stardance has been extended another month! the new deadline is october 31 :)

You are browsing as a guest. Sign up (or log in) to start making projects!

marina

@marina

Joined June 25th, 2026

  • 20Devlogs
  • 3Projects
  • 2Ships
  • 20Votes
Open comments for this post

@marina on My personal site!!!

1h 24m 48s logged

began to create my personal site! today i added a leaf animation a stardance star icon andddd a about me and projects section aswell as a pumpkin cursor!

0
0
26
Open comments for this post

@marina on project mari

3h 33m 36s logged

Added this to a portable .app image now, the python is now a bundle and works out of the box with options to edit for a custom config

0
0
72
Open comments for this post

@marina on project mari

1h 47m logged

Today i removed old code that i used ai to assist with in the beginning the majority of it is now hand written along with using code from the riko project made by JustRayen, the main part made by ai is the gpu passthrough on my server as both vulkan and the other fix for amd cards wasnt working and i hadnt commited any of my new code on github to fix things like the memory bug and visual effects with the hair clipping into clothing

0
0
33
Open comments for this post

@marina on project mari

1h 57m 11s logged

today i fixed the gpu passthrough for an amd gpu, it currently took a long time to respond when using vulkan it now responds a few seconds quicker after first prompt

0
0
74
Ship Pending review

@marina on project mari

I made Marina, an anime character who lives on your Mac desktop in a transparent window. You talk to her or type, and she answers in her own voice, lip syncs, and acts out what she writes. She starts speaking before she’s finished thinking, you can interrupt her by talking over her, and sometimes she starts the conversation. Everything except the language model runs locally on an 8 GB MacBook Air. It’s built on rayenfeng/riko_project, which had a voice pipeline but no avatar. the hard part was fitting it all into 8 GB of RAM, which is why the language model runs on a separate GPU server. Making her feel responsive was the hardest part. Streaming replies a sentence at a time meant making sure she never split a stage direction or a number like 3.5 in half. Letting you interrupt her was tricky too, because her own voice goes back into the mic.

The quickest way to see her is https://skep13.github.io/marina/ where you can type to her in the browser. Replies there are recorded lines in her real voice.
The app only runs on a Mac. Keep the folder at ~/marina, not in Desktop or Documents, or it will hang.
Run ./setup-mac.sh first.
Bring your own .vrm avatar and put it at app/models/model.vrm.
For the model, install Ollama, run ollama pull llama3.2:3b and set base_url to http://127.0.0.1:11434/v1 in the config.
Hover over her to see the buttons.

  • 7 devlogs
  • 14h
Try project → See source code →
Open comments for this post

@marina on project mari

1h 57m 43s logged

i fixed the ability to interrupt it mid conversation to feel more realistic, she now speaks once every 25 minutes of not being interacted with, she can set timers, check battery and other similar stuff and finally she speaks before the prompt is finishing generating which feels more natural and reduces thinking time from 23s to 0.2s to first token.

0
0
90
Open comments for this post

@marina on project mari

24m 40s logged

today i decided to combine both whisper and kokoro to have a more accurate voice model that i can run locally, i also added a new design inside of vroid studio that will be added later on :)

0
0
116
Open comments for this post

@marina on project mari

3h 36m 45s logged

today i typed up, the requirements file to have all of this running, tommorow i will be setting up the api link to my server. As well as a selectable menu to choose local or api hosted AI

0
0
20
Loading more…

Followers

Loading…