A potential fix for MOSS training issues....
So I compared MOSS's training sigmas with Krea 2's and it was like looking at MiniMax all over again.
Proof1 community thread
Key points
- I went to bed, then got up again, because something from my MiniMax H3 training wouldn't leave me alone - on H3 I had to shift the share of training steps at the clean end of the noise range (where faces and fine detail get refined) before likeness came good.
- I've just pushed an update (v8.7.0) with a new Clean-end share option for MOSS training, on by default at 60%.
- My test run is going as I write - previews are holding their composition between epochs, and at epoch 12 it had already passed the best likeness my previous settings got to in 30 epochs.
- Numbers so far - same 53-image dataset, rank 16, 0.25 MP, LR 1e-4, same seed, ArcFace likeness vs 3 reference photos:
Sources (1)
- [1]A potential fix for MOSS training issues....r/StableDiffusion (top, day) · Oct 19, 01:53 AM
So I compared MOSS's training sigmas with Krea 2's and it was like looking at MiniMax all over again.
I went to bed, then got up again, because something from my MiniMax H3 training wouldn't leave me alone - on H3 I had to shift the share of training steps at the clean end of the noise range (where faces and fine detail get refined) before likeness came good.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Basalt: NVFP4 + faster prefill for Qwen3.8 Flash-Next on Blackwell (single 5090)
- Oct 17, 2026Qwen Flash Next on Mac Mini is runnable without maxing out your memory - results of different memory usage for ssd streaming
- Oct 17, 2026Blown away by Minimax H3 sometimes. It nailed this one!!!!
- Oct 17, 2026MOSS-Image LoRA training is now in AcademiaSD LoRAlab Trainer Studio (Base and SFT, under 6 GB VRAM)
- Oct 16, 2026FLUX 3 vs KREA 2 vs Z image turbo vs Qwen 2.1
- Oct 15, 2026What a £300, 35W Snapdragon mini PC can actually do 👀: 35B coder at ~30 tok/s, and an 80B model that doesnt even fit in RAM running at 17 tok/s
