Fizgig cuts image LoRA training from 49 to 20 minutes on RTX 5090

shootthesound · reddit · 2026-11-08

The open-source LoRA & fine-tune studio Fizgig (689 GitHub stars) ships a major training speedup:

On an RTX 5090, a MOSS-Image LoRA run (94 images, batch 1, Muon, 70 epochs) now takes 20 minutes instead of 49
The Muon optimizer step now handles a LoRA's matrices together instead of one at a time: 12x faster Muon math with identical results — the main source of the batch-1 speedup
PyTorch 2.15 upgrade with gradient checkpointing cuts VRAM sharply: Krea 2 drops from 22.9 GB to 14.9 GB, MOSS-Image at batch 3 from 21.7 GB to 7.1 GB
Batch training is now fully supported; other models also got faster, with Moss the most dramatic

More from Multimodal