How to Get Model Maker: by FDG
Guide made by FDG on discord
Prep
You will need to following to submit a model:
- General information about your model
- The .pth (weight) file of the model
- The index file for the model
- A HuggingFace account
- At least 1 COPYRIGHT FREE audio demo
Preparing your model files
Your model needs to be packed into a zip file with its final .pth and index file
Google Colab users can get this file by running Step 4 on the colab, then checking /RVC_Backup/Finished/ for the model zip.

For MAINLINE local users, you can find these files in the /assets/weights folder and the /logs folder, respectively. People not using mainline may instead find the pth file in /weights, rather than /assets/weights.

Once you have located both of these files, they need to be packed into a zip file.
We will not accept files packed as anything else rar and 7z will be rejected.

Uploading the files to HuggingFace
Create a HuggingFace account if you haven't already. Once you are logged into HuggingFace, click on your account icon in the top-right corner, and hit new model

From here, you will be asked how to set up your model repository, which is where ALL of your models will be stored. You can name your repository as you please.
Make sure it is set to public and the license is openrail.

Once the repository has been created, locate the Add file button above the file list and select Upload files

You will be prompted to select your files to upload. You can either click on the upload box to browse to your zip file, or simply drag and drop it into the box. After queuing your zip file(s) for uploading, Hit the Commit changes to main button at the bottom.

Once its done uploading, you should automatically be brought back to the file list. From here, right-click on the download icon next to the model's name and click on copy link.

Submitting the model to AI Hub
Once your model is ready, you can submit it in #get-model-maker using the/submit command of QCBot.
While using the command, please fill in the following information about your model:
- modelname: The name of your model
- rvc: The version of RVC this was trained on (this will almost always be 2)
- extraction: The extraction method used for training (crepe and RMVPE are accepted. HARVEST and others ARE NOT)
- epochs: The number of epochs your model trained for
- link: the HuggingFace link of your model that we copied earlier
- image: An image of what your model is representing. (a character or person)
- demo: An audio demo of your character either speaking or singing (this MUST be copyright free AND not contain ANY POST-PROCESSING)
- note: (optional)
Please note that you can only attach one audio demo file using the command. If you wish to include more, you can just send them after the bot confirmation response.
If everything went fine, your submission will be added to the queue and the bot will send a confirmation message, containing your submission ID. You'll need this ID if you want to:
- Check your submission's number in queue with
/queue [ID] - Cancel your submission with
/cancel [ID]
From here, you can wait for a Model QC to verify your model. You'll be notified once it has been reviewed. You can then re-post your model (and any future models) to the forum #voice-models

Extras
Did your model not get accepted? Here are some reasons why.
You forgot one of the files in the zip
The zip file needs to contain both the model weights (.pth) AND the index file (added_...index). Please make sure both of them are included when uploading and submitting.
You used an outdated extraction method
RVC has many extraction methods, we only accept two of them: crepe and RMVPE. Any other extraction method will not get your model accepted.
This includes, but is not limited to: harvest, dio, crepe-tiny and pm.
mangio-crepe is fine.
The model is low quality
Your model may be rejected if one or more of the following is noticed:
- The voice is scratchy/screechy
- The voice is muffled
- The voice is inaccurate to the source
- The voice cannot hit certain notes
- The voice has slurred speech or otherwise cannot pronounce words in its intended language
The model may also have artifacting. Artifacts are when white noise, reverb or other unwanted sounds (popping, clicking, etc) are present in the model. This is usually due to a bad dataset, but training a model for too little can also cause this to happen.
If your model is affected by any of the above, consider cleaning your dataset or getting a better one, and retraining.
The audio demo has an instrumental
Due to concerns over copyright, it is highly recommended that you do not include ANY instrumental AT ALL when submitting a model. Any demos with an instrumental will likely be rejected.
The audio demo is altered
You may desire to edit your demo with reverb and EQing before you send it, please don't do this. It is not representative of the model and will be rejected. You can mix and master however you want for your covers, but for the model demos please make sure you use the raw output from RVC.
Trimming out silence at the beginning and/or end of inference is fine.