The ULTIMATE Local AI Update Just Dropped! ACE-Step 1.5, Paints-Undo & Whisper Premium Will WOW You

[FurkanGozukara]

The ULTIMATE Local AI Update Just Dropped! ACE-Step 1.5, Paints-Undo & Whisper Premium Will WOW You

Full tutorial: https://www.youtube.com/watch?v=hzKSt5WUAm0

In this tutorial I show the newer ACE-Step XL 1.5 Premium features, especially the corrected remix workflow that was missing from the previous video. You will see how to remix songs properly, convert lyrics and language, tune remix strength and melody retention, regenerate only selected parts, use the Library metadata system, install the upgraded Paints-Undo pipeline, reduce VRAM usage, and fix repeating lines in Whisper Premium transcriptions.

Important links:

Download ACESTEP XL Premium files: [ https://www.patreon.com/posts/ace-step-1-5-xl-premium-157675060 ]

Discord: https://discord.com/invite/software-engineering-courses-secourses-772774097734074388

Google AI Studio: https://aistudio.google.com/

Previous ACE-Step tutorial: [ https://youtu.be/9C_6qNKjgpA ]

Windows requirements/setup tutorial: [ https://youtu.be/DrhUHnYfwC0 ]

Whisper Premium tutorial: [ https://youtu.be/4lAk6sf1qF8 ]

Download Whisper Premium App Files: [ https://www.patreon.com/posts/whisper-webui-premium-145395299 ]

Video chapters:

00:00:00 ACE-Step XL 1.5 Premium update, remix focus, and Paints-Undo preview

00:00:49 Billie Jean remix example: checking style preservation and vocal change

00:01:16 Lower remix strength idea and Gangnam Style Korean-to-English demo

00:01:36 Hearing the Korean-to-English result and why extreme remixes sound strange

00:01:53 Stronger remix settings with higher strength and melody retention values

00:02:12 How to update ACE-Step XL 1.5: download ZIP, extract, and overwrite

00:02:37 Run Windows install/update.bat and rebuild the virtual environment if needed

00:02:56 Library tab overview: daily categories, saved generations, and metadata

00:03:18 Loading old songs from Library with lyrics, parameters, and JSON restored

00:03:33 Starting a proper remix: select the SFT model and open Advanced -) Remix

00:03:45 Uploading the source song and learning the two key remix parameters

00:04:04 Remix presets explained: different lyrics, same lyrics, medium and big change

00:04:19 Why lower melody retention lets the model generate completely new lyrics

00:04:30 Torch compile speed tip and preparing the target style caption and lyrics

00:04:42 Using Gemini in Google AI Studio with the ACE-Step lyric instruction file

00:04:59 Editing or writing lyrics and matching the vocal language accurately

00:05:15 Launching a live remix generation and measuring local generation speed

00:05:27 Live timing result: around 33 seconds for a complete remix generation

00:05:49 Use generated result as source to repair or improve selected song parts

00:06:03 Selecting remix start/stop points and regenerating only the chosen section

00:06:17 How section patching works: full remix generated, only selected part replaced

00:06:36 When to keep lyrics/style the same and when to change them for a section

00:06:47 Comparing the full output against the newly generated section preview

00:06:58 Iterative remix workflow for perfecting each part of the composition

00:07:14 LoRA training progress, future voice accuracy, and language-swap limits

00:07:27 Fast local iterations, multiple attempts, no watermark, and usable outputs

00:07:55 Final remix reminders: SFT model, proper Windows setup, and compile mode

00:08:11 Tuning dramatic changes with percentage values and fixed seed comparisons

00:08:41 Paints-Undo upgraded intro: new pipeline, faster speed, and better results

00:08:52 Download, install, and start Paints-Undo with windows_startup.bat

00:09:08 First launch model downloads plus new xFormers Triton attention support

00:09:21 GPU compatibility, Torch 2.12.1, CUDA 13, and upgrades over the original

00:09:34 Upload an image, generate the prompt, and tag it with the WD14 tagger

00:09:52 Operating steps, keyframes, Tiled VAE options, and low VRAM preparation

00:10:04 24GB vs 7GB VRAM usage and how the new memory-saving options help

00:10:27 How keyframes become the drawing video before final video generation

00:10:52 CUDA 13, Torch 2.12, Triton attention, diagnostics, and speed improvements

00:11:10 Automatic attention fallback plus Linux/cloud installer compatibility notes

00:11:27 Full generation time, possible torch compile addition, and result preview

00:11:46 Why keyframes matter, why results vary, and why not every image works

00:11:57 Fixing out-of-VRAM errors by lowering resolution and using supported ratios

00:12:14 Whisper Premium update: transcribing videos with all Whisper model options

00:12:31 Faster Whisper quality mode and the repeated sentence problem

00:12:42 Using large-v1 and repetition penalty to prevent repeated subtitle lines

00:12:54 How to tune repetition penalty carefully so transcription is not skipped

00:13:06 Example result: highly accurate subtitles generated from the new video

00:13:19 29-minute video transcribed in 1.5 minutes, 20x real-time, and closing

This video is for users who want fast local AI music remixing, better generation iteration, image-to-drawing animation, and high quality subtitle transcription. Follow the timestamps to jump directly to ACE-Step remix settings, Paints-Undo installation, low VRAM options, or Whisper repetition penalty tuning.

