Forget Suno: Run the Ultimate AI Music Studio LOCALLY (100% Free)

[FurkanGozukara]

Forget Suno: Run the Ultimate AI Music Studio LOCALLY (100% Free)

Full tutorial: https://www.youtube.com/watch?v=9C_6qNKjgpA

Full ACESTEP XL 1.5 Premium guide for local AI music generation, remix, repaint, stem extraction, audio processing, SAM Audio segmentation, Windows installation, RunPod, Massed Compute, SimplePod and Linux cloud workflows. This tutorial walks through the entire practical pipeline from first launch to final output management: generating fast songs, comparing Turbo/SFT/Base models, reusing prompts and seeds, remixing with reference audio, repainting selected sections, improving generated tracks, splitting vocals/drums/bass/other stems and adding instruments back with Lego mode. You will also see how to trim silence, export timelines for editing software, use SAM Audio with text prompts, process batches.

Essential links:

📥 App/latest zip: https://www.patreon.com/posts/ACESTEP-XL-Premium-SAM-Audio-157675060

▶️ Windows requirements guide: https://youtu.be/DrhUHnYfwC0

💬 Discord/help/community: https://discord.com/servers/software-engineering-courses-secourses-772774097734074388

Video Chapters:

00:00:00 Intro: ACESTEP XL 1.5 Premium local music, segmentation and processing tutorial

00:00:52 Fast song generation examples across styles in under one minute

00:01:55 Output manifest proof, 40-second generation time and supported models

00:02:29 Turbo/SFT/Base models, LoRA support, GPU presets and Torch Compile boost

00:03:10 Remix feature preview, same-lyrics requirement and responsible usage note

00:04:16 Repaint mode: regenerate and merge only a selected song section

00:05:38 Extract mode: stems, silence trimming, all-stems and batch folders

00:06:30 Lego mode: add an instrument stem such as guitar into existing audio

00:07:25 Audio Processing presets and manual enhancement controls for AI songs

00:08:35 Auto-Editor silent trim for tutorials, videos, audio and workflow export

00:09:48 DaVinci/Premiere/Final Cut/ShotCut/Kdenlive timeline export demo

00:11:01 SAM Audio Segment: BF16 models, VRAM presets and advanced segmentation

00:11:47 SAM outputs demo: vocals, drums, bass, remaining audio and saved files

00:12:47 Custom SAM prompts, semicolon batch segmenting and speech cleanup example

00:14:19 Batch processing, load metadata, manifests, saved settings and presets

00:15:09 Why local open-source models matter and where to run ACESTEP

00:15:55 Windows install begins: Patreon zip, changelog, attachments and download

00:16:53 Windows requirements tutorial before Python/CUDA/C++/FFmpeg setup

00:17:29 Extract zip safely, avoid bad paths and run Windows_Install_or_Update.bat

00:18:24 Automatic VENV, FFmpeg, UV install, model downloads and hash verification

00:19:24 Turbo default vs all-model download for SFT/Base and BF16 safetensors

00:20:32 First Windows launch, default Generate Song test and CMD progress

00:21:44 Model recommendations, VRAM tiers, languages, vocals and MP4 image output

00:23:29 Torch Compile setup for faster repeated generations

00:24:05 Outputs folder, model switching and full remix setup workflow

00:25:24 Practical remix loop: adapted lyrics, strength, reference audio and seed lock

00:28:03 Repaint workflow with source range preview, generated result and comparison

00:29:13 Recap: extraction, Lego, audio processing and SAM text-prompt usage

00:30:20 Windows wrap-up, LoRA training teaser and move to cloud installs

00:31:16 RunPod setup: credits, template, CUDA filters, GPU choice and storage

00:34:53 Upload zip in Jupyter Lab, extract, run instructions and handle installs

00:35:43 RunPod errors, resume behavior, model downloads and hash verification

00:38:04 Start ACESTEP on RunPod with Gradio Live, proxy ports and persistence

00:40:18 Add 7860/7861 ports, verify storage reuse and rerun installer after resume

00:42:10 RunPod connection troubleshooting and Gradio Live recommendation

00:44:12 Fix corrupted VENV/stale handle errors, reinstall safely and retest

00:47:24 Successful RunPod relaunch, default generation, nvitop and loading tips

00:49:26 RunPod first load vs fast inference, 15-second second generation example

00:51:02 Download outputs and delete RunPod pods/storage to stop spending

00:53:30 Massed Compute setup: coupon, Creator image, GPU prices and ThinLinc

00:57:13 Massed install from extracted folder, Linux notes and ultra-fast downloads

00:59:18 Start app on Massed Compute via localhost or Gradio Live

01:00:23 Default Massed generation, nvitop, faster loading and speed test

01:02:03 Sync/download outputs and delete Massed Compute instance safely

01:03:25 SimplePod setup: template, persistent volume, pricing and GPU choice

01:06:39 Jupyter upload, direct file browser, install command and model downloads

01:08:21 Start SimplePod, Gradio Live, default generation and one-time load errors

01:09:31 nvitop monitoring, newer driver/CUDA details and generation completion

01:10:42 Direct output/model downloads through SimplePod file browser

01:11:42 Delete instance, keep storage, relaunch GPU and verify install

01:13:15 Discord, subreddit, changelog, update guidance and support links

01:14:30 Final cleanup: terminate servers, delete storage and LoRA training outro

#ACESTEP #AIMusic #LocalAI #RunPod #MassedCompute #SimplePod #SAMAudio

Updated Full Written Tutorial Package (v5.3 Wildcards, DOCX, HTML, PDF and 20 Page PNG Guide)

I also prepared a polished written tutorial from the full ACESTEP XL 1.5 video and updated it with the new ACE-Step XL 1.5 Premium v5.3 Wildcards feature. It uses real screenshots from the actual application, includes the YouTube tutorial thumbnail/link context, and keeps the complete video chapters and transcript below this section.

Download / Open

• Full YouTube video tutorial: https://youtu.be/9C_6qNKjgpA
• Patreon app/latest zip and premium post: https://www.patreon.com/posts/ACESTEP-XL-Premium-SAM-Audio-157675060
• Written DOCX tutorial: https://github.com/FurkanGozukara/Stable-Diffusion/raw/main/Tutorials/assets/forget-suno-ace-step-xl-15-written-tutorial/ACESTEP_XL_15_Written_Tutorial.docx
• Written HTML tutorial: https://github.com/FurkanGozukara/Stable-Diffusion/raw/main/Tutorials/assets/forget-suno-ace-step-xl-15-written-tutorial/ACESTEP_XL_15_Written_Tutorial.html
• Compact 20-page PDF tutorial: https://github.com/FurkanGozukara/Stable-Diffusion/raw/main/Tutorials/assets/forget-suno-ace-step-xl-15-written-tutorial/ACESTEP_XL_15_Compact_20_Page_Tutorial.pdf
• Compact 20-page HTML tutorial: https://github.com/FurkanGozukara/Stable-Diffusion/raw/main/Tutorials/assets/forget-suno-ace-step-xl-15-written-tutorial/ACESTEP_XL_15_Compact_20_Page_Tutorial.html
• All high-resolution tutorial PNG pages: https://github.com/FurkanGozukara/Stable-Diffusion/tree/main/Tutorials/assets/forget-suno-ace-step-xl-15-written-tutorial

What The Written Tutorial Covers

• Windows setup: downloading the Patreon zip, extracting safely, running Windows_Install_or_Update.bat, downloading models, and launching with Windows_Start_App.bat.
• First generation workflow: prompt, lyrics, model choice, duration, guidance/steps, seeds, output folders, metadata, and reusable settings.
• Wildcards: use [option A|option B|option C] in Style, Advanced Music Caption, or Lyrics so one option is picked at generation time; nested wildcards are supported.
• Batch folder processing uses the same Wildcards behavior: batch jobs can vary instruments, moods, hooks, or lyric phrases across outputs without manually editing every run.
• Wildcards caveat: keep Auto/Enhance Style and Auto/Enhance Lyrics disabled when you need exact wildcard expressions preserved, because those improvement tools may rewrite the prompt text.
• Model and speed choices: Turbo, SFT, Base, BF16, VRAM presets, Torch Compile, and when each choice matters.
• Creative tools: Remix, Repaint, Extract, Lego mode, Audio Processing, Auto-Editor timeline export, and SAM Audio text-prompt segmentation.
• Output management: MP3/WAV/FLAC/MP4 results, manifests, stems, trimmed files, SAM segment folders, and batch processing.
• Cloud workflows: RunPod, Massed Compute, SimplePod, Jupyter uploads, Gradio Live links, persistent storage, restart behavior, nvitop monitoring, downloading outputs, and deleting rented resources.
• Practical troubleshooting: bad extraction paths, missing requirements, first-load delays, stale/corrupted VENV handles, connection issues, model verification, and safe reinstall/update behavior.

v5.3 Wildcards Screenshots

Click either screenshot to open the full-resolution PNG. Embedded previews use height="700" with no fixed width.

20 Page Visual Tutorial

Click any page image to open the full-resolution PNG. Embedded previews are intentionally shown at height="700" with no fixed width, so GitHub keeps the page readable while still allowing full-size inspection.

