InfernoSaber Model Repository
Overview
Welcome to the InfernoSaber, an Automapper for BeatSaber with fully adjusteable difficulty.
Model List
Recommendation of the models/branches in the following order:
- expert_15: Current favorite, trained on curated high difficulty maps (8+ nps) with a like/dislike rate >90%. Usually generates good flow.
- easy_15: Trained on curated low difficulty maps (5- nps) with a like/dislike rate >90%. More creativity but less flow, especially for high difficulty maps.
- pp3_15: Trained on random maps from my personal collection, mainly ranked high PP maps.
- fav_15: Trained on all personal favorite maps throughout the years, mainly in the NPS area of 5-8
Model Details
- Model Name: InfernoSaber
- Model Version: v.1.7.0
- Architecture: Multiple custom DNN and Autoencoder with TF v.15
- Training Objective: Classification/Creation of song maps
- Language: Python
- License: Model and source code are free to use
- Repository: https://github.com/fred-brenner/InfernoSaber---BeatSaber-Automapper
Intended Use
This repository contains the trained models for inference (go to the other branches at the top: "Files and versions", click on "main" -> select "expert_15" or similar).
All models are trained on different datasets with mixed genre.
The models are required by the source code at github (and automatically downloaded if not found).
Inference and Training
The models can be run on most systems and do not require GPU for inference.
For training / creation of new models, the following specs are required:
- 10-20 GB free RAM per 50 maps (mainly depending on the creativity of the maps)
- 8-15 GB VRAM per 50 maps (mainly depending on the creativity of the maps)
Limitations and Ethical Considerations
The models and corresponding source code are free to use.
Selling the generated maps is prohibited and labeling as human made is against the beatsaver policies.
This project is not meant to be replacing human maps, but as extension to cover more unkown songs and massively reduce the amount of human effort required.
Contact Information
If you have comments, suggestions, or want to contribute to new models/features feel free to open a discussion.
Cheers!