Video Transcription

• 00:00:00 Greetings everyone. Today I am going to show you  newer features of the ACE-Step XL 1.5 premium  

• 00:00:06 application and how to do remixes properly. It was  lacking and inaccurate in our previous tutorial.  

• 00:00:15 I assume that you have watched that. If you didn't  yet, please watch it. This application is amazing.  

• 00:00:20 It has so many unique features that you would  like. Today you will learn how to do remixes  

• 00:00:25 properly, and also I will show another application  called as Paints-Undo upgraded. This application  

• 00:00:32 is very unique. You provide an image, and it  reconstructs it like you have hand-drawn it. It is  

• 00:00:38 not perfect, but as you do more iteration, you may  get a really good video. So it is like starting  

• 00:00:43 from the 0 to repainting it into the full perfect  quality. Okay, so let me show you a few remixes.  

• 00:00:49 The first 1 will be famous Michael Jackson song,  Billie Jean. So let's listen a little bit of it.  

• 00:01:00 So you see it is keeping the style. The vocal  has changed because we are doing heavy remix  

• 00:01:06 right now. However, it is working really, really  good. You can also do lesser degree remixes.  

• 00:01:16 So if you make lesser degree, it won't be too  much changed. I will show all of them. Another  

• 00:01:21 demonstration will be entirely changing Korean  song into English song. The famous Gangnam Style.  

• 00:01:36 It sounds weird because the song is weird,  and we are changing from Korean to English,  

• 00:01:41 but still successful. As we have LoRA training,  hopefully, it will be even better. You can even  

• 00:01:47 do more powerful remixes by increasing the remix  strength and the remix melody retention values,  

• 00:01:53 like this case. So this is even more powerful  remix. These are the lyrics of the song itself.  

• 00:02:12 Yeah, it is weird, but this is how it is done. So  how you are going to update to this version? The  

• 00:02:18 link of the application will be in the description  of the video. You already know it. Download the  

• 00:02:22 latest zip file, put it where your installation  is. If you are making a fresh installation,  

• 00:02:27 it is the same. Right-click, right-click,  and extract all the files, and make sure  

• 00:02:32 that you are overwriting all the files. This  is the proper way of upgrade. For upgrading,  

• 00:02:37 Windows install or update.bat file, it will  fully upgrade. If you encounter any issues,  

• 00:02:42 you can also delete the virtual environment folder  and run the installer again. So this will remake  

• 00:02:48 the virtual environment. Everything will be  kept, only the virtual environment where the  

• 00:02:51 libraries are installed will be remade. It will  be really fast. This is the way of upgrading.  

• 00:02:56 So before I show you how to make remixes from  scratch, there is 1 another thing that I want to  

• 00:03:01 show you which I missed in the previous tutorial.  When you go to Library, you will see all of your  

• 00:03:06 generations. You can category them by the day,  as you are seeing right now. It is very useful,  

• 00:03:12 and when you click on the song, it will update  all the lyrics, all the parameters that were used,  

• 00:03:18 the full metadata JSON. So this is very useful.  You can play the songs from here. So the library  

• 00:03:28 feature is really, really useful. I had forgotten  that in the previous tutorial, but you know it  

• 00:03:33 now. So how to make remixes? Make sure that you  have selected SFT model. The remix works best,  

• 00:03:40 maybe only works with SFT model, we can say that.  Then go to Advanced tab, select the Remix tab,  

• 00:03:45 upload your song. Your base song will be here.  Let me show you. So this is the original song. The  

• 00:03:53 most important 2 parameters are remix strength and  remix melody retention. As you make these lower,  

• 00:03:58 it will make more dramatic changes. And we  already have 4 presets: different lyrics,  

• 00:04:04 same lyrics medium change, and same lyrics big  change. Since this is a Korean song and I want to  

• 00:04:09 generate it in English, I am selecting different  lyrics. So you see it is setting the remix melody  

• 00:04:14 retention to 15 percentage. Therefore, it will  be able to generate entirely new lyric. If you  

• 00:04:19 make this like 25 percentage, you won't see  lyrics changed. This is how it works. You need  

• 00:04:24 to make a few tests to understand how it works.  I am also using torch compile which speeds up  

• 00:04:29 my generation significantly. Then you need to  type your style caption and lyrics according to  

• 00:04:34 whatever you want to turn the song into. So I used  Gemini and this file to generate the target style  

• 00:04:42 and the lyrics. You can use the Gemini for free  on Google AI Studio. So type Google AI Studio to  

• 00:04:47 Google, go to this link, aistudio.google.com,  and then upload your files, and also upload  

• 00:04:53 this file that comes with the zip file, which is  ACE-Step lyric generation instruction for LLMs.  