Video Transcription

• 00:00:00 Greetings everyone. Today I am going to show you  how to install and use the ultimate local music  

• 00:00:09 generation application. This application, ACESTEP  XL 1.5, is a combination of applications that not  

• 00:00:16 only generates music but also supports SAM Audio  Segmenting and Audio Processing. I will not only  

• 00:00:24 show how to install and use it on your Windows  computer but also on RunPod, on Massed Compute,  

• 00:00:32 and on SimplePod. I will show it on all 4  different platforms; Linux users can use  

• 00:00:39 the Massed Compute installation procedures and  scripts. So what is this application about? This  

• 00:00:46 is a music generation application mainly, but  with so many other features. So let's see them  

• 00:00:52 quickly. You can generate widely different style  musics under 1 minute. Let me play some of them:  

• 00:01:04 "Hands in the dirt. Put the work on track.  Built from the ground, never life on lack.  

• 00:01:26 Woke up with the fire with the weight  on my back. No gold on my wrist till  

• 00:01:31 I'm dressed.". "Woke up with the fire.  Put the weight on my back. No gold on my  

• 00:01:39 wrist till I'm dressed in fact. Hands  in the dirt. Put the work on track.  

• 00:01:53 Woke up with the fire.". So, as you have seen,  I have generated several different style music.  

• 00:01:59 Every one of them is amazing and they only took  like 50 seconds. Let me show you. When we open  

• 00:02:05 the outputs folder, we can go to generation and  we can open the generation manifest file. It  

• 00:02:12 will show us how much time it took; it took  only 40 seconds to generate this full song  

• 00:02:20 which is 3 minutes 11 seconds. How? Because this  application is very professionally developed. It  

• 00:02:28 supports all 3 models: ACESTEP XL 1.5 SFT, Turbo,  and Base models. It supports full LoRA training  

• 00:02:37 and LoRAs, though that will be another tutorial,  not this one. We have GPU optimization presets  

• 00:02:42 automatically selecting; therefore, it is fully  optimized for my GPU. And most importantly, when  

• 00:02:49 you go to ACESTEP Advanced, you will see that we  have a compile model. Therefore, it speeds up the  

• 00:02:57 generation speed up to 50%. This is my addition  to the public repositories that you can find about  

• 00:03:03 this application. But believe me, this ACESTEP XL  1.5 premium is nothing like you will find anywhere  

• 00:03:10 else. So what other features it has? It can remix  full songs amazingly. For remix to work, you need  

• 00:03:17 to have the same lyrics. By the way, before I  start showing you all these features, I need to  

• 00:03:24 tell you that this application needs to be used  respectively, and this tutorial is for research  

• 00:03:31 and education purposes. Okay, first I will play  from the original then I will play from the remix.  

• 00:03:41 beauty queen from a movie scene. So one of the  best part of this application that it is really  

• 00:04:01 fast. Therefore, you can generate many iterations  and get what you want exactly. So, it is so easy.  

• 00:04:09 I will show how to use, install, and everything.  This was the remix feature. Let's see the repaint  

• 00:04:16 feature. Now, the repaint is like a remix, but you  generate only a certain part of the song. And to  

• 00:04:24 be useful, we need actually LoRAs. When we train  LoRAs with our vocals and with our audio—you know,  

• 00:04:32 with our own speaking—it will be perfect. For now,  it can be still used to generate fully new parts;  

• 00:04:40 however, it won't be very consistent for lyrics  generation. For instrumental regenerations it can  

• 00:04:47 work, but let me show you for now. And as I said,  it will be much more useful hopefully with LoRA  

• 00:04:53 training and with vocals training. I'm working on  that, so first I will show the original area then  

• 00:04:58 I will show the repainted area. I don't need the  noise. I don't. So this way I can totally generate  

• 00:05:12 new sections of the song. However, currently,  it is not consistent, so it is looking like  

• 00:05:18 this. Let me show you: we have selected the  parts between 19 seconds to 23 seconds. Just  

• 00:05:30 as you have seen, it has successfully repainted  the selected area and merged it like this. So the  

• 00:05:38 next feature is extraction. This is extremely  useful to extract stems from any given song  

• 00:05:45 or audio file. The normal extraction extracts  everything like this—the full song—or you can  

• 00:05:51 use our Auto-Editor trim output and just get the  extracted section and trim out the silence. Let me  

• 00:05:58 play some of them to show you. So you can extract  vocals, woodwinds, brass, fx, and everything that  

• 00:06:09 this ACESTEP XL model supports. You can  extract all stems at once. This is very useful,  

• 00:06:15 or you can do batch folder processing to extract  from multiple given files. Everything is working  

• 00:06:21 perfectly. To change the configuration and  features of the auto editor trim output, just  

• 00:06:26 change the settings set here and it will follow.  So the next feature is Lego. Lego is adding a new  

• 00:06:35 stem to the existing audio. So let me demonstrate  to you how it works: first I will show the  

• 00:06:41 original then I will show the Lego painted area.  So in this example, I have added a guitar stem to  

• 00:06:57 the track. From here you can select the stem that  you want to add and it will add that stem. It will  

• 00:07:03 paint that stem into your audio. It can be a full  song, it can be just only vocal, whatever. And it  

• 00:07:09 will generate and add it properly. It will follow  the music caption; however, not the lyrics. For  

• 00:07:16 lyrics to be valid, you need to use vocals, but  you may not get the performance you want. You need  

• 00:07:21 to test it; however, it is working amazing at the  moment. The next feature is the audio processing  

• 00:07:28 tab that we have implemented. So this audio  processing tab has several different features.  

• 00:07:34 The default feature is processing presets which  changes the audio enhancements that you will  

• 00:07:41 see here. This is very useful to improve your AI  generated songs. You can read their descriptions,  

• 00:07:47 you can play with their values, or you can use the  presets from here and it will update the values  

• 00:07:53 you will see here. And when you click the process  file, it will generate the processed audio.  

• 00:08:06 with the weight on my back. Since the  original song is already pretty good,  

• 00:08:21 you may not notice the difference, but if you max  all of these, you will notice the difference. So,  

• 00:08:26 play with it and see if it is really improving the  quality for you. This is a very useful tool that I  

• 00:08:32 have implemented upon the request of someone.  So the next feature, which I like very much,  

• 00:08:38 is auto editor trim silent. This is extremely  useful. I use this daily in my tutorial videos  

• 00:08:45 to trim out unspoken parts. You know when  I am thinking something, I am not speaking;  

• 00:08:51 therefore, that section is silent and I want to  trim that part. So I have a video tutorial here,  

• 00:09:00 one of the last tutorials that I have recorded;  it is 20 minutes 31 seconds because it has a  

• 00:09:05 lot of pauses. Then I process this file and it  becomes 16 minutes. So it trimmed out 4 minutes of  

• 00:09:14 silent parts or maybe when I am taking breaths or  some other stuff where I wasn't talking properly.  

• 00:09:23 This is extremely useful. You can use this  feature: upload a video and get trimmed output.  

• 00:09:29 You can set the re-encode values from here, but  it will automatically try to match the original  

• 00:09:35 video and you don't even need this; you can export  only audio or you can export as a workflow which  

• 00:09:44 I am using normally. Davinci Resolve Export. For  DaVinci Resolve export, select processing preset  

• 00:09:50 'none' if you don't want to process the audio,  obviously enable 'auto editor trim silent',  

• 00:09:55 and select the editor workflow, then use 'browse  local file' so that it will generate accurate file  

• 00:10:03 paths. Then click 'process file'. At the end, you  will get a file that you can import into DaVinci  

• 00:10:10 Resolve like this from here. Let me show you.  Moreover, everything you have generated in our  

• 00:10:16 application will be saved. When you click to open  outputs folder, you can see the SAM audio process,  

• 00:10:22 audio processing results all here. They are  all saved. And in DaVinci Resolve, I will go  

• 00:10:28 here and I will click timelines, import this file,  select the file and okay. And it will fully import  

• 00:10:35 accurately with accurate file path references  and resolution, whatever I want. And you can see  

• 00:10:41 that it is properly split like this. So I can use  the DaVinci Resolve as I want with this timeline.  

• 00:10:48 This auto editor workflow supports DaVinci  Resolve, Adobe Premiere Pro, Final Cut Pro,  

• 00:10:54 ShotCut, and Kdenlive. It is extremely useful;  I use this at my every tutorial. It is super  

• 00:11:00 useful. The next feature is SAM audio segment. SAM  audio segment is Facebook Meta's published model  

• 00:11:08 and it is a state-of-the-art audio segmentation  model right now. I have implemented SAM audio  

• 00:11:14 with all of the extra features and optimizations.  The downloaded models are all BF16 by me. We have  

• 00:11:22 VRAM presets automatically set according to your  GPU. This is a very heavy model. It is processing  

• 00:11:29 with segmentation. You see, we support as low as  10 GB GPUs and even 8 GB GPUs as well. Moreover,  

• 00:11:36 it can extract any audio arbitrarily and it  supports quick prompts. You can select multiple  

• 00:11:42 quick prompts from here as well to extract  multiple values at once. So, let me show you  

• 00:11:47 some of the results. It also supports the audio  editor trim output and compile model to improve  

• 00:11:54 the speed. So, this is segmentation output, this  is vocals output, and also drums and bass. It will  

• 00:12:02 only show the first one; the other ones will be  saved inside the outputs folder in the SAM audio  

• 00:12:09 segmentation. You see with suffixes of drums,  guitar, or vocals, whatever you want to extract.  

• 00:12:19 So it is like this; let me also play some of the  others from here. As you can see, it is amazing.  

• 00:12:35 You can also play the remaining audio; it also  saves that. This model is very, very powerful.  

