Fizgig 8.8.1: a new MOSS-Image training technique
Quick update on MOSS-Image training in Fizgig (free, open source LoRA trainer).
Proof1 community thread
Key points
- The video above shows the new controls.
- Slide starts at 60% of steps there until epoch 7, eases down to 30% by epoch 15, then puts every 2nd epoch back at 60% so likeness doesn't fade.
- Tick it on the Samples tab and your Base previews get much of MOSS-Image SFT's look, without swapping models, so they're a lot quicker.
- Not as good as SFT itself, but a good guide.
Sources (1)
- [1]Fizgig 8.8.1: a new MOSS-Image training techniquer/StableDiffusion (top, day) · Oct 20, 06:09 PM
Quick update on MOSS-Image training in Fizgig (free, open source LoRA trainer).
The video above shows the new controls.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0
- Oct 20, 2026How to train MOSS-Image LoRAs locally (8 GB VRAM): settings, tips and workflow
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 19, 2026Moss image speed boost with spectrum moss node
- Oct 19, 2026Krea 2 Turbo @ 4 steps (w/ my 4-step LoRA) vs MOSS SFT at 4 steps (ww AK's 4-step LoRA) - same prompt set you've seen me publish in the past (headline: MOSS is really good! and great job by AK on his LoRA)
- Oct 19, 2026Fixed the vertical colour stripes in MOSS-Image SFT with a 1.9 MB LoRA (extracted, not trained)