Heersin / rf3_tuna_ai

Rune Factory 3 -- Tuna, Generate By so-vits-vic 4.0

rf3_tuna_ai

Rune Factory 3 -- Tuna, Generate By so-vits-svc 4.0

Instructions

1. raw_data from tieba

2. filter by wav.title attributes which contains 'myce'

3. trained by so-vits-svc

   • about 58000 steps, V100 GPU * 1 for 10+ hours (cloud server)
     • cuda 12 and torch 1.13.1
   • pretrained model and env from hugging_face
     • D_0.pth and G_0.pth
     • 44k
   • kmeans cluster used

4. music vocal extraction, try the following:

   • online : melody ML and vocal remover
   • local: deploy uvr5-gui on linux, 3080Ti 3-4 minutes per song
     • window size : 320
     • VR-HP-3 model

5. inference command: python inference_main.py -m "logs/44k/G_[step].pth" -c "configs/config.json" -n "[song].wav" -t 0 -s "[speaker]" -cm "logs/44k/kmeans_10000.pth" -cr 0.4

6. Mix vocal and intrumental/bass/drums ...: Audacity

7. Picture Generate NovelAI

8. Video maker : kdenlive

Result

无知时诋毁图恒宇
懂事时理解图恒宇
成熟时成为图恒宇

Celebration

Rune Factory3 for NS ! Looking forward to new game also.