• 00:12:40 It's a heavy model compared to ACESTEP extraction;  however, this model is more powerful and it also  

• 00:12:47 supports, as I said, custom prompts. For example,  for custom prompts, you can enable batch segment  

• 00:12:53 as well with semicolon separation. You can write  the stems or segments that you want to extract  

• 00:12:59 and it will iteratively process every one of  them and save them in the outputs folder. It  

• 00:13:04 is working amazing. This was a feature that  you have requested, so I added it. Moreover,  

• 00:13:09 you can use this SAM audio segmentation for  other tasks as well. Let me demonstrate to you.  

• 00:13:14 This is an everything video: "Hey, hang on.  Can you hear me? Super loud. You know what?  

• 00:13:21 You know what actually looks better? You will  not believe the day that I have had. Seriously,  

• 00:13:26 no way. I can't wait to tell you all about it when  I get home. I'll be over in like...". Now imagine  

• 00:13:31 that I can clean this very well and extract only  the speech. To extract only the speech, all I need  

• 00:13:39 to do is select the quick prompt 'speech' from  here. I also enabled auto editor trim to trim out  

• 00:13:46 the silent parts, and let me play you the cleaned  audio. This is amazing. This is magnificent:  

• 00:13:52 "Hey, hang on. Can you hear me? Super loud. You  know what? Do you know what? Actually, what's  

• 00:13:57 better? You will not get the day that I arrived.  Seriously, no way. I can't wait to tell you all  

• 00:14:03 about it when I get home. I'll be over in like 10  minutes.". As you have seen, it is just amazing,  

• 00:14:08 amazing—the extracted cleaned audio. So, we have  all the optimizations, quantization, attention,  

• 00:14:15 whatever you want. We support so many features  and you can use the SAM audio. It also fully  

• 00:14:21 supports batch folder processing so that you can  process so many files at once in a given folder  

• 00:14:28 and extract multiple stems, extract multiple,  you know, segments, whatever you want. Another  

• 00:14:34 feature is load metadata. I save all the metadata  of all generations in the outputs folder. You see  

• 00:14:41 everything is properly saved like this and you  can select the manifest and it will load all the  

• 00:14:47 settings that were used to generate this song or  audio processing or segmentation, whatever. You  

• 00:14:53 can also have full custom presets; it is fully  working. You can save every setting, load them,  

• 00:15:00 you can delete your preset, so whatever you want  to do, and they will be also remembered when you  

• 00:15:06 next time restart your application, very useful.  This local model is now more important than ever  

• 00:15:13 if you know the recent incident that Anthropic  stopped serving its very best model. So anytime if  

• 00:15:23 you are dependent on the cloud providers' cloud  services, they may stop serving you. However,  

• 00:15:29 this model is local. You can use it as much as you  want. It is fully commercially usable. Therefore,  

• 00:15:35 this is the way of the future: having local  models, open-source models, using them on your  

• 00:15:40 local computer. If your local computer is not  powerful enough, then you can use it on private  

• 00:15:46 cloud services, private GPU providing services  like RunPod, like Massed Compute or SimplePod.  

• 00:15:52 I will show all of them. Since we have seen all  the features, now we can begin the installation  

• 00:15:59 and the setup, using it on Windows first then the  cloud services. So you need to go to this link;  

• 00:16:06 this link will be in the description of the  video below and also in the comment section  

• 00:16:11 of the video. Download the latest installer zip  file. It will be here or at the bottom when you  

• 00:16:17 scroll down in the attachments you will see it.  You see this post is very long. Why? Because I  

• 00:16:23 write every update as a change log here. If you  read everything here, then you will learn how this  

• 00:16:30 app is developed—the timeline, what new features  were added, how it is working. This is very, very  

• 00:16:37 useful to read, so I recommend you to read it.  And at the very bottom, let me show you—yes, it  

• 00:16:42 is very long, I agree with that—you will see the  attachment. So you can also download from here.  

• 00:16:48 Let's go to top again, click download. It will  download the zip file, but do not start installing  

• 00:16:55 yet. First of all, you need to follow the Windows  requirements tutorial. If you haven't followed it  

• 00:17:01 yet, do so; if you have followed it previously,  then you are ready. So this is the tutorial that  

• 00:17:07 you need to follow first. Its source is fully up  to date, the tutorial is fully up to date. You  

• 00:17:13 can still skip watching this tutorial and try to  make it work. It may work depending on your setup;  

• 00:17:18 however, if you want to use torch compile  especially, you need to follow that tutorial. Then  

• 00:17:24 move the downloaded zip file into any disk where  you want to install. So let's install it into  

• 00:17:31 our Q drive and extract the file like this and  enter inside the extracted folder. Make sure that  

• 00:17:37 your folder path doesn't have space characters or  non-English characters; if you don't want to have  

• 00:17:44 any issues, I recommend that. Then double-click  Windows install or update.bat file and run. This  

• 00:17:50 will start the installation. It will generate  a Python 3.11 virtual environment, so you need  

• 00:17:56 to have Python 3.11. It is also installing local  shared FFmpeg runtime. So the installation will  

• 00:18:05 be fully automatic as long as you have followed  the Windows requirements tutorial. It will also  

• 00:18:10 automatically download all the models that you  need and everything will be inside this folder and  

• 00:18:17 inside the virtual environment that it is going  to generate in a moment. First, it is downloading  

• 00:18:23 the necessary shared FFmpeg file so that you  won't have any issues. Okay, virtual environment  

• 00:18:29 generated and the FFmpeg are also downloaded like  this: ffmpeg shared and ffmpeg shared download.  

• 00:18:36 The installation will be super fast because we are  using UV installation, not depending on anything  

• 00:18:41 else. So the installation already completed.  You can verify the installation with scroll up,  

• 00:18:46 and it will start downloading the models  automatically into accurate folders with accurate  

• 00:18:51 names. My downloader has 16 connection download  and it also does hash verification; therefore,  

• 00:18:59 all the downloaded models are 100% accurate. They  can never be corrupted. It fully supports resume  

• 00:19:06 as well. When you run the Windows install again  or for update again, it will verify the status and  

• 00:19:12 continue wherever it was left. This will download  the SAM audio segment model, audio processing  

• 00:19:18 related models, and it will also download the  ACESTEP XL 1.5 turbo model. If you also want  

• 00:19:25 to use ACESTEP XL 1.5 SFT and base model, then  once this installation has been completed, you  

• 00:19:32 need to run the Windows download all models.bat  file. I will show that after this download is  

• 00:19:39 completed. You can also see the speed; it is  100 megabytes per second, which is my maximum  

• 00:19:45 internet connection speed. Normally shared models  were FP32; however, we are not doing inference at  

• 00:19:52 FP32. Therefore, I have generated BF16 versions  which is our inference precision. And so you see  

• 00:20:00 the models are now half size—faster to download,  not faster to load. It uses lesser RAM memory;  

• 00:20:06 everything is much better right now. You see the  ACESTEP model is just a single safetensors file. I  

• 00:20:13 also converted the Pytorch PT files into secure  safetensors files. Okay, so all the 57 models  

• 00:20:21 downloaded, everything is accurately downloaded,  hash verified. If you wonder how much space the  

• 00:20:27 models are taking, it is currently 34 GB which is  very reasonable. Let's start the application with  

• 00:20:34 Windows startup.bat file and while generating some  songs I will also download the remaining models.  

• 00:20:42 So I will run the Windows download all models; for  cloud installation it will automatically download  

• 00:20:47 these remaining models, so you won't be needing  that since they are fast. So the new installed  

• 00:20:54 application started. This is an easy-to-use  interface which I have developed. First,  

• 00:20:58 you can hit generate song to verify everything is  working without changing anything else. This has  

• 00:21:05 a default style prompt to give you an idea  and lyrics. In the zip file, you will also  

• 00:21:11 find a file named as ACESTEP lyric generation  instructions for LLMs. You can use this file:  

• 00:21:19 give it to your favorite LLM like ChatGPT and make  it format your lyrics or your style accordingly.  

• 00:21:27 These ACESTEP XL models really work well to format  the lyrics or the style. So you can always watch  

• 00:21:35 the status in the CMD window rather than depending  on the Gradio interface itself which I recommend,  

• 00:21:41 and the remaining models are also being downloaded  right now which are ACESTEP XL SFT and base  

• 00:21:48 models. When you go to the Advanced tab you will  see which models are recommended and supported  

• 00:21:54 for which task. So the turbo model is mostly for  generation only because it is fast, high quality,  

• 00:22:02 and the other models are for other tasks. It  will automatically select the GPU optimization  

• 00:22:07 preset. If it be too slow, you can reduce your  tier to generate fast. You can also provide an  

• 00:22:13 image from here and it will generate an MP4 file  with the song it generated and with the image you  

• 00:22:21 have provided. It is automatically generating  with the selected resolution and it is keeping  

• 00:22:27 the original aspect ratio. You can also set  your song instrumental from here, change the  

• 00:22:32 vocal from here, or change the language from here.  You see, it supports so many different languages.  

• 00:22:37 This model is currently state-of-the-art. It is  certainly better than the Suno free version. It  

• 00:22:43 is rivaling the Suno 5, Suno 5.5, especially with  LoRA training. Hopefully, it will be the next  

• 00:22:48 tutorial. It will be better than even the paid  Suno 5.5. And you know, Suno is very likely to be  

• 00:22:55 putting some watermark into your generations;  therefore, platforms like YouTube know that  

• 00:23:02 song is AI generated. Okay, we can listen to the  default song. Woke up with the track. Woke up with  

• 00:23:17 the five. Put the weight on my back. I'm blessed  in fact. Okay, it is working amazing as expected.  

• 00:23:27 If you are going to generate multiple songs and  if you want to speed up, go to advanced setup and  

• 00:23:31 enable compile model. For this to work, you need  to have CUDA installed properly and also MSVC C++  

• 00:23:40 compiler to be installed properly. Everything is  explained in the Windows requirements tutorial.  

