- Author
- Not stated (published on the personal site zhinit.dev, GitHub user zhinit)
- Published
- 16 July 2026
- Kind
- Personal blog post describing a hobby project, with audio samples
- Original
- www.zhinit.dev/blog/training-a-kick-drum-diffusion-model
- Subject
- Audio models and analysis
- Themes
- Small means, Art through code
What it is
The post comes from the developer behind the zhinit.dev site. It explains KickWithReverb, a web app that generates kick drums. The post teaches the pipeline step by step for newcomers.
What it says
- The post says diffusion runs on a small compressed latent, not raw audio.
- The post says it trained a VAE, a diffusion U-Net and a HiFi-GAN vocoder.
- The post says the VAE squeezes each kick to 352 numbers, about 63 times smaller.
- The post says many keywords gave poor kicks. The app always uses "hit house".
- The post says the kicks have a grainy artifact that works as texture.
- The post says generation takes about 2.5 seconds warm and about 10 cold.
- The post says each model took about a day to train.
What evidence it gives
The evidence is the author's own build, with code, weights and a live app linked. Audio examples are embedded, but the post reports no listening test. Figures such as 13,615 kicks and the timings are self-reported.
Interests and limits
The author built and hosts the app, so the account is a self-description. No blind listening test or comparison with other kick generators appears. Results rest on one personal sample library. The author says the models needed more hyperparameter tuning. The text keywords were the weakest part.