Phosphene 3.2.4 fixes Ideogram 4 for slower Apple GPUs. It now runs on M1 and M2 Macs, not just the fast ones.
Big thanks to cocktailpeanut, who tracked down exactly what it took to get Ideogram running on an M1 Max and sent the fix.
What was wrong
On slower Apple GPUs, Ideogram's large fp8 kernels overran the macOS Metal command-buffer watchdog - macOS killed the GPU work as "too disruptive to interactivity" and the render crashed (SIGABRT). Fast M4 machines slip under the watchdog; M1 and M2 didn't.
The fix
The Ideogram subprocess now runs with MLX_DISABLE_COMPILE and MLX_METAL_FAST_SYNCH, so MLX submits shorter Metal command buffers that stay under the watchdog. The default quantization is now -q 6 instead of raw fp8 - same legible text, much faster, much less RAM.
Also in this release
- Generates 1 image by default instead of 4 (the old default was quietly 4x the time and memory).
- Live progress while it renders, instead of sitting on "Loading pipeline".
- The "reveal outputs folder" button now opens the actual folder, with your latest render selected.
- Smaller image sizes (768x432, 512x512, 432x768) for weaker machines.
- Fast mode (-q 4) for low-RAM Macs.
No setup anymore
Ideogram used to need a Hugging Face token and a license click. It now works out of the box - an un-gated mirror of the weights, no token, mflux bundled. Your first render just downloads the model (about 26 GB, one time).
Apple Silicon only. Update via Pinokio - one click. Free and open source: https://github.com/mrbizarro/phosphene