• 00:23:44 This application will find accurate CUDA and the  installed C++ compiler automatically and use it  

• 00:23:51 to torch compile. The first run with torch compile  will be very slow since it will compile, but the  

• 00:23:58 subsequent ones will be very fast, and it may  need to recompile depending on the settings  

• 00:24:03 that you can change. So this is how you generate  songs. Everything will be saved inside the outputs  

• 00:24:10 folder. When you open the outputs folder, you will  see that it is saved like this. You can change the  

• 00:24:16 models from here. You can go to advanced, select  remix. For selecting remix, I recommend you to  

• 00:24:23 change the recommended model to SFT and upload  your generated song or whatever song that you want  

• 00:24:29 to remix, and use the same lyrics here. Change the  music caption like rap or whatever you want, and  

• 00:24:36 then you can also select the remix source start  and end. It shows the preview here and generate  

• 00:24:43 music. By the way, if you get an out of VRAM  error, you can restart the application. Oh, you  

• 00:24:48 see we have just downloaded the model, it didn't  see it. So, I need to restart the application. So,  

• 00:24:54 as I said, if you get an out of VRAM error, I  recommend you to restart the application. However,  

• 00:24:59 I made a lot of improvements so it shouldn't be  necessary; it should automatically switch between  

• 00:25:05 models without restarting your application. So  select SFT, go to advanced, select the remix,  

• 00:25:13 and have the same lyrics as the remix song to  not have any lyrics issue. Change the caption,  

• 00:25:20 then hit generate music and it will generate  the remix. Okay, as you have seen the remix  

• 00:25:27 results were not great. However, one of our users  just messaged me and explained to me how he makes  

• 00:25:35 amazing remixes. So I'm now going to show you how  he makes them. He made a production loop. It's a  

• 00:25:43 loop; therefore, you need to repeat it. First of  all, we are uploading our original audio. Then  

• 00:25:49 write an adapted lyric, not a literal translation.  Keep line length, syllable count, vowel shape, and  

• 00:25:56 stressed syllables close to the original phrase.  Replace meaning with sound compatible wording  

• 00:26:02 when needed. Set remix strength between 0.92 and  0.99; start in the middle of the range. Listen,  

• 00:26:10 then move tighter or looser depending on whether  the vocal follows the original too much or not  

• 00:26:16 enough. How you set it? If you pay attention to  the remix strength, it is here. So, by default,  

• 00:26:21 it is 1. However, as recommended, try with 0.92  to 0.99 to keep the wording more accurate. Add  

• 00:26:30 a clean 30 second voice reference. Now, this is  important. This is something that I didn't test,  

• 00:26:36 but he tested and figured that out: use it for  timbre vocal character. Keep it clean, dry,  

• 00:26:42 and center it on the voice you want to be copied.  So, where you are going to edit? Obviously, it is  

• 00:26:47 here. So, provide the reference audio here to make  the remixed audio even closer and better to the  

• 00:26:55 original song. And the next step, generate until  the base take feels right. So as I have shown,  

• 00:27:01 the generation is fast. So generate again and  again and again until you get a good base result.  

• 00:27:08 Then lock the seed and disable random. How you are  going to do that? I have updated the application  

• 00:27:15 and moved the seed to this place. So it is  much easier now and it will automatically  

• 00:27:22 set the last generated seed value. So once you  have the accurate and reasonably good remix,  

• 00:27:30 just uncheck this and keep working on the same  seed value. Moreover, I did set the default remix  

• 00:27:38 strength to 0.95, so with the latest update,  you are going to have both of the features.  

• 00:27:45 Then change one word or phrase at a time and keep  working until you get the desired result. He is  

• 00:27:52 also doing some additional stuff like separating  stems, tuning vocals, then mixing all; it is  

• 00:27:58 all up to you to the level that you want to go.  For repaints, same way: upload to source audio,  

• 00:28:04 set your caption and lyrics, and select the  repaint start and end like 19 to 24. It will  

• 00:28:11 show the preview here. However, currently,  we are generating remix; therefore, it is  

• 00:28:15 waiting for that, so we need to first wait for the  remix to be completed. Okay, change it to preview.  

• 00:28:21 You can watch the status in the CMD window. As  you are seeing, I am not changing any optional  

• 00:28:27 parameters or advanced settings because they are  all set to maximum quality with the VRAM presets  

• 00:28:34 we have. As you play with this application, you  will understand how to use it, how useful it is,  

• 00:28:39 how advanced it is. It has so many amazing  features. So, our generation is almost  

• 00:28:43 completed. Yes, it is completed. The result should  appear in a moment. And it is here. So, the latest  

• 00:28:50 generated result is here. The latest remixed area  is here. You can also see the original inputs.  

• 00:29:03 Yeah, this didn't work very well, so I probably  need to change my caption or generate more. This  

• 00:29:08 is just a very basic caption, but this is how it  works. When I go to repaint, it's the same way.  

• 00:29:13 For extracting, I need to change the base model  to this one. The base, the Lego, and extraction  

• 00:29:20 will become available. I have already shown how to  use them. And as you have noticed, everything has  

• 00:29:26 very good explanations on the Gradio interface,  so read every one of them to understand. I also  

• 00:29:31 can use audio processing: just upload the file and  click the process file, or I can use the auto-trim  

• 00:29:39 editor. We already saw them. SAM audio is same  way: just select your file and select the options  

• 00:29:46 whichever you want to do like text prompt. I  never tried the span actually, so I can't say  

• 00:29:52 that it is working as expected and it is very hard  to use, or the visual mask. I also never used it,  

• 00:29:58 but the text is working perfect. These two others  are very hard to use; they require a lot of manual  

• 00:30:03 work. However, text is working perfect. Type your  custom prompt whatever you want to extract as I  

• 00:30:09 have shown in the beginning of the tutorial and  it will work. My installer has properly compiled  

• 00:30:16 Flash Attention; therefore, it should work on your  Windows or on cloud services. So, this was how to  

• 00:30:24 do inference on this application on Windows. The  LoRA training is also fully working and it will  

• 00:30:31 come as a next tutorial. Hopefully, I am still  in research; I will try to train a vocal voice  

• 00:30:37 and generate with that voice consistently and  with the style, of course—the singing style. It  

• 00:30:42 will be an amazing, epic tutorial hopefully. So  now I will show how to install and use on RunPod,  

• 00:30:48 on Massed Compute, and on SimplePod. If you  don't have a powerful GPU, if you want to  

• 00:30:53 generate faster, if you want to scale up your  generation, then you need some cloud services  

• 00:30:58 and these are the best services. So in the zip  file, we already have all the instructions.  

• 00:31:03 Let's begin with the RunPod. You see we have  a RunPod SimplePod ACESTEP instructions.txt  

• 00:31:09 file. They are all in the zip file; you need to  first download it and extract it. When I open it,  

• 00:31:14 it will show me all the instructions. So  let's begin with the RunPod. First of all,  

• 00:31:19 please use this link to register; I appreciate  that very much. Once you register and sign in,  

• 00:31:24 click this plus icon and add some credits. You see  I am currently spending some money because I have  

• 00:31:31 some storage. Okay, let's delete them because I  will show you how to use the storage as well. Man,  

• 00:31:38 if you forget this, you will spend money like me.  This is the homepage, the newest homepage of the  

• 00:31:44 RunPod. You registered, you added some credits,  then as a next step, you will see that we have the  

• 00:31:49 template. All of my application installers work  the same way; so once you understand the logic,  

• 00:31:56 you will be able to use every one of them. So  click this; it will open the accurate template  

• 00:32:03 and select it. This may change depending on  the application. Since we are using CUDA 13,  

• 00:32:08 there is one important thing which is go to  filter and select the CUDA version as 12.8, 12.9,  

• 00:32:17 and 13. Why? Because RunPod is not updating its  NVIDIA drivers. The issue is the NVIDIA driver,  

• 00:32:24 not the CUDA version. Because we can install the  CUDA version, but we cannot install the NVIDIA  

• 00:32:29 driver. This way we are going to get a machine  that has the accurate NVIDIA driver that will  

• 00:32:35 run the CUDA 13. After that, you can also play  with the other filters like secure / community—I  

• 00:32:41 recommend secure—the RAM amount, disk type,  whatever, and then apply filters. It will  

• 00:32:47 select it. Now if you don't select any permanent  storage, it will create a temporary storage that  

• 00:32:54 you will use on this instance. If you want to  use permanent storage which you can resume later,  

• 00:32:59 you need to add a volume disk. So you see  currently we don't have a network here;  

• 00:33:04 therefore, I will go to storage first and create  a network volume in here. You need to select a  

• 00:33:09 region because it's region-specific and depending  on the GPU that you are going to use, you need to  

• 00:33:14 select your region accurately. So I am going to  select a region which has a lot of RTX Pro 6000,  

• 00:33:21 which is one of my favorite GPUs, or RTX 5090.  For this one, it would work too. Okay, let's see  

• 00:33:28 what we have here—high performance. The US regions  are usually better. So this is a Europe region but  

• 00:33:35 it has a lot of RTX Pro 6000. It will be slow  but I need to select this. I hope it doesn't  

• 00:33:41 be too slow. Let's say example, I will make this  200 GB and create network volume. Now my volume  

• 00:33:47 is ready. Therefore, I will click this again to  get the accurate template. Template is selected,  

• 00:33:54 my filters are kept, very nice. And now I am  going to select my GPU. Wow, the prices are all  