• 00:04:59 Then tell whatever you want to do. So this is  the style, and there is this lyrics. This lyrics  

• 00:05:05 is modified version of the English lyrics that  were published. You can also type your lyrics  

• 00:05:10 yourself. Then select your vocal language. This  is super important. Whatever the language of the  

• 00:05:15 song that it will be turned into or the original,  you need to set it accurately from here. This made  

• 00:05:21 huge difference. Then you are all set. Let's live  generate 1. The generation is really fast. Let's  

• 00:05:27 see how much time it will take. Okay, we are  almost done. So 20 seconds, and in 25... okay,  

• 00:05:34 in 30, 31, and 33. In 33 seconds, I got the remix  generated. So I can generate as many as needed  

• 00:05:43 remixes and pick the best 1 that I love. Okay,  let's say you didn't like a certain part of the  

• 00:05:49 song. Now we have a new feature: Use generated  result as a source. Click this button. It will  

• 00:05:56 update the input audio and also this preview  audio. Then from here, select the remix start  

• 00:06:03 and stop wherever the part you want to regenerate.  Once you select it, hit Generate Music again. It  

• 00:06:09 will regenerate the remix and only change the part  you selected. This way, you can quickly iterate  

• 00:06:17 different sections of the song and get the perfect  output as you want. So how this works is that it  

• 00:06:23 regenerates entire remix but only changes the  selected part from newer remix into your previous  

• 00:06:30 song. So don't change your lyrics or style unless  you want to change your lyrics, of course. If you  

• 00:06:36 want to change your lyrics, change entirely, keep  entirely, and that section will be regenerated and  

• 00:06:42 copy-pasted. So you see this is the entire song,  and if you want to only listen the new generated,  

• 00:06:47 it is also here. It is shown here. Let me show  you. So this way you can iteratively regenerate  

• 00:06:57 and regenerate and get the perfect composition  of each part you want. This is super useful. This  

• 00:07:03 is really, really good feature, and hopefully,  with the LoRAs, we will get much better quality,  

• 00:07:09 much more accurate audio, much more accurate  voice. It will be like the original singer.  

• 00:07:14 The LoRA training is hopefully coming soon. I am  working on it, but we have pretty much perfected  

• 00:07:20 the inference. Now everything is ready. Of course,  changing the entire singing from Korean to English  

• 00:07:27 is hard. It is far from perfect. You need to  work on it, make multiple generations. But the  

• 00:07:32 generation speed is incredible, like 30 seconds  on my local Windows computer, so that you can  

• 00:07:38 iteratively improve song until you get a perfect  song. And 1 advantage of this application is  

• 00:07:43 that it won't have any watermarks, so it will be  fully original to you. You can use it anywhere you  

• 00:07:50 want. I will say 1 more time, make sure that you  have selected the SFT model. This is for remix.  

• 00:07:55 If you have proper setup of your computer with  following the Windows tutorials, you can enable  

• 00:08:00 compile mode. This will really speed up subsequent  generations. The first 1 will be slow since it  

• 00:08:06 will compile, but then it will become very fast.  And if you want more dramatic changes, reduce  

• 00:08:11 these values. As you reduce it, it will do more  dramatic changes. Like 10 percentage will change  

• 00:08:17 it entirely, like 15 percentage will change it a  little bit lesser degree. So you see we have also  

• 00:08:23 same lyrics medium change, like 25 percentage.  Play with the percentages and see the impact it  

• 00:08:29 is making. If you want to see direct impact on  the same song, you can also disable random seed  

• 00:08:34 and set a certain seed to keep trying on the same  seed value to get similar results. Okay, the next  

• 00:08:41 application I'm going to show is Paints-Undo  upgraded. This is an older application,  

• 00:08:46 but today I have updated it entirely into new  pipeline. So it is much faster, much better now.  

