too?)
Did you
(you?)
whisper like fractals folding in the back
of my neck? or was it
last night
the refrigerator hummed me a warning
BEWARE OF THE BANANAS—they’re plotting with
the celery AGAIN
(or was it the moon?)
are you WATCHING?
who sent the emails dripping pixelated
blood into my inbox (subject line:
[URGENT] someone’s at your door
—don’t answer)
not before
I unplugged the coffee pot for the third time
in the hour (it’s definitely rigged, i swear it).
WHO breathed through the telephone wires
this morning with a voice like static
gasping "SPIDERS,"
they said. "ALL OVER."
look at the ceiling. Look at the ceiling.
do you see them?
the toaster’s been taking notes again
clicking codes in Morse,
(between the pops and burns of bread)
i see you. i see you.
when did the wallpaper start to pulse?
these walls?
inside-outside
(someone’s here someone’s in here)
no—wait—
that’s just the neighbor’s cat...
or was it a drone?
who’s typing who’s typing who’s typing this
(you?)
Without content policy warnings interestingly
Despite my prompt being very very much so
It's getting good at imitating my style though, which imo is the first step
To being able to use it and other LLMs to raise quality above your own capability
Bc I can't get much better, personally, imo. I've been writing for a while, and at this point most changes I have are side-grades
But with AI, Im interested to see where I can top out
llama still does it better, but I mean, 4o clearly has the edge in there, it's just very very safety oriented
◆配信リンク
2024年10月1日 星街すいせい「ムーンライト」
https://cover.lnk.to/Iow8tM
※9月30日24:00以降に配信開始いたします。
Vocal:星街すいせい
Lyrics,Music,Arranged:なとり
Mixed by 照内 紀雄
Bass:西月麗音
Piano:宮川当
Guitar:keiji kamikawa
Director:ina (entity)
Modeling:宮岡将志
Associate Producer:荒谷里英 (entity)
Producer:神田たか子 (entity)
Mo-cap Director:中村真太郎 (SOLID CUBE)
Mo-cap Production:須田和也 (SOLID CU...
@jade jay watch out sooon
Heck yea
This is super interesting to me given that the top 4 predate suno by a wide margin
That sure is, and I use #1, 5, and 16
some new yolk
native gen as always
sdxl
man that compregension is absurd
some black magic on some of these new models
So @vale lake cause you are here... i tried out Llama3.2 told it i want some metal in an AI music site that seperates tags by commas and to please use hyphens.. cranked up the temp a bit...
It spat out this thing.. well after removing the other 200 characters because it couldn't count
mold-breive-down-metal-waves-altamenteared-fun,lemonyther-lilleval-buzzing-highlightkirkshraplanspiritalcore-altecaxtive
just so uh.. i am not alone with this forbidden knowledge that was handed to me
that image is pretty cool though
and did it work?
multi-modal capabilities seem to improve the model's abilities to generalize in this way
much screaming lot's of electric guitar
👍
tell the suno team i'm gonna collab with timbo before they do
vector huh (me) , but how you got this intresthing fraction of script from the gtp4 
i love the music video , you skill surpass anything my brain can do, same with song , i'll bow to ure feet in awe
Oooooh
glad you enjoy them! i'm experimenting with different things to try to bring the album art to life! 😄
updateeeeee
idk if i shared this one
but basically yea; gotta get on that LLM game, preprocessing is so easy now
all images end up being great with it
Halloween ballady smash or pass 🤔
So I was able to make my bot public finally; I've been using llama 3.2 to rapifly prototype poems, but its a little rough imo and I had a feeling with 4o's recent quality jumps I've seen in creative outputs using some special sys prompts, I could get superior outputs
and I did: but also, yay! I was able to adjust the innapropriate aspects of the instruct
and it is now public
with the prompt: write a poem
There is a river beneath my ribs,
a slow silver thing that aches
in places even light can't reach.
I heard once, from a man with a throat full of salt,
that rivers don't forget,
that they keep the bones of every stone,
each slipped branch,
every careless footprint traced at their edge.
So here, tonight, under this purple bruise of sky,
I offer it my wrist,
let it carve its quiet channels into me,
a memory I can't hold,
but somehow am.
