Fizgig 9.2 Cuts MOSS-Image LoRA Training to 20 Minutes
Fizgig 9.2 moves to PyTorch 2.15 and a faster Muon step: a MOSS-Image LoRA trains in 20 minutes instead of 49 on an RTX 5090, at a third of the memory. LoRAs load in ComfyUI.
Fizgig 9.2, released on November 8, makes every training run faster and lighter, and MOSS-Image most of all. On an RTX 5090, a MOSS-Image LoRA (94 pictures, batch 1, Muon, 70 epochs) now finishes in about 20 minutes instead of 49, and likeness scores came out the same or better (release notes).
Fizgig's Start tab. The trainer runs on its own, then saves kohya-format LoRAs for ComfyUI/models/loras/.
The numbers
| MOSS-Image LoRA on an RTX 5090 | Before 9.2 | 9.2 |
|---|---|---|
| 94 pictures, batch 1, Muon, 70 epochs | 49 min | about 20 min |
| Memory, batch 3, gradient checkpointing | 21.7 GB | 7.1 GB |
At batch 3 the trainer now matches OneTrainer's 1.07 s a step on the same card while using less memory (6.9 GB against 7.4 GB), so batched runs are a practical option rather than a compromise.
A run on the Training tab, with the queue in the sidebar.
What changed
- PyTorch 2.15. With gradient checkpointing on, memory drops hard: Krea 2 at batch 1 goes from 22.9 to 14.9 GB. On a 16 GB laptop 4090, Fibo 2.0 trains 22% faster, Klein 13% and Krea 2 11%.
- A faster Muon step. It now handles a LoRA's matrices together instead of one at a time: 12x faster, same results. LoKR training is 4 to 6x faster on every model.
- Faster batches. Captions of different lengths no longer slow attention down: 18% faster steps on MOSS-Image, 12% on Krea 2, 9% on Qwen-Image 2.1.
MOSS-Image Turbo Train
Version 9.2.5 adds an experimental MOSS-Image Turbo Train preset: the Fast preset (rank 16, 0.25 MP, Muon) with Adaptive LR running between 2e-4 and 4e-4, and per-image tracking so each picture gets its own pace. It is a quick way to see what a LoRA could become before committing to a long run, and it holds up on a good dataset.
Presets for MOSS-Image also moved to a new Clean-end share, 30:7:60:8:3: 30% of steps on nearly clean images until epoch 7, then 60% from epoch 8. Load a preset again to pick it up, since saved settings keep the old schedule.
Updating
The update downloads PyTorch 2.15 (about 3 GB) and needs Python 3.11 or newer. NVIDIA drivers must support CUDA 13, version 580 or newer; if Fizgig cannot see the GPU after updating, update the driver. On AMD the bundled ROCm PyTorch is unchanged, and rented pods pick up the new version after a restart.
.safetensors files that load with the stock ComfyUI LoRA loader.
Comments
Sign in with GitHub to join the discussion.