Fizgig 8.8.1: a new MOSS-Image training technique
retained text
Quick update on MOSS-Image training in Fizgig (free, open source LoRA trainer). The video above shows the new controls. **What changed** - **New default Clean-end share: "Slide".** MOSS-Image's own schedule spends very few steps on nearly clean images, which is where faces and fine detail are learned, so likeness was slow and composition hopped around between epochs. 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. On the dataset that gave me the most trouble, it was the best of eight settings I tried: around 70% ArcFace likeness from epoch 20 right through to epoch 60. For context, that's a very high score: different real photos of the same person often score lower. It's tuned on datasets up to about 55 images; with a lot more, a steady 40% (Single) may suit you better. Still early days, so please share what works for you. - **Base previews with the SFT look.** 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. It downloads itself the first time (about 2.1 GB). I'd still leave the Turbo LoRA off for previews. - **Every enlarged preview now says how it was made** (SFT, Base with the SFT look, Base only, Turbo or not). - **Krea 2 on 16 GB cards:** Auto no longer picks heavy block swapping when NF4 fits, which was eating CPU. **Also: update ComfyUI.** A lot of people have told me their Fizgig LoRAs work better since ComfyUI's update this week. - GitHub: [shootthesound/Fizgig](https://github.com/shootthesound/Fizgig) - Release notes: [v8.8.1](https://github.com/shootthesound/Fizgig/releases/tag/v8.8.1)
document fetched 2026-10-21 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xazk0f/fizgig_881_a_new_mossimage_training_technique