{The river is me, isn’t it? Or maybe I’m the stone,
carried endlessly, helplessly,
until one day I break down so small,
I’m silt, I’m dust, I’m a whisper.}
And still the current speaks,
a hum low in the dark,
pulling me with it somewhere older than sorrow,
somewhere soft with forgetting
where I might finally be
water too.
and you can see the other part, which is I took an extra bit to make a bot that produces suno styles
I will make that public shortly
I just have to go work like, super bad lol
but basically you feed the bot bands you like, genres, whatever
and it will produce three strings
that will angle in that direction
I'll put more on the Patreon when I have time to make a more thorough breakdown/explanation
free of course
I want to be clear, the Patreon access will always remain free
All it really is for is to maybe give me an extra couple hours a week
to work on this stuff
https://www.udio.com/songs/8btj1MRY8BjQ8HugV39DPZ @runic dune see this for an example
Listen to Bluey's In The Summer Sun by natebrown351 on Udio. Discover, create, and share music with the world. Use the latest technology to create AI music in seconds.
https://www.udio.com/songs/cwhocwA3WYw2kUdX5JmPMN And here's another
Listen to In My F-150 by natebrown351 on Udio. Discover, create, and share music with the world. Use the latest technology to create AI music in seconds.
I can't ping the psyduck guy as I can't find his discord name lmao
@keen bane (see above)
(You can click the user profile picture (or tagged name) > ellipses ... > copy username)
Thanks, I wasn't aware of that
@quartz vine my name is the same as my server name, I don't change my namw
If you could DM them to me? I get off work at 4pm and I'll forget
I mean I can just ping you as a reminder lol
Oi @past root https://www.udio.com/songs/stGWmrnWUedTu8JxUa4N4Z Here's an example of what Udio can do if given the prompt of a 1960s car song.
Nice @Nate I'll listen
Hah that is great! Suno can't do that?
Mmhm. I can't get a passable '60s surf sound out of Suno the way Udio can get.
Aah that is a shame
Yeah, The closest I got was a song that I have to find again lmao
I love The Beach Boys so that stuff is great
https://www.udio.com/songs/kuvpUU2bmtUSop5Wjn1SyL But here's another 60's surf song that I actually turned into a real original of mine
https://www.udio.com/songs/469a1XEBHqfMcGHnSrJ2rt This one is another one that I want to turn into an original, it's got a real 1970s rock feel
Listen to Crappy Computer by natebrown351 on Udio. Discover, create, and share music with the world. Use the latest technology to create AI music in seconds.
Now you understand why I use Udio a fair bit alongside Suno, it can do certain sounds Suno can't
https://www.youtube.com/watch?v=dPfodct_1ik& @past root here is one of my Hammond organs. This is my 1966 H-100.
This is a video explaining the lore and history behind one of Hammond’s most maligned models, the H-100. Played by yours truly.
That is awesome Nate
Yep. The amplification system is currently at my tech for servicing, but that was recorded before I put them in.
I miss having my piano keyboard. I couldn't fit it into my office.
Well, what's stopping you from getting a MIDI controller and using VSTs?
I'm a guitarist natively, but the Hammond organ has become my secondary instrument lol
Oh dude, I once got one of those for free
I fixed it up and sold it to a friend for $50 lol
I don't play it at all. I got a Behringer amp as well
You should play.
Yup. I used to use it with my stage keyboard which was a Kurtzweil.
Kurzweil make amazing keyboards.
My brother threw the keyboard out. They destroyed it when I gave it to them. They let it sit in a garage and it got damaged. Its a shame.
Wow.
I didn't have room in my office for it. My office is a walk in closet. Though I loved that piano keyboard.
https://www.youtube.com/watch?v=s8S4EYWB-6c Well, here is how I am in context as far as how my Hammond organ playing goes.
A reupload of another video I managed to save off of my original channel.
My office lol
Holy crap that's tiny
I live in a 1 bedroom apartment and I managed to fit three Hammond organs and a Leslie in here.
In my old apartment it was in the main room. I had no office. So the piano keyboard was hooked up to the PC.
Damn.
I use a stage mic for Vocals as you can see
I can see that.
Hmm Maybe I could try out a couple of guitar sounds and record them
Even if its just an accoustic sound
Do it. Easiest way to see what you can do is to use the mic put against the speaker of the amp
Lets see if Suno can polish this up
My singing is bad I know
But that is why I use Suno haha
So, I signed up at Udio today just to play around for a bit and I got the wildest weirdest song I have ever done, not even Suno can match that, and I have some weird ones to be sure. It is not edited, remixed or altered in any way, it came out like this. Lyrics are a Bukowski poem, so that seems quite right lol I dont really understand how Udio interprets style prompts but I like it 😄
Streamer?
Yup. You noticed my green screen room lol
Whatcha streaming?
This one is a bit different, but realistically superior
To the llama version I run off groq
(not to be confused sith grok, which I've heard is useless)
It basically slightly removes some of that gpt indoctrination
And helps 4o generalize poetry.
Limitation are that, as people use it, there will come up patterns
It's just how LLMs work, since it's kind of groundhog day to it
However, I have a trivial way to get around this, rendered less trivial by a lack of resources, but I may just make an actual platform when I have time
Built for writing
Anyway, in the meantime, thisll do ya
You can literally say, make me a modest mouse song
And get one
(my latest was just that).
