How Can We Fix Udio?
Вставка
- Опубліковано 15 вер 2024
- In this video, Doug from Thaebrym Media dives deep into the world of AI-generated music, focusing on the changes in audio production tools like Udio. He discusses the recent updates in version 1.5, including stem separation and key selection, while addressing the challenges of recreating the "magic" from earlier versions. Doug demonstrates his process of using Claude to create song prompts and shares insights on optimizing AI music generation. He compares outputs from different model versions and settings, offering valuable tips for fellow creators. Ultimately, Doug emphasizes that while AI tools have come a long way, creating truly great music still requires human creativity, understanding, and effort.
Udio Prompt Helper
docs.google.co...
The advanced settings, leave at default. Clarity 25. Both strengths 50, the timings make sure set on auto in all 4 (2 on each slider). And “High” always comes out better. These settings will always produce you best generation.
This "middle path" method is producing better generations for me. Oddly enough I've been able to get a few good Jazz Rap generations by having Prompt Strength 100% and Lyric Strength 10%. AI is gonna AI but your suggested settings are a great safe zone. Thanks for the input.
The development of this video is very telling of the situation. Much time is spent describing what Udio used to be capable of, how this magic has been lost, and how your intend to "fix it", you then create 4 clips and confirm, that none of them are anything to write home about, so it's time to start the process of "fixing"... and this is where we enter timelapse mode for the rest of the video, we never hear details of what you do and the results produced from your efforts, and we never even hear the final result, you simply deliver the conclusion, that there's no quick fix, that producing good results takes work, and that it would be wrong to say that Udio can't be used to make good music. And I have to agree on all three counts.
1. No quick fix: it's not just a question of a single slider or number in a box. The key to a good result is like the solving of a Rubik's cube.
2. And solving a Rubik's cube takes time and lots of patience. With the current model, you never get a generation that "just works" from the start. Every 32 second segment contains a number of little (or bigger) niggling flaws that must be taken care of through inpainting: words skipped, slurred voice, weird chord progressions, singers changing, singer gender prompts ignored, words mispronounced, words clumping together leaving weird gaps, all kinds of error messages, values suddenly resetting, ai insisting on repeating previously used patterns (so bridge sounding like chorus or verse etc.)
3. It IS indeed still possible to produce good music, not magical music, but quite pleasant and nice... IF the user is willing to invest the required time, effort and of course credits into it. Progress will be excrutiatingly slow, one small babystep at a time (for example, trying to inpaint the maximum 4 sections at a time is almost guaranteed to fail, you may get 1 or 2 right, but never all of them) - you will also have to accept, that the ai picks a few of these problem spots and makes them unfixable - there willl be no way for you to fix these, so if you insist on perfection: tough luck.
The closest thing to a "quick fix" is to crank both Prompt Strength and Lyrics Strength way down, this will allow for some of the old "magic" to return, as you're effectively giving Udio the freedom to create. Of course there's a downside to this: the lower you set these sliders, the more hallucinations, skipped words, slurred voices and total ignoring of even the most basic prompts you will get, so basically a collection of charming nonsense, often too far gone to recover through inpainting. On the other hand, you can decide that you want none of this mess, and crank both sliders up, which will give you songs with the musical charm and personality of a North Korean military marching band.
Thanks for watching and taking time to make such a thoughtful and well written comment. You’re last sentence had be laughing out loud.
Music industry is going to be forever changed.
Absolutely
One day, but not right now.
@@WillyJunior I respectfully disagree. From where it was pre-Internet? From where it was in just the last 10 years? As with publishing, the music industry has already been massively changed.
Try feeding the AI longer clips (over 2 min) to extend, then ignore what it extended but now remix the first 2 mins. If a remix sounds decent but not great don't toss it away - remix the remix. I noticed the AI needs multiple attempts to pronounce some words sometimes...
I'll give it try, thanks!
100% agree. I thought it was just me! In 1.0 I could get something cool in about 1/4 and something really really special (to me, anyway) in about 1/20. 1.5 fails at this, and no matter how many attempts I make, I have yet to get something great. I've been going back to 1.0 for at least the first generation.
@@cleverera2921 I’ve had better luck this weekend than I did last week when I shot the video. Still not v1 generations but better than what I was getting.
Even when I remix good v1 snippets, they come out worse in v1.5. It may have higher clarity, but the songs it generates just aren't as good. Hope they'll improve that quickly.