• 00:34:00 increased because of the demand, and persistent  storage is network volume. Okay, now I have it,  

• 00:34:07 but it doesn't show whether it has selected it or  not. It says automatically create; oh, here. Yes,  

• 00:34:12 from here I can select my persistent storage like  this. So don't make it automatically create. But  

• 00:34:19 I see that we could create from this as well,  so the interface keeps changing. Pay attention  

• 00:34:24 to the interface. The most crucial part is  using the accurate template and deploy port  

• 00:34:30 and also selecting the accurate filter for  NVIDIA driver CUDA 12.8, 12.9, or 13. Okay,  

• 00:34:38 this template was already used, very nice, so  we didn't wait. When you go to the details or  

• 00:34:45 telemetry it will show the driver; yes, this  is the driver version. This driver supports 13,  

• 00:34:51 therefore it should work very well. So go to  connect and click Jupyter Lab. If you don't  

• 00:34:55 see this is enabled, you can refresh this page.  The Jupyter Lab is starting. Then I will drag  

• 00:35:01 and drop the downloaded zip file into here like  this. You can also use this upload icon to upload  

• 00:35:10 your zip file. Once the upload is completed, it  will show the upload status here. Right-click and  

• 00:35:15 extract archive, then it will extract like this.  Then there is RunPod SimplePod instructions and  

• 00:35:22 there is this installation command. Always read  these instruction files. Open a new terminal  

• 00:35:27 from this plus, copy-paste it, and it will do all  the installation and model downloads. This will  

• 00:35:33 download all of the models, not only the turbo  model, because this is a cloud service; therefore,  

• 00:35:38 it is fast. All you need to do is now wait,  depending on your port. It may take a lot of time  

• 00:35:44 or it may be fast. However, RunPod may be broken  at many times; so if you get any error, just get  

• 00:35:50 a new GPU. Unfortunately, there is no other way.  Currently, this one is looking like it's working,  

• 00:35:56 but you can never be sure, and it is already slow.  So the installation started and it is installing  

• 00:36:02 the libraries. We are using UV installation and  even with UV, RunPod is unfortunately slow. So  

• 00:36:10 depending on your chance, the server, and the GPU  you get, you may wait a lot of time or it may be  

• 00:36:17 fast. During the installation, unfortunately, an  error occurred. This is an operating system error;  

• 00:36:24 that means this is a RunPod error. To fix this  issue, run the installer again and it will try to  

• 00:36:32 resume from wherever it was left, so it should be  faster when you run it the second time. As I said,  

• 00:36:39 RunPod is very erratic. It may throw errors, not  all GPUs work, so it is totally unpredictable in  

• 00:36:50 performance. However, it is the most widely  used one and it has so many different GPUs;  

• 00:36:56 therefore, this is like a trade-off between some  features and some errors or unpredictability. You  

• 00:37:05 see the second time running installation is much  faster. I hope this time it won't have any issues;  

• 00:37:11 we should see. But this operating system error  is 100% related to RunPod's shared storage and  

• 00:37:19 network system itself. Now it is starting  to download all of the models. You don't  

• 00:37:24 need to run an additional command. If you get  any errors during the model download as well,  

• 00:37:29 just run the installation again; it just resumes,  it doesn't start from the beginning. The model  

• 00:37:34 download speed is looking decent. Since I am  using 16 connection downloads, it is merging them;  

• 00:37:41 therefore, the downloads are really fast and  optimized. It's also verifying the hash values  

• 00:37:46 so that you will never have corrupt model issues,  because corrupt models are very annoying where  

• 00:37:52 everything looks normal but it doesn't work or  it produces inaccurate results. Yes, the speed  

• 00:37:58 is also decent, equal to 100 megabytes per second.  All right, so the installation has been completed.  

• 00:38:03 All the models have been downloaded. Now we  are ready to start launching the application.  

• 00:38:10 For launching, return back to the RunPod SimplePod  instructions.txt file. All of my applications have  

• 00:38:17 that. Copy the start command, open a new terminal,  and paste the start command. This will start as a  

• 00:38:25 Gradio live shared link. The starting may take  time depending on your port speed. Okay, so the  

• 00:38:32 application started. We can see that Gradio live  is here. If you don't want to use Gradio live,  

• 00:38:39 you need to add a port to here; it will restart  your port, so be careful with that. So you need  

• 00:38:46 to edit this port and add a port here to connect  from the proxy of the RunPod. I will show that  

• 00:38:54 after this. Okay, so the application started. Now  it is the same as using in Windows; I'm not going  

• 00:39:01 to repeat. However, I'm going to show you how  you can resume using without making everything  

• 00:39:07 from your persistent storage permanent storage.  To resume it, I'm going to delete this port. You  

• 00:39:13 see when I stop this port, it will be like this,  and since I was using permanent storage, I will  

• 00:39:19 be able to resume it. Let's verify that. Okay,  it doesn't show, so I'm not sure if it started  

• 00:39:26 with my permanent storage now. Okay, it says that  yes, it was using this, so it should start. Okay,  

• 00:39:32 I'm going to terminate now. You know, when you  terminate it, everything will be deleted unless it  

• 00:39:37 is in your permanent storage, so let's terminate  the port. And our template is still selected,  

• 00:39:43 but let's begin from the beginning to verify:  so open the instructions.txt and double-click  

• 00:39:49 the template. You see it is selected, then as  a filter, I'm going to select the disk and I'm  

• 00:39:56 going to make the CUDA filter. I also recommend  selecting maximum RAM and this disk. Okay, apply  

• 00:40:04 filters. So I'm also going to select my network  volume. You see when I select the network volume,  

• 00:40:10 the options will be lesser because it will  filter based on my network storage region. Okay,  

• 00:40:18 selected; everything is ready. But before I start,  I'm also going to add the port; the port was 7860.  

• 00:40:27 So let's also add it. However, am I going to  add the port? So for adding the port, I'm going  

• 00:40:33 to click set overrides and add it here: 7860 like  this and set overrides. Even though you are using  

• 00:40:45 my template, you can add some stuff like this  and deploy port. Now it will be almost instant  

• 00:40:52 to start the application. You will not be spending  the whole time to reinstall, and every file you  

• 00:40:58 have generated, everything you did, will be kept.  This is how the permanent storage system works.  

• 00:41:04 SimplePod is exactly the same as RunPod; I will  also show that. Okay, now when I click the port  

• 00:41:10 instance, you see there is also the 7860 port.  First of all, connect with Jupyter Lab so that we  

• 00:41:16 can start the application. I recommend you to run  the installation again to not have any issues. So  

• 00:41:22 open a new terminal, run the installation. This  time the installation should be almost instant.  

• 00:41:28 Let's see. Okay, it's saying that requirements  are verified; everything will be just verified.  

• 00:41:33 It will only install the FFmpeg again. Okay, it  is verifying the packages, everything is getting  

• 00:41:40 verified right now. All the models it is skipping  since their hash values were already verified. So,  

• 00:41:47 it will take like 1 minute to verify everything  and get ready. This is the recommended way with  

• 00:41:52 my old applications; I recommend you run  the installer again to be sure to not have  

• 00:41:57 any issues after you start your permanent storage  again. Okay, everything verified. It is now just  

• 00:42:04 reinstalling the FFmpeg, then we will be ready.  Yes, ready. Now return back and start. So this  

• 00:42:10 time we will be able to connect both from proxy  and Gradio live. However, I recommend connecting  

• 00:42:18 from Gradio live because RunPod proxy always  causes issues for me—it doesn't work very well.  

• 00:42:25 Therefore, I recommend the Gradio live option  always. Okay, it has started on the local port  

• 00:42:30 and also on Gradio live. So let's open the Gradio  live, and to open the local port, you see there is  

• 00:42:36 this 7860 which is the port it starts. Okay,  looks like we have to use 7861; sometimes it  

• 00:42:44 requires you to start with the plus 1. This is how  the RunPod proxy works. So let's also show that;  

• 00:42:51 I will just repeat the steps. You see from here I  can also edit the port and I will just add 7861.  

• 00:42:58 When I do this, it restarts the port; therefore,  I will be needed to start it again, but once you  

• 00:43:06 restart, you don't need to run the installation  again. Restarts keep the workspace installations;  

• 00:43:13 so after the restart—okay, it is done—I will just  connect back to the Jupyter Lab interface. Yeah,  

• 00:43:20 it will take some time. Okay, it started. This  time I will just run the start command, not the  

• 00:43:26 installation, because I only did a restart. When  you stop the port or terminate and start again,  

• 00:43:32 you need to run the installation; if you just  restart, you don't need to run the installation  

• 00:43:37 again. It is restarted; I will just run the  start command. This start command also updates  

• 00:43:46 the application if there are newer versions. Okay,  the application started. So, let's open the Gradio  

• 00:43:51 live and we should be able to connect from 7861.  Yes, yes, I know this is awkward, this is weird,  

• 00:43:59 but this is how the RunPod proxy works. Okay.  I don't know, maybe I should block or unlock;  

• 00:44:05 let's try block. Okay, Gradio live starts. Yeah,  proxy didn't start, but this is the way you can  

• 00:44:11 try it. Let's generate a song with the default  values on RunPod. It will be slow to load models;  

• 00:44:17 their network storage system is slow. Okay,  we got an error. But don't need to panic;  

• 00:44:24 I know the reason, I know the solution. Remember  when we were installing we had this network RunPod  

• 00:44:32 related error? The reason for this error was  that UV installs with multi-threads based on  

• 00:44:39 the number of CPU cores, and this machine has 256  CPUs. Therefore, it was spawning so many threads  

• 00:44:48 to install and it was causing this issue and it  corrupted our virtual environment. Therefore,  

• 00:44:54 we got this error. So, how are we going to solve  this error? First of all, restart your pod. Then  

• 00:45:01 we are going to delete our virtual environment  and run the installer again. Also don't worry,  

• 00:45:06 I have updated the zip file; in the future,  the installation will be limited to 4 threads.  

