Easey Glyph

--- license: mit tags: - generative-art - audio-reactive - flow-matching - live-performance - vj - glyph pipeline_tag: unconditional-image-generation --- # EASEy-GLYPH: Audio-Reactive Generative Glyph Visuals Initially released as a Stable Diffusion experiment, the Effortless Audio-Synesthesia Experience ([EASE](https://github.com/kevinraymond/ease)), was my first semi-deep trip into serious audio reactive visuals. It could have its place, but definitely is too heavy for many non-CUDA machines. Enter [EASEy-GLYPH](https://github.com/kevinraymond/easey-glyph). The core design goal was to make real-time generative visuals accessible on Apple M1 hardware, which I quickly discovered was not going to be any typical diffusion-based solution. The entire architecture works backwards from that constraint: a 34M-parameter flow-matching model generates 32x32 glyph grids (not pixels) that render into stylized, low-resolution visuals at 60fps on modest hardware. The low resolution isn't a limitation - it's an intentional aesthetic choice that produces chunky, graphic visuals with real transparency and color depth. An optional "super-resolution" CNN can upscale to 256x256 when the hardware budget allows. Read more about the app in its repo. This HF repo is for the models I'm including so you can try it out right away! Training your own models is relatively quick as well with a decent GPU, with scripts in the app repo. _NOTE: LLM generated content below_ --- ## Sample Outputs <table> <tr> <td align="center"><img src="images/abstract.png" width="200"><br><b>Abstract</b></td> <td align="center"><img src="images/nature.png" width="200"><br><b>Nature</b></td> <td align="center"><img src="images/ukiyoe.png" width="200"><br><b>Ukiyo-e</b></td> <td align="center"><img src="images/albums.png" width="200"><br><b>Albums</b></td> </tr> <tr> <td align="center"><img src="images/pixel.png" width="200"><br><b>Pixel</b></td> <td align="center"><img src="images/botanical.png" width="200"><br><b>Botanical</b></td> <td align="center"><img src="images/darkpsy.png" width="200"><br><b>Darkpsy</b></td> <td></td> </tr> </table> ## Base vs Realtime Models Each variant ships with two FlowUNet models: - **Base** — unconditional generation. Good for ambient visuals, pool-based morphing, and general exploration. - **Realtime (CFG)** — trained with classifier-free guidance on audio...