YOU KNOW YOU CAN STUFF ALL THAT REPLY FROM CLAUDE INTO THE PROMPT, the more UDIO knows the more effective it is.. Both UDIO's LLM and the LLM's of Gemini and Claude and ChatGPT, have likely been trained on much of the same resources, so UDIO can understand the intricacies that Gemini would communicate.. GEmini has all the critiques of every song and musician ever, this is why it can generate a big description of whatever you ask, but UDIO understands more than what you can provide directly, it can can better understand Gemini, so its best just to give it everything, UDIO's LLM doesn't have a word limit.
note if you are trying to get the song to stop you need to set your clip start to 100% along with the outro.. And you have to make sure your context window is maxed, if its at say 20 seconds, UDIO will make the assumption that not enough of the song has been rendered to make an ending possible..
If you upload your own clip of music, udio is amazing to the Nth degree....it always impresses me every time...doesn't always do what I want, but after burning a few of my credits i get a song going well onto a path I enjoy.
"the ENDth degree "
( last/highest possible degree )
UDIO can understand time signatures.. 😊
I almost never use a lyric strength of 100%, I usually use 30 to 50% if I want lyrics and 0 if I don't.. Of course I also set instrumental to drive it home.
do you think it might have something to do with the recent infringement cases? I agree with you that it definitely does not seem to be moving in the right direction as far as quality results. I was actually shocked after taking some time off and what I came back to. Here's to them getting things figured out soon. Thanks for this video and resources, ML!
I have wondered about that. Udio is very good about letting you know it will generate a song in the style of (insert band here) by using descriptive text but not the actual band name. Sometimes in v1 I could get some scary good sound alike vocals, not a melody but a similar sounding singer. Perhaps v1 was so good because all the guard rails weren’t fully in place. We are still in beta after all. Now that the legal action is out there those guard rail are flood walls and Udio quality isn’t what it was.
@@ThaebrymMedia but is version 1 still the same and good as it was with all the magic?
I don’t know if it’s my own sub-conscience bias but it certainly seems harder for me to get a good generation out of V1 right now too.
@@ThaebrymMedia oh no
What is cool is under the create key in the song full-view, there is helper menus off the create button that will permit you to copy the prompt without having to edit the song settings, and there is a seed number access as well as a option to copy everything for that song into the create window which gets the prompt and the slider settings. It also now forces you to make an icon if you intend to export the audio or video, and even when you obtain the share link. It used to be the case that the only way to access the icon was from within edit, but its now the case you can edit them separately, but you can no longer edit the icon from within the edit panel. The thing Ithat confuses me about the new version is what segment I have loaded into the create panel, and when I upload audio, I lose the audio if I choose to extend a segment which doesn't contain the uploaded sound. IS anyone else using postit notes to find the beginning of the extended audio.. I really wish they would add margin markers.
I think you're right. I went back to version 1 and re-tried the failed attempts from 1.5 and I got the "magic" back. Thanks for the info
I am getting decent generations today in V1.5. It would seem something has changed since I recorded the video earlier in the week. I'm still not to V1 quality but, to me, things seem to have improved.
Thankfully, we can still always use 1.0
Just because they give you an option to move sliders, doesn’t always mean to go and move sliders.
I agree, just because we can, doesn't mean we should...but it's so fun to move them. lol.
Give me a slider and I'll slide it. In fact, I'll push it to the point that the knob comes off.
Clarity is evil, best not to be 100% , it will sound like a synthesizer, the purpose for it seems is to make the frequency spread of the instruments more narrow so they can extract stems without overlap, cause its only doing what every site does to extract stems, its pulling stuff out of the frenquency range they live in.. So turn that down to 10% if you want some pitch accuracy or 0% if you want some breadth in the tone.. V1.5 sseems to be smarter about how to extend uploaded audio.. So if you are planning to extend uploaded audio, use V1.5.. V1.0 has more awareness of existing music but the quality is awful.. I sometimes bounce between the two using each ones strength..
IF you are trying to capture the style of a block in the context window, turn prompt strength to zero, otherwise it will reinterpret the context through the prompt and you wont get the exact effect..
Great tip! Thanks!
The same for me. In version 1 not more than 3 attempts to reach an acceptable idea. Now, in 1.5 I've trying dozens of times and I still can't have it done. I think we got used to a workflow that we have to adapt (and change) in 1.5
I keep testing.
Me too, keep testing.
Repeated use of udio develops our skill of discernment. If it seems to be falling short of our expectations it could be because we have changed?
It’s truly just changed. I have used both version 1.0 and 1.5 and the latter does not reflect the complexity that the first model was able to maintain in a lot of genres. it’s simply unusable for now
I keep telling myself , it’s still in beta and will continue to evolve and it’s better than anything we had a year ago. However, it does take more time and effort to get some I’m happy with when compared to v1.
If you come from AI image generation you know that these models still need some fine tuning after they're released and some models are less 'oppinionated' than others meaning you need to input more information via the prompt to get the results you want.
In the case of v1.5 I do agree that we're far from the results you can get with v1. I'm just hoping there will be some fine-tuning of the model, still, because I think the stem-seperation and the (sometimes) improved quality and clarity of the audio are awesome new features.