• 00:45:11 If your machine is better, like if you're  on SimplePod or if you trust your machine,  

• 00:45:16 you can increase this to improve the speed of  installation; however, on RunPod, I recommend  

• 00:45:22 making it 4. So first we will delete the virtual  environment. For deleting the virtual environment,  

• 00:45:26 we are going to use this command. Let's connect  back to the Jupyter Lab interface. Open a new  

• 00:45:32 terminal, copy-paste. It will delete the virtual  environment; this is the only way to fix when the  

• 00:45:37 virtual environment is corrupted. Then copy the  installation command again. It is still deleting  

• 00:45:43 the virtual environment; when it deletes the  virtual environment, it will disappear from here.  

• 00:45:47 You can refresh to see. You see the RunPod network  storage system is slow; therefore, Massed Compute  

• 00:45:54 is a much better alternative. However, they don't  have a permanent network storage system; that is  

• 00:45:59 their disadvantage. Or you can use SimplePod; I  didn't encounter such issues, and their network  

• 00:46:04 storage system is also fast. Okay, it is deleted.  Now I will open a new terminal and run the install  

• 00:46:10 command again. This time I will limit it to 4.  So if you get such errors, you can reduce it  

• 00:46:15 even to 2 or 1. This is the way of preventing such  operating system errors. You see stale file handle  

• 00:46:24 errors on RunPod or any shared network having  systems; in my Windows computer or in Massed  

• 00:46:32 Compute, I never encountered this issue, nor on  SimplePod. However, this is extra information  

• 00:46:37 for you in the future to fix such issues yourself.  Obviously, since we are running 4 threads instead  

• 00:46:44 of the maximum, it will be slower to install,  but it should work perfectly fine. Okay,  

• 00:46:49 the installation went smooth this time; we do not  see any errors. Everything is looking perfect. The  

• 00:46:57 speed was decent, not very bad, but it was slower  obviously, and since we had downloaded everything  

• 00:47:04 they are just skipped—already verified—and now  I can start the application again. So this was  

• 00:47:12 important information for any other application in  the future where you may encounter this problem,  

• 00:47:20 and now you will know the solution. Okay, the  application started; let's open the Gradio live  

• 00:47:26 link. You see it is like this. If you encounter  any issues with Gradio live, you can open it  

• 00:47:32 in a private window; it may help sometimes, or  restart your browser entirely. As a rule of thumb,  

• 00:47:38 always run the application with default values  then verify it is working. You see it said it  

• 00:47:44 could not parse the server response, but then it  started because we clicked it too early. Probably  

• 00:47:50 the Gradio is taking some time to load, especially  when you change the model, waiting to load values  

• 00:47:56 accurately. So the processing started; we have  no error this time and it should work. Let's  

• 00:48:01 see. Once the model is loaded, the subsequent  generations will be very fast, but the initial  

• 00:48:07 model loading will be very slow. And if you want  to monitor the VRAM usage, you can open a new  

• 00:48:13 terminal. Type pip install nvitop like this. Then  type nvitop and it will open the nvitop window. It  

• 00:48:23 shows the driver version and the CUDA version of  the driver. This CUDA version is the version of  

• 00:48:28 the driver, not what the template has. So this is  important. So it is starting to load the model; we  

• 00:48:35 have to wait because model loading is slow because  the hard drive system on RunPod is slow. But it is  

• 00:48:42 loading. We have full optimizations to speed up  this process both on Windows, on Linux, and on  

• 00:48:48 cloud machines. We can also follow the status  here. You see it is using Flash Attention;  

• 00:48:53 I have compiled this myself so that it supports  every cloud GPU. This compile took over 12 hours;  

• 00:49:01 it was really brutal to make it right. We are  using torch 2.11 with CUDA 13. Hopefully, I will  

• 00:49:07 upgrade applications and compiles to torch 2.12;  I'm waiting for it to mature. We are also using  

• 00:49:14 torchao; this is just torchao loading, but torchao  is used in the inference. So, this is a fully  

• 00:49:21 optimized and highest quality, highest performance  programming application. It took me weeks to make  

• 00:49:28 this application and get it to this point. I  will show the second run as well. You see the  

• 00:49:34 first run is taking like 200 seconds because it  is all taking time with loading the model. Okay,  

• 00:49:40 the inference started. So the models were loaded;  inference is really fast as you can see. Wow,  

• 00:49:47 really, really fast. It is using like 24 GB of  VRAM; you can perfectly run this on an RTX 5090  

• 00:49:53 as well. So the first song has been generated.  It may take a little bit of time to appear here  

• 00:50:00 because of the Gradio live. You can click this  download; if it doesn't appear here, you can also  

• 00:50:05 download from outputs which I will show. Let's  listen to it. Excellent, the default is working.  

• 00:50:17 Let's generate another song and see the speed.  Okay, generation started; it should be ultra-fast  

• 00:50:25 since it will use the already loaded model. We are  really fast, I can see that already. So, let's see  

• 00:50:32 how many seconds. Okay, it is already done. Oh my  god, yep. You see how many seconds it took? Let's  

• 00:50:39 go to outputs. This is the 3rd song; let's see the  manifest.json. So, it took literally—let's see—15  

• 00:50:48 seconds. So, on this GPU, you can generate 4 full  songs in 1 minute. In 1 hour, you can generate 240  

• 00:50:58 songs and it costs only 2 dollars per hour. And  how to download every generation? Go back to your  

• 00:51:06 workspace, go to ACESTEP Premium, and right-click  outputs and download as an archive, and it will  

• 00:51:12 zip it and start downloading everything like  this. You can of course download from here;  

• 00:51:17 click this and it will download. So how to  terminate the machine properly? You can stop  

• 00:51:22 the machine. When you stop the machine, it will  still use some credits. You see it shows 0. Why?  

• 00:51:28 Because it is using the credits in my storage.  So currently it is using 14 dollars per month.  

• 00:51:35 You need to also delete this if you want to not  use any storage. I'm going to just delete this  

• 00:51:41 because I am done with it. But this is how you use  the permanent storage system. How to use RunPod,  

• 00:51:48 how to solve issues. Basically, we have explained  everything. You see I have to remove this first,  

• 00:51:54 so I will just terminate the pod. So what is  the difference between permanent storage and  

• 00:51:58 the regular network volume? So let's go to pods;  everything is selected. Let's select this GPU and  

• 00:52:04 let's select the volume disk. So you see network  volume uses my permanent network; volume disk uses  

• 00:52:11 the volume, and volume disk also uses some credits  by itself. And this time when I stop this you will  

• 00:52:19 see that it is going to use this much money.  However, I can resume this if there is a GPU  

• 00:52:24 later. And to make it 0, I need to terminate  the pod. Okay, let's also delete my permanent  

• 00:52:30 storage too. So I will not get money wasted and  now I am spending 0. Let's verify that. Okay,  

• 00:52:38 it still shows my spend rate; let's refresh. Wow,  it is taking some time to update, I guess. But  

• 00:52:46 I am spending 0 right now because I don't have  any pods or I don't have any storage here. Wow,  

• 00:52:53 it still didn't delete it; man, the system became  very slow. Oh, I generated this inaccurately with  

• 00:52:59 the network volume now accidentally. Yes, why I  had generated it accidentally? Because let me show  

• 00:53:05 you so you won't make the same mistake:  if you select the network volume here,  

• 00:53:11 it will automatically create it. This is actually  misleading; this is how I wasted money. Yes,  

• 00:53:16 I just understood it. So you need to select volume  disk if you don't want such accidental spend;  

• 00:53:23 if you make network volume, it will automatically  create unless you select your existing network  

• 00:53:28 volume, or you need to select volume disk. Okay,  this was everything about the RunPod part. Now  

• 00:53:33 I am going to move into Massed Compute. So for  Massed Compute we have Massed Compute instructions  

• 00:53:41 read.txt file. Open it. Please use this link to  register; I appreciate that, this is important.  

• 00:53:49 After registration, add some credits to your  account from billing. Then go to deploy. In the  

• 00:53:56 deploy menu, it will show you available GPUs like  this. Now the main difference of Massed Compute is  

• 00:54:02 that it is always super fast; its speed is like  20 times faster than RunPod. You won't get such  

• 00:54:09 errors and it is cheaper than RunPod. However, it  doesn't have a permanent network storage system;  

• 00:54:16 so every time you have to install or you have  to download models. So let's use the same GPU.  

• 00:54:22 From category, select 'creator'. From image  select 'SECourses'. And you see currently it is  

• 00:54:27 2.19 dollars per hour; it is the same as RunPod.  However, we have a coupon 'SECourses'. This is  

• 00:54:34 working on every GPU that you can see here,  every one of them. So, this is a great GPU;  

• 00:54:40 you can also take even better ones like H200 NVL  or H100 or like H200 NVL. But this is probably  

• 00:54:49 best for inference as a price-performance balance.  Let's verify. And it is now 1.64 dollars. You  

• 00:54:56 see it is 25% cheaper than RunPod. Then click  deploy. The initial generation may take a little  

• 00:55:04 bit of time; you have to wait for initialization.  Meanwhile, let's install the necessary application  

• 00:55:10 to connect, which is ThinLinc Client. You can also  connect from the browser, but I don't recommend  

• 00:55:15 that. So, open this ThinLinc Client link from  here. Let's allow all downloads according to your  

• 00:55:22 operating system; since I am on Windows, let's  download it. Installation is so easy: open it,  

• 00:55:28 click Yes, then click next, accept, next, install.  That's it; everything is default. Run ThinLinc  

• 00:55:34 Client. Then you need to set the shared folder; so  you can share the zip file. Options—you can also  

• 00:55:41 read the other options here—go to local devices; I  only enable clipboard synchronization and drives.  