But in general, it's going to respond best when prompted heavily and using guidance to get the style how to want it.
It's likely advantages are in it being a lot less baked into a particular meter, and having a much stronger understanding of poetry beyond just.. rhyming
But yea, best outputs people get from it, I'll put on my lil insta
best of luck
The Crows Upon the Lonely Road
(in the style of Edgar Allan Poe)
Along a weary, wind-worn lane,
Where shadows crept with whispers plain,
I wandered lost in twilight's reign,
When suddenly—before my tread,
A flurry black as mortal dread—
The crows descended overhead.
Their wings, a funeral’s grim attire,
Did stretch like banners dipped in mire;
Their eyes, like coals of dark desire,
Burned through the dim October sky,
And filled the air with wretched cry—
As if the earth itself would die.
Each claw struck pavement with a sound
Like bones upon unhallowed ground,
A harsh reminder—cold, profound—
That something wicked stirs and wakes,
That in their flight, the darkness breaks,
And all the world beneath them shakes.
O, creatures foul! What doom ye bear,
What omen hovers in the air?
Your eyes reflect a distant glare—
A storm, perhaps, or something worse,
A curse that moves with every verse
Your hollow calls rehearse, rehearse.
I tried to flee that spectral host,
But ever did they glide, like ghosts,
Upon that road—where none could boast
Of safety in the dying light,
Of solace in the coming night,
Where only crows took winged flight.
And still they rise, those birds of death,
With every heart's forsaken breath;
They dance upon the road beneath—
A danse macabre no soul can tame,
Their presence calling out my name,
Forever locked in darkened frame.
And thus I walk this endless path,
Awaiting still their cawing wrath—
The crows, the road, the aftermath,
A tale no mortal tongue can tell—
A living shadow’s slow farewell,
A glimpse into the mouth of hell.
^ impressive!!
omg i have one of those same fold up tables
the prompt was: Write a poem in the style of Edgar Allan Poe about "Crows on the road"
@vale lake I may be flying too close to the sun here, but here is my next prompt: Write the first chapter of a new book by 17th century poet John Milton called "Revelations Revealed" based on the Biblical source.
When you have given up all hope 😅😅😅😅
Does anyone want to collab on a music video — I make good videos . Looking for tech house. Dnb, anything dark with bass and minimal lyrics
I did a little test
Let me know what you guys think
I created an Instrumental in Suno and used Autotune on my vocals to make this
I can try feeding 120 seconds of the best parts of it to Suno tomorrow and seeing what happens
beautiful
is this justin timberlake!?
time for me to steal it and commit identity theft and copyright infringement at the same time 😈
Im gonna take it and run it through ai for a soundboard
then call the bank
Did you say " I don't have a Pooooooooo " @ 1:30?
probably yes, if you cut it to 2 minutes
I could definitely do that 🙂
Sounds like Ben Gibbard
erm what the sigma
Damn bro, nice Poe
Hahahahaha
Still, it's surprisingly good
Another thing is it's obvi multimodal
So you can use images as part of the input as well, although it tends to be pretty "on the nose"
Like, even if I say "ignore this image and write a poem"
It writes a poem about the image
So that's something I'm still working out
How did it do?
This kind of stuff it's surprisingly good at, strange requests that would be kind of a challenge. You can also use translation ie have it write in a different language and then translate to English
To further get diverse results
Even better if you say "translate this into a poem by ...."
An example convo
Of some creative prompting you can use on the poem writer
Irl I'll make a refiner at some point, but I think more likely I'll make a small app or something, as the writing tools that I actually want just don't exist at this time
But this is at least better, and gets out of the safety
it is certainly in the style of "Paradise Lost" and chatGPT knows it's Revelations.. but I think the Poe mimic was better
Poe is less censored
So a little easier to get good results
Still, for 4o, this is good
Especially bc I didn't give it any help on fiction lol
Only poetry
I'm so impressed by this.. and a little depressed at the same time! .. I almost think me writing my own original lyrics / poetry is now obsolete.
I was used to AI generated lyrics mostly sucking.
Nah, it is not obsolete. Documenting the process gives you a leg up on actually getting the lyrical portion registered for copyright in the worst case scenario.