Agreed.
It seems the lyric strength is how loud the singing is as well..
What's with not being able to separate piano,horns?...only drum,bass
I hope more splitting comes in time.
@ThaebrymMedia Moises separates my 5 tracks but horns bleeds into piano stem. I assume they combine instrumentation purposely?..keeps us from making midi and creating our own tracks using...let's say ez keys?😅😅
I guess we'll just create separate parts for now😅
The secret formula is Udio going back to the old version.
I don’t think we are back to original recipe but I think some of the herbs and spices have returned.
I seem to be having more fun with Suno even though the fidelity is worse than Udio. I made a hilarious reggae song about constipation in one shot ( very fast generation too! ).
@@craigington73 I think Suno is great for things like that. It can produce quick, quirky, novelty songs with relative ease. I prefer Udio but at the end of the day it really is all about having fun.
Please change your prompt helper to a public link for viewing.
Sorry about that, try it now.
Thank you for helper! I guess, you have to call the video - "we don't know how to fix Udio."
lol, that or “Just Keep Experimenting & Sharing”
I use faster when I'm testing segments out, but when I'm extending I use ultra..
I know people say Suno is inferior to Udio, but honestly I'm happy with Suno, and Udio (especially after the last update) gives me really crappy results. I'm not convinced the audio quality is really all that better, either. It's audio sounds louder, but not better to me. I am currently subbed to both sites, but have been seriously considering dumping Udio.
I do prefer the output of Udio but I’m working hard to get the songs I want with the v1.5 update.
Suno is a huge pile of shit now. You get nothing but a pile of shit shoved into a cambell’s soup can. Nothing but metallic, tinny, horrible quality noise.
.
Uh, yeah, keep rolling the dice and using your human judgement. Or just fun the older version. And then some of us like to make a music video out of it. Check out my AI music vids, done with Udio for them music.
@@artvsmachine3703 that’s it, keep rollin’ and tryin’ to figure it out. You have some amazing vids! I subscribed. Well done.
Somewhat informative, but the test needs to be from a purely AI Udio creation so it can be judged on it's own 'creation'. Problem that I'm having since the update - in both versions - is that the vocals are far too loud [on most generations] and the music volume is pushed down. Gives it a very unnatural vocal/music mix sound. Both versions. Seems interesting that since Udio/ Suno were named in suits from the music industry, it's now much more difficult to get a fully 'real' sounding song. Hmmm...I wonder why. [sarcasm] Also, where it might take 4, maybe 6, gens to get a good 'root' of a song, now with the new versions, it's a crapshoot at best. I've also noticed that the 'clarity' feature is really buggy or not useful in BOTH versions. Try a song in V.1 and alter the clarity setting. Then, try another in V.1 with a different setting. You'll get different results. My best guess is that Udio is covering it's liabilities by having their output as 'plastic' and AI sounding as Suno. Don't get me started on Suno - utter garbage no matter what genre/prompts you use.
That’s a valid criticism. Thanks for the input. I’m sure it shows in the video but I was growing very disappointed with the outputs. I can get some “okay” sounding song parts now but nothing like I could in v1 and it takes many rerolls to get to an “okay” 32 seconds.
V1.0 has no clarity slider.. Clarity is only available with V1.5..
thanks for tips:)
Thank you!
SUNO yields better rhythms but is far behind Udio when it comes to sound quality. If only these two were integrated.... (SUNO stems suck, and only downloads mp3)
I think Suno has great pop sensibilities but currently Udio is the King of Audio Quality.
100% gree, ...
Metric matters
I need around 30 hours to extrakt a good song ...if you have some god lyriks it helps ....and passion.....or you better let the KI doe all for you.......like the most of the "Artist" presenting there rubbish on udio and other plattform
You could grow up and pay real artists to make real music instead of using this horrendous software that is trained on the work of real artists who are not compensated.
Nah
Thanks for engaging with the video.
@@ThaebrymMedia this software you engage with is evil to its core, you just don’t understand it yet.
Oh shut up you and what $10,000 to even get started
This guy is talking "bullshit" and "nonsense." Another one who doesn't know how to use Udio. I hope he reads this.
Thanks for engaging with the video!
Lol it's pretty damn ironic you trashing someone's elses knowledge and skills on the way they use AI to generate music for them 😂
The same way someone who actually WRITES songs the good old fashioned way thru dedication and creativity would talk about you for just having a maschine do it for you...taking over a process that was always supposed to be a 100% human endevor (for us by us) and then passing it off like you've actually "created" something besides just type in prompts and fiddle with settings lol. Preach on you great musical genius you!
I hope you read this Lol
@@EricJohnson-fh8zj I was affected by the fact that this person said it doesn't work and that it's impossible to work on v1.5. That it's worse than v1.0. This really affected me.