• 00:08:52 Download the latest zip file, the link will be in  the description of the video as usual. Put it into  

• 00:08:56 wherever you want to install. Everything is same  in my applications actually. It is using Python  

• 00:09:02 3.11. Just install it. So the installation has  been completed already very quickly. For starting,  

• 00:09:08 use the windows_startup.bat file. It will download  the necessary models automatically for you. It  

• 00:09:14 is using special new xFormers Triton attention,  so it is really fast. It supports all the GPUs,  

• 00:09:21 so this application that I made is much better  than the official Paints-Undo application, many  

• 00:09:27 times better. And it works with Torch 2.12.1 with  CUDA 13. Upload your image. I have a test image in  

• 00:09:34 the zip file. Generate prompt. It will quickly tag  the image. So you see it is first time downloading  

• 00:09:40 the model automatically. The tagging is like  almost instant with WD14 tagger. Okay, it is done.  

• 00:09:46 So this tag will be used for operating steps.  You can change these steps. It is like generating  

• 00:09:52 keyframes. And if you are on low VRAM, if you get  out of VRAM errors, enable Tiled VAE keyframes and  

• 00:09:58 Tiled VAE video. I have entirely coded these 2  options. And then there is also Low VRAM mode.  

• 00:10:04 If you have a 24 gigabytes GPU, you shouldn't  need to enable any of these, but you can also  

• 00:10:09 enable. They don't cause massive speed degrades.  It is almost same speed but much lower VRAM usage.  

• 00:10:15 When I enable all 3, it is as low as 7 gigabytes  of VRAM. And the base repository was like 24 plus  

• 00:10:22 gigabytes of VRAM. So you will see your keyframes.  You can also upload your keyframes if you want.  

• 00:10:27 What these keyframes do is that then it will turn  that into a video with Generate Video. It is like  

• 00:10:34 you are drawing it slowly. This was developed  by the legendary developer of the ControlNet,  

• 00:10:40 lllyasviel. You know him maybe if you are in  the generative AI projects. You can read more  

• 00:10:46 description here. Our application is many times  better than the original version with supporting,  

• 00:10:52 as I said, newest Triton attention, CUDA 13,  Torch 2.12. I needed to do a lot of programming  

• 00:10:59 to upgrade all of them. And the speed is amazing,  as you are seeing, compared to before. It is also  

• 00:11:05 showing peak VRAM and other stuff. So you see it  is also showing whichever the attention is being  

• 00:11:10 used. It has automatic trying of the attentions.  If 1 fails, it uses the next 1. But this  

• 00:11:16 attention is really, really fast. And also the  installers are working on RunPod, Massed Compute,  

• 00:11:21 SimplePod as well. So if you are a Linux user, you  can use the Massed Compute installer easily. Okay,  

• 00:11:27 we are almost done. So the entire thing is taking  like 100 seconds. I may even add torch compile if  

• 00:11:34 anyone requests. So it is like, you see, drawing  from scratch to it. It is like sketches to the  

• 00:11:40 final video like this. Okay, we are watching  it right now. So the keyframes are very crucial  

• 00:11:46 here because it is trying to match them. It is far  from perfect, and not every image will work. Also,  

• 00:11:52 if you get out of VRAM error, you can reduce the  resolution. This makes really significant impact  

• 00:11:57 on the speed and VRAM. It is like this. You cannot  set it any ratio as you want. It is like this. You  

• 00:12:03 need to use this to set your resolution. And  generate multiple keyframes, multiple generate  

• 00:12:09 video to get the perfect output that you want,  like you are drawing it. Oh, 1 more thing that  

• 00:12:14 I want to show. You know that we have Whisper  premium application which supports all Whisper  

• 00:12:19 models with extra features to transcribe videos.  If you are using this with the latest version,  

• 00:12:26 you see we have Whisper, Faster Whisper. This is  the really good quality. And if you get repeating  

• 00:12:31 output, sometimes because you are getting  repeating output, like repeating sentences, I  

• 00:12:36 recommend you to select model large version 1. And  the most crucial thing that you are going to make  

• 00:12:41 additionally to default settings is repetition  penalty. Set this like 1.2, and it should prevent  

• 00:12:47 your repeating sentences. You can try 1.1, 1.2,  1.3. As you increase it, it may skip some of the  

• 00:12:54 transcriptions. So making this very high is not  recommended. So slowly increase this until you do  

• 00:13:00 not get any repeating sentences. This way you  can generate amazing accurate transcriptions,  

• 00:13:06 amazing accurate subtitles. For example, the new  video that I have transcribed is looking perfect.  

• 00:13:11 I don't notice any errors at all. And the speed  was amazing. 29 minutes video was transcribed in  

• 00:13:19 1.5 minutes. You can see that, 29 minutes video  transcribed in 1.5 minutes. So the real-time  

• 00:13:26 speed was 20x with the highest quality options.  Thank you so much. Hopefully see you later.