• 00:55:46 Then click details, and in here add a shared  folder like this with read and write permission.  

• 00:55:52 So this is my folder on my PC. And okay. So copy  the file into your shared folder; so it is in  

• 00:55:59 my shares folder right now, the installation zip  file. Then I need to just wait for initialization  

• 00:56:04 to be completed. Sometimes you may refresh  this to verify whether it is done or click  

• 00:56:09 this running instance and it will refresh, but  it should automatically show you as well. Okay,  

• 00:56:14 so our machine is ready to connect. You see, click  details, copy the login URL, paste it into here,  

• 00:56:21 copy the username, paste it into here, and copy  the password and paste it into here. Then connect,  

• 00:56:28 click continue. Just wait a little bit. This  ThinLinc Client may be slow for big file transfers  

• 00:56:35 but for small files it should be fast. For big  file transfers, you can use like your Google  

• 00:56:40 Drive or your OneDrive. So this is the interface:  go to home, go to thin drives. You will see there  

• 00:56:46 is your shared folder, whatever the name you have  given. And inside this, I will see my zip file.  

• 00:56:51 You can also open Google Chrome here, login to  your Patreon, and download from here. So it is  

• 00:56:57 basically like your Windows but it is running on  a cloud machine and this is a Linux Ubuntu system.  

• 00:57:04 So our installer zip file is here; drag and drop  it into downloads. Do not run anything here;  

• 00:57:10 you have to drag and drop them into your downloads  folder. Right-click and extract here. Enter inside  

• 00:57:15 the folder, open Massed Compute instructions  read.txt file. Copy this command, return back  

• 00:57:22 to the downloads, enter inside this folder,  and open in terminal. You see this location  

• 00:57:29 is so important: I am in this folder where these  files are located. Right-click and paste and hit  

• 00:57:35 enter. Then it will start the installation. The  installation on Massed Compute will be ultra-fast;  

• 00:57:40 let's watch it real-time because their disk system  is really fast. When you get this software updated  

• 00:57:46 information just click cancel, you don't need it.  So it is installing Python 3.10, then installing  

• 00:57:53 packages with a virtual environment. If you are  a Linux user, Mass Compute installation is what  

• 00:57:58 you need. It should work perfectly fine on Linux  as well on your local Linux system as well. So  

• 00:58:04 for Linux users, I recommend my Massed Compute  scripts. So the installation almost completed;  

• 00:58:10 I mean this is like 100 times faster than RunPod.  Yes, it is done. You see it took like 1 minute to  

• 00:58:16 install. Then the download begins and the download  speed is also amazing on Massed Compute; their  

• 00:58:21 network and their disk speeds are unchallenged.  There are no other cloud services like them. So  

• 00:58:27 you are seeing it in real real-time: it is just  downloading, verifying, downloading, verifying.  

• 00:58:32 Download speed is around 500 megabytes per second,  even faster. The merge speed is very fast since it  

• 00:58:38 is downloading with 16 different connections  then doing hash verification. I'm trying to  

• 00:58:43 unify scripts as much as possible; whatever is on  Windows is also on Linux and on cloud services,  

• 00:58:50 I'm all trying to use the same stuff. Even the  requirements and other stuff are the same. So,  

• 00:58:55 it is just downloading and verifying really,  really fast. Yes, I mean yes, 800 megabytes;  

• 00:59:02 it means like it is 5 gigabit internet connection  speed—it is just amazing. All right, so the  

• 00:59:08 installation has been completed. All of the models  are already downloaded. We can even see the full  

• 00:59:13 size of the folder; so it is 65 GB. Then return  back to Massed Compute instructions read.txt file.  

• 00:59:22 Copy this starting command, return back to the  folder, and open a new terminal. The location of  

• 00:59:28 the terminal is super important; right-click and  paste. Then it will start the application. We can  

• 00:59:33 see that the start of the application will be also  really fast. There is one difference with this:  

• 00:59:38 you see it is starting like this, so I can  open the localhost as well here since it is  

• 00:59:46 a desktop-based operating system. So you see  now this is running inside the Massed Compute,  

• 00:59:52 not in my computer. So I can use it locally like  this or I can copy this Gradio live link and open  

• 00:59:59 it in my computer. This is the way I prefer so  that I have the full smoothness of the interface.  

• 01:00:07 It is like running on my own Windows computer but  it is running on the server. Okay, we got this  

• 01:00:13 error again, so it was because of the loading;  let's refresh. It is also happening because of  

• 01:00:18 the Gradio live; sometimes it can happen, but  refresh of the page should fix it. Yes. So the  

• 01:00:23 model is selected; let's make a quick generate  song. Always do default generation then start  

• 01:00:29 playing with other stuff; verify that it is  working first because your GPU could be broken,  

• 01:00:34 your machine could be broken—anything can  happen—and the usage is exactly the same as  

• 01:00:39 on the Windows tutorial part. Once you reached  this point, when you return back to your CMD,  

• 01:00:44 you can see the status. You can open a new window  and type nvitop and you can see the progress on  

• 01:00:52 the GPU. It is loading; the loading is much  faster on Massed Compute as expected since  

• 01:00:57 their disk speeds are much faster. Once the model  is loaded, the subsequent generations will be  

• 01:01:02 really fast. Okay, generation is proceeding with  amazing speeds. Yes, really fast. Almost done.  

• 01:01:09 And since this is the turbo model, it is only 8  steps inference. Yes, it is already generated and  

• 01:01:16 the song has appeared. Since it is Gradio live, it  is taking some time, but it is here. Let's listen.  

• 01:01:27 So I can generate as many as songs I want and  use the perfect one. The next generation will  

• 01:01:33 be—let's see how many seconds—it should be around  15 seconds, maybe even faster. Okay: 7, 8, 10, 12,  

• 01:01:42 13, 14. Yes, in 15 seconds I got a new song and  it should appear. Yes, you see this way I can  

• 01:01:50 generate as many as I want and use the best one. I  came with This way you can get your perfect song;  

• 01:02:00 you can try different styles, use any lyrics  you want, and everything else is the same. When  

• 01:02:05 you click here, it will download the song,  or you can go to your installation folder,  

• 01:02:10 go to outputs—everything will be here. You can  copy them into your thin drives. Enter here;  

• 01:02:18 when you generate something here or when you  copy something here, it will be synchronized  

• 01:02:22 with your shared drive on your operating system.  So my shared drive is here; you see the outputs  

• 01:02:27 arrived. It is also copying; right now you can  see that it is synchronizing. It is not very fast,  

• 01:02:34 but for small files it should work fine, and  now they are on my computer; they are all  

• 01:02:39 synchronized. So how are you going to stop your  machine on Massed Compute? There is no stop; I  

• 01:02:45 mean when you stop your instance, it will continue  spending your money. Therefore, once you are fully  

• 01:02:51 done and you did back up your data and everything,  you need to delete it like this. If you need to  

• 01:02:58 transfer big files, search 'wget' on our channel  and watch this tutorial. This way you can upload  

• 01:03:05 your folders and big files onto Hugging Face  and quickly download them back into your Massed  

• 01:03:11 Compute instance. You can also back up your data  or big files in your Google Drive or OneDrive and  

• 01:03:16 quickly download them back to your Massed Compute.  Everything is possible once you learn it; it is so  

• 01:03:21 easy to use, and this is the tutorial that you  need. So now I will show on the SimplePod. Open  

• 01:03:27 back the RunPod SimplePod instructions and you  will see that there is this link to register.  

• 01:03:31 Please use this link to register; I appreciate  that. Once you registered and logged in,  

• 01:03:36 add some credits to your account. Then return  back to the instructions and you will see that  

• 01:03:42 there is this template; you need to use this.  Double-click it; it will select the template. Now  

• 01:03:48 it may be a little bit different in the selection  part compared to RunPod, so I will explain.  

• 01:03:54 This SimplePod also supports a permanent network  storage system. You see there is 'Add persistence  

• 01:04:00 volume'; if you select this, it will use it. If  you don't, it will use a temporary disk. So let's  

• 01:04:05 select 'Add persistent volume', then click 'use  template' or 'edit and use'. I prefer you to do  

• 01:04:11 first 'edit and use' to add permanent storage. So  on this screen, you see there is 'Add new volume'.  

• 01:04:18 I already have a volume, but you can add a new  volume. So when I click 'none', it doesn't work,  

• 01:04:23 so I have to first generate the volume. Let's  generate the volume and repeat: go to storage. I  

• 01:04:29 am going to delete this one and I will generate a  new one. For the tutorial, they only have a single  

• 01:04:36 data center right now; so the biggest disadvantage  of SimplePod is that it doesn't have as many GPUs  

• 01:04:42 as RunPod, but it is cheaper and faster. So  let's make this 200 GB and save. You see the  

• 01:04:47 200 GB per month price is 6 dollars; it was  14 dollars on RunPod. Save. Then return back  

• 01:04:54 and double-click this link; it will open the  template. Let's say 'Add persistent volume',  

• 01:04:59 'edit and use'. Select the volume from here:  Tutorial. The mount point is accurate. You can  

• 01:05:04 add the ports like 7860 and 7861. You can give the  name like this, and everything else is automatic.  