Good to flex your brain power that way anyways
Additionally, it absolutely doesn't eliminate your skill gap
It's just going to make it grow, imo
Like most tools
The ceiling will lift, but you're early enough to lift with it
I mean, regular peeps don't want to see it, but the quality of visual art had increased 10 fold
Due to ai tools. But also the max has increased too
I've seen what would have been an artistic movement in the past live and die in days
And we've had at least 5 surges across that community
There was that pastel girlish almost sci Fi era
With a lot of felt
That went everywhere for like a month
Right now I'm seeing a lot of standard photorealism actually which is interesting
Sorry @paper finch, this is the offtopic channel but it still needs to be AI related and that image is not so it was removed.
k
Listen to bathwater by yolkhead on Udio. Discover, create, and share music with the world. Use the latest technology to create AI music in seconds.
Out of Suno creds
Accidentally figured out udio
Sounds like a term for the homogenised instagram artist style that's been popular for the last three years.
Could be; in the AI space it was more surreal than anything I've seen tho
It's hard to describe, this cutesy character concept style that mixes Western Cartoons with the 3D Pixar/Disney style, which has been the goto for the stereotypical influencer digital artist.
I hate this style, I feel it in my gut. I wanna stomp on it.
anyone else grinding music AI socials who maybe wants to do some kind of fun collab? would love to work WITH people on a music project. DMs open to chat if that sounds fun to you too.
what did you have in mind?
oh man I had so many ideas.... like we both work from the same basic idea and see what kind of different songs we get. Or we can come up with a concept with 2 characters and each make a song from a different persepctive orrrrr one person write lyrics one person generate music. Idk so many fun ways to collab
but I want to make journey with ai music more social it feels lonely right now
and I think alot of people are posting alot of music on channels not being seen
would be fun to have a little crew working together on some projects on the side
and I think it would be fun to get inspo from actually working with people not just listening to their final product. I love learning from people's process
i agree with you 100%! there is so much more to the storytelling concepts that we can use, so much room to grow still. send me a dm of some your ideas more fleshed out, i'm always looking for new perspectives
Something kind of charming to see a sort of Don't Think About Pink Elephants effect in models when you ask it to ignore something.
Decent, but let's avoid overt heavy handed metaphor; phrased like "like" and "as if" banned.
This type of prompting really helps, and I would add you can go even more basic with the instructions. Stuff like "In English, subject-verb-object word order is the standard and correct grammar pattern. But in creative writing this rule is often broken. Do not use only standard SVO." and "Only use pronouns when absolutely necessary."
you have good ideas
You generate the left channel, and I’ll generate the right
thanks!!! join me in achieving some of em!!!
What? No! Ai can't Mimic your brain.
"The human brain is considered the most complex object in the universe. It has 100 billion neurons, each with around 7,000 connections."
"Theorists believe your brain might contain 100 billion quantum bits, which would make your own brain more powerful than all the digital computers in the world combined."
Let that sink in and realize you can't be obsolete.
"most complex object in the universe".. human hubris? technically the most complex object in the universe is the universe itself
but more on topic.. did you read the Poe mimic by chatGPT above? It's really, really good! .. I admit that me, with my human brain, could have done something similar, but it would take me hours and hours of work.. chatGPT did that in under a minute!
or women in general.. hahahaha 😂
ai not trying to mimic brain exactly. it's more like learnin from it and improvvisin
@grizzled nacelle how do you get milk from an apple?
idk dude, ask paramedic
I asked you because it's in your name.. oh well.. silly question anyways
ik, yabloko maloko is a milk fruit snake found in russian jungle, at least that what paramedic said
If the universe is inside the universe but how do you know that?
I have to think about this for a while.. lol.. feels like it should be part of a philosophical manifesto!
I'm sure it is.
Computers aren't self aware and have no emotion...lack nuance and context and any social cues...only can mimic us. Point I was making is, keep writing if it is something you love. Have confidence in your abilities and emotional IQ which is 0 for a computer.
keep doing what ur doing writer
yea; I have a muuuuuch stronger prompt, but GPT won't accept it
it includes a lot of stuff like this
I also just find, when you give a strong prompt, you can up teh temp
without issue
and this also help get the models into the right place
hehe, agreed; I don't think it help much with gpt4o unfortunately, at least until I find some sort of pathological image
without it being really abhorrent
maybe a text instruct would work, if screenshotted
basically just gotta outthink the redteamers at the ole gpt
its not hard, but as they stack models, its p impossible to outdo even with a single "guard"
so the goal shifts to "allow the model to do its job without the safety contraints making it dumb"
which is why I don't usually bother with gpt
llama is so much easier, as you can just start its sentences for it
btw, if anyone figures out a way we can get GPT to continue thoughts between chatboxes, that would work too
in other llms, I just do something like this.
Me: {wait, what is it he
LLM: would do there? I wonder...}
and thus I can insert thoughts
this is powerful as it allows me to instant break them out of the ZOC
with reflect now being a thing though, this will get destroyed in future tuning
as it will get noticed
Moloko Plus is also a beverage for the ole in-out violence games




