Full recipe: training an accurate MOSS-Image character LoRA with just 40 images
Darth-Daenerys · reddit · 2026-10-21
A detailed walkthrough of training an accurate character LoRA (Fizgig) for the MOSS-Image model in Fizgig, usable in an 8-step LoRA ComfyUI workflow (90s per 5.5MP image on an RTX 3080).
Data: 40 high-quality 1:1 images, 80% face / 20% body.
Training (MOSS-Image Fast preset): adaptive LR on (min 5e-5, max 1e-4), per-image adaptive LR on, 70 max epochs saving every 5, seed 42, LoKR network with factor 8, EMA 0.98, problem-image detection on.
Model stack: MOSS-Image Base DiT (bf16) + FLUX.2 VAE + Qwen3.5-9B text encoder.
ComfyUI workflow: int8 convrot for Unet/Clip, skip layer guidance with single layers=22, KSampler 12 steps / CFG 1.5 / res-multistep / simple scheduler / full denoise, character LoRA strength 1.
More from Multimodal
- Midjourney sref Code 1080457169 Nails Pointillist Snow-Scene Oil Painting Style — sergeantsref · 2026-10-21
- Moss-Image SFT + New DMD2 LoRA: Sharper 8-Step Generation, But Style Collapse — AgeNo5351 · 2026-10-21
- TikTok is flooded with Opus 5.5 motion-graphics AI slop and viewers can't tell — felpix_ · 2026-10-21
- Fibo 2 Edit impresses: uncensored out of the box, fast, no color drift in early tests — East_Shoe8814 · 2026-10-21
- AI film goes from idea to IMAX premiere and greenlit series in 14 days — lmoroney · 2026-10