• 01:05:15 Save and use. Now it is not started yet; we only  did set our accurate template. You see there is  

• 01:05:21 'change template', there is 'edit template', and  the template is selected. The filters are here;  

• 01:05:26 you can apply them. So currently I am going  to select RTX Pro 6000. You see the price is  

• 01:05:33 1.6 dollars; it was 2.15 dollars on RunPod.  Almost the same price as Massed Compute,  

• 01:05:41 so you can choose either one of them. You can also  use the half RTX Pro 6000—this is shared between 2  

• 01:05:48 people, so you are getting half performance and 48  GB. This is lesser performance than an RTX 5090;  

• 01:05:55 you can also pick an RTX 5090, the price is  amazing and the speed and quality is amazing.  

• 01:06:00 Since this model fits into 32 GB with maximum  quality, you can also pick this one, and when  

• 01:06:07 you scroll down, you will see all of the features  here and run. Then it will start the machine;  

• 01:06:14 the machine is getting composite. Once you get to  this screen, the rest is exactly the same as the  

• 01:06:21 RunPod. Let's just wait; the data and everything  will appear here. You can also see the logs  

• 01:06:25 here. You see system general crash logs are also  shown; it is showing my current spending. Okay,  

• 01:06:31 console and Jupyter links appeared. We can also  see the volume: you see it is using our tutorial  

• 01:06:38 permanent storage system, so everything will be  automatically saved. Click the Jupyter direct;  

• 01:06:43 my browser is particularly giving a warning—I  will just say I want to continue and confirm.  

• 01:06:48 This is because it is using some HTTPS link,  therefore it is giving that error, but it is fine;  

• 01:06:55 this is running on HTTPS, so it is secure. Drag  and drop your zip file, same as on RunPod; you can  

• 01:07:01 also use this upload files icon, and right-click  and extract archive once it is uploaded. One big  

• 01:07:08 advantage of SimplePod is that there is a direct  file browser. This is ultra-fast and it allows you  

• 01:07:15 to upload and download very big files very fast;  RunPod doesn't have this feature but they have.  

• 01:07:20 So once the extraction is completed it should  be instant. Open RunPod SimplePod instructions.  

• 01:07:26 You will notice that it is currently installing  with concurrent installs set to 4. On SimplePod,  

• 01:07:32 you can make this 8; on RunPod, we make it 4 to  be secure because their network storage system is  

• 01:07:38 slow. On Massed Compute, we make it default;  it always works. This should be really fast  

• 01:07:43 to install on SimplePod. So the installation  started; it is quickly installing right now,  

• 01:07:49 same as on RunPod. Okay, so the installation has  been completed. Let's verify if there are any  

• 01:07:56 errors or not; all looking great so far.  I am scrolling down slowly and it started  

• 01:08:02 downloading the models and is still downloading  the models, I think. Let's go to the very bottom;  

• 01:08:09 yes, still downloading the models. Let's just wait  a little bit more. All right, so the models have  

• 01:08:14 been downloaded. Installation has been completed;  we can see that downloaded 75, all completed. Now  

• 01:08:21 we can start the application. So return back to  the instructions.txt, copy the start command here,  

• 01:08:28 open a new terminal and paste it. If you get  any installation error as happened in RunPod,  

• 01:08:34 you can delete the virtual environment and  install again. So the application is starting; we  

• 01:08:39 just need to wait a little bit more. Okay, so the  application started quickly; let's open the Gradio  

• 01:08:44 live. You can also connect from this link: you see  this is the HTTP port which we have enabled. Let's  

• 01:08:53 see if it will work accurately. Okay, let's try  plus 1—maybe as it is in the RunPod proxy. No, it  

• 01:08:59 didn't work on SimplePod, but it doesn't matter;  we are going to use the Gradio live. As I said,  

• 01:09:05 make a default generation. This error is not  important; this is happening because of the Gradio  

• 01:09:12 live, let's click again. If you get that error,  okay, the second time clicking started. Yes, it  

• 01:09:19 is happening because of the Gradio live, but it is  not important; it happens one time until the page  

• 01:09:25 is fully loaded. The generation started; to follow  everything, I think we can open nvitop. Okay,  

• 01:09:31 we need to install first: pip install nvitop, then  nvitop. Yes, so you can monitor your VRAM usage  

• 01:09:38 here; you can see your CPU usage and memory usage.  The loading of the model is much faster than  

• 01:09:45 RunPod. Once the first time is loaded, the second  and subsequent generations will be much faster;  

• 01:09:50 but as a rule of thumb, generate one time with  the default values. Verify it is working then  

• 01:09:56 proceed to use other features. Everything  is shown in the Windows tutorial part,  

• 01:10:01 so watch it to learn if you skipped to this point.  Okay, generation should be completed quickly once  

• 01:10:08 all of the models are loaded. Okay, models loaded  and it is generating with great speed right now.  

• 01:10:15 You can also see this GPU usage, but I see that  this one is limited to 500 Watts. This should be  

• 01:10:22 600 Watts if I remember correctly; I will tell the  developers of SimplePod. Their driver version is  

• 01:10:28 very up to date—you see, 595, whereas on RunPod  it was 580. Their driver is also supporting CUDA  

• 01:10:36 13.2; this is a big disadvantage of RunPod.  Okay, generation completed, the song appeared  

• 01:10:42 here. Everything is same: you can click here to  download, you can go to ACESTEP Premium from your  

• 01:10:53 Jupyter Lab interface, right-click outputs and  download as an archive; but SimplePod, as I said,  

• 01:10:59 supports direct download. Go to file browser;  you can even download the models from here or  

• 01:11:04 upload here. When I scroll down to workspace,  I will see everything is here. Double-click it:  

• 01:11:10 ACESTEP Premium. Double-click it and let's  download the outputs folder. So I click it  

• 01:11:14 and I will click this download and it will zip and  download; it will be super fast. Keep it. You can  

• 01:11:21 even download models: go to models, let's download  one of the big models. For example, this one; this  

• 01:11:26 is like 9 GB. Click download; you see the download  will start immediately with full speed, amazing  

• 01:11:33 speed as you can see. Same way you can upload very  big files: go to the folder wherever you want,  

• 01:11:38 click this upload icon. You can upload a file or  folder, upload whatever you want. So to prevent  

• 01:11:45 your credits from being used, you need to delete  the instance. There is no stop button in here;  

• 01:11:50 when you delete your instance, if you were  using your permanent storage volume like this,  

• 01:11:55 everything will be kept there. Delete instance,  confirm. Now my storage is back here, it is fully  

• 01:12:02 working. To use again and to start again is the  same as the RunPod part: return back to the RunPod  

• 01:12:07 SimplePod instructions and double-click the  template link. It will appear like this: add  

• 01:12:12 persistent storage, edit and use. Make sure your  volume is selected from here and it will continue;  

• 01:12:18 save and use. Then select your GPU; you can use  any GPU like this, for example, let's use the 5090  

• 01:12:25 this time. Scroll down and run, it will start  the machine and it will use my storage. Okay,  

• 01:12:31 machine started back; click direct Jupyter Lab,  show details, I want to continue. Since this was a  

• 01:12:37 full stop, I recommend running the install command  again; it will be really fast. The second time,  

• 01:12:43 it will just verify the installation; it should  be almost instant. It will fix the extra libraries  

• 01:12:51 and dependencies that need to be installed  into the instance, not into the shared storage,  

• 01:12:58 like FFmpeg. Yes, it is verifying everything.  You see it is just verifying the models,  

• 01:13:03 not redownloading, not reinstalling. So the  installation verified, then just start again  

• 01:13:10 and start using it back. So this is it; thank  you so much for watching. You can always join our  

• 01:13:17 Discord channel; joining the Discord channel is  my number 1 recommendation to contact me. You can  

• 01:13:23 make a reply; the Discord channel link is here.  You can join our subreddit, follow me on LinkedIn,  

• 01:13:30 and to see all of our scripts, you see we have our  Patreon exclusive posts index. Click it and you  

• 01:13:36 will see all the scripts we have with their links  and details. You can use Ctrl+F to search anything  

• 01:13:44 here like the ComfyUI, F3 to switch between  them. Our subreddit is getting bigger and bigger;  

• 01:13:51 you see we have a massive amount of subscribers,  visits, and views. You can also leave a comment  

• 01:13:57 here. I recommend you to read everything  here to understand how we are developing,  

• 01:14:03 what features we are adding, and how to use  them. I have done so many different updates,  

• 01:14:08 so this is so important. The zip file also updated  to version 5.1. The zip file may not get updated  

• 01:14:15 all the times whenever I update the version of  the application because it patches updates from  

• 01:14:22 a remote repository. So you just need to run the  Windows install or update file if the zip file is  

• 01:14:30 saved. So the new instance also started; this  was how we reuse on SimplePod with a permanent  

• 01:14:37 storage system. Now I am going to terminate  everything to not spend any money; currently,  

• 01:14:42 I'm spending like this much. So go to my  servers—you can also have multiple servers here,  

• 01:14:47 it will show in 'my servers'—I will just delete  it. If you want to open it back, just click the  

• 01:14:52 instance name and it will happen; also you can  switch them from here too. So let's just delete;  

• 01:14:57 there is no stop button. Then I need to delete my  storage too. So I will delete that as well. Okay,  

• 01:15:02 thank you so much and hopefully see you later in  the LoRA training tutorial of the ACESTEP 1.5.