-
-
Notifications
You must be signed in to change notification settings - Fork 12
🎰 Success Story
... people just started to become crazy & aware of the KaraFan's power :
Karafan 4.5 is Amazing !!!!!
I obtained a very good result on a title which has always been complicated to separate!! The main problem is the violin which tends to start in both VOICE and INSTRUMENTAL parts instead of INSTRUMENTAL alone!
Karafan 4.5 passed the challenge, where all the other models I tested failed
Well done Captain Flam! ✅ 😍
I like your results, they compare very closely to my best efforts using what I know, so it's a definite time saver. Thanks for making it.
I been playing with karafan, man o man. The instrumental quality is so damn crisp and clean, it blows Every other instrumental separation tool out the water. Good shit, bro 💯🐐🐐
For your information :
Anjok is THE BIG BOSS of all of us. 🥇
He's the creator of Ultimate Vocal Remover 5 (UVR).
He has trained (with Aufr33) the majority of MDX A.I models that we can use as the last state of Art !
So, I am very grateful for his glowing comment :
I did a deep dive into KaraFan and compared the results to those in UVR using different settings. I tested on several genre types with all the presets, and here's what I found.
- KaraFan does a great job leveraging the models to minimize bleed. However, some tracks I tested did have more unwanted vocal bleed in the instrumental than UVR. Some results were similar to UVR using Max Spec or Average in ensemble mode. Though looking at the code, I can see KaraFan implemented a clever, nuanced approach to minimize bleeding with the ensemble code, which includes pitch shifting, high/low pass filters, etc. That implementation did give better results for some tracks I tested.
- For non-MDX23C models, I noticed that KaraFan relies on the older chunking method instead of using batch mode with the native chunks the model uses (basically, it's double chunking, which I found hurt results when UVR used this method in the past). KaraFan could improve if that was implemented.
- Overall, your mileage will vary depending on the track you are processing. If you don't get your desired result from UVR, use KaraFan and vice versa. Between KaraFan & UVR, there are many different ways to get good results.
- For some tracks, one of the biggest hurdles, it seems, is getting more "official" sounding instrumentals that don't sound muddy. I've actually been cooking up a new way to fix this that involves filling in the spectral holes with generated data. It's still a work in progress, but so far, my results have been promising for heavier music (metal, hard rock, etc.), where this issue is more pronounced.
I'll be releasing some new models soon, including Inst HQ 4.
So, I'm excited to see how KaraFan leverages those to get good results.
@Captain FLAM : Keep up the good work! 👍
This 12th of September 2023 (on Discord), I've receveid my first compliment from one of KaraFan users :
It's weird why Source Separation Ratio (SSR) doesn't exist, it is also a big part of defining quality.
Like how well the model separates from it's original signal, how often it false detects other instruments, and stuffs.
I can tell not all model are the same and have different SSRs,
but tbh, if talking about source separation ratio, the KaraFan wins from ALL models
Salute.
...
eyyy, I find it cool tho where it actually mute the signal when there is no vocals detected.
MDX23C 8k FFT usually have around -70dB separation ratio, but for KaraFan, it's beyond 120dB or in short mutes the signal which is insane !
We were talking about SDR score that define the quality of the separation.
... and « MDX23C 8k FFT » is actually the BEST model with highest SDR score !!
But ... I'm only talking aboout HIFI quality for my ears ... once again.
And also :
I answered Bas Curtiz (who made the SDR tests, thanks to him !) in private message :
I could trick the SDR algorithm if I would 😉
Remember all of you : I'm a hacker ... (white hat)