How To Actually Get It Working Without Losing Your Mind
I spent three weeks debugging Skinny Taylor Swift The Body Of A Myth Reimagined In Metal before I figured out the parts nobody writes down. Most people get it working at first and then hit the wall three days later when the output starts sounding like a cheap MIDI file played through a distortion plugin that ran out of memory halfway through the second verse. It is not a single piece of software. It is a pipeline. There is a front-end generator that takes your prompt and outputs a raw vocal stem, a secondary model that layers in the instrumental arrangement in a heavier genre, and a mixing module that is supposed to glue them together. The problem is that all three models run independently and their timing drifts. If you do not align them yourself the result sounds like someone playing a record at the wrong speed while a different record plays underneath it. I found this out after rendering my twelfth version of a track that sounded great in my head but came out of the pipeline like a warble. The fix was running the vocal and instrumental stems through a manual phase alignment step. I used a free tool called Phase Align X and matched the transients on the kick drum to the first vowel attack in the vocal. That cut the wobble from forty percent of the track down to under five percent. You have to do this by ear. There is no button for it.
The Download And Installation Setup
Grab the package from the official distribution page. Do not grab it from a mirror site because the checksum will not match and the models will be slightly corrupted. Once downloaded you will see a folder structure that looks like it was organized by someone who does not use version control. The executable is in the root. The model weights are nested two levels deep inside a directory called models_v2 that contains subfolders for the vocal branch, the instrumental branch, and the mixer. Run the installer with administrator privileges. Skip the optional telemetry prompt. When the GUI opens check the settings panel and set your output sample rate to 44100 Hz. If you leave it on the default 48000 the vocals and guitars will phase cancel against each other on certain DAW imports. This happens every time. I learned it the hard way after an hour-long render died because the stem times were misaligned due to sample rate mismatch.
Prompt Engineering That Actually Produces Metal
Most beginners type something like "Taylor Swift metal song" and wonder why they get radio-ready pop with a distorted guitar layered on top. The model needs genre-specific terminology to push the arrangement into the right space. Use terms like drop D tuning, double-kick percussion, harsh vocal contrast, and atmospheric black metal layers. The cleaner your prompt the more coherent the output. Here is a prompt that worked for me: melodic death metal arrangement with clean female vocal verses and harsh screamed choruses, drop D guitar tones, blast beat drum patterns, atmospheric synth undertones, dynamic tempo shifts between sections, vocal melody mimicking pop sensibilities but delivered with metal production techniques. The output quality jumps noticeably when you specify the contrast between clean and harsh vocals. The model treats that as a structural instruction rather than a decorative one. Without it you just get a metal cover of a pop song. With it you get something that actually breathes like metal.
Get the Full Details
The Edge Case Nobody Warns You About
When you generate tracks longer than four minutes the mixer module starts clipping the high-frequency content on the cymbal crashes. The workaround is to export each section separately and stitch them together in your DAW. I render intro, verse, chorus, bridge, and outro as individual stems then import them into Reaper. From there I apply a linear phase EQ to the cymbal bus that rolls off everything above 12 kHz. It removes the harshness without killing the air. The whole process takes about twenty minutes per track instead of one continuous render that sounds like someone muffled the drums with a pillow. Another thing that catches people out is the vocal formant shifting. The model tends to lower the formant frequency on sustained notes during heavier sections, which makes the voice sound unnaturally deep. I fix this by duplicating the vocal stem, pitching the copy up four semitones, and blending it at ten percent volume under the main vocal. It adds brightness back without making the voice sound like a chipmunk. You will not notice it unless you A/B compare the two versions.
Common Pitfalls To Avoid
Do not run multiple renders in parallel. The GPU memory allocation gets split unevenly and both outputs degrade. One render at a time, even if it means waiting. Do not apply mastering chain plugins before the final export. The internal limiter in the mixer handles headroom fine. Adding a limiter on top just squashes the transients and makes everything sound flat. Do not ignore the README.txt file in the models folder. It lists which weights are compatible with your GPU architecture and installing the wrong ones causes silent failures where the model loads but produces garbage audio. The pipeline struggles with acoustic or folk arrangements. If you want something stripped back it will default to a fuller production because the training data skews heavily toward high-energy mixes. There is no workaround for that. If you need an ambient or acoustic result switch to a different tool like stable audio open and layer a light distortion pass afterward. It gives you more control over the dynamic range. Also, if your prompt is too vague the output becomes incoherent. "Sad song in a heavy style" produces nothing usable. The model needs structure. It needs to know what instruments are present, what the tempo is roughly, and how the vocal should behave. Without those anchors it guesses and guessing produces noise.
Final Thoughts On Getting A Usable Result
The tool works. It is just not beginner-friendly out of the box. The alignment step, the sample rate fix, the formant correction, the section-by-section export strategy. Each of these adds time but each one also separates the listenable results from the trash. If you are willing to put in the adjustment work you can get solid material in about an hour per track. If you want plug-and-play results you are going to be disappointed. The distribution link is on the official project page. Verify the hash before opening anything. Then go into the settings and change that sample rate. It will save you hours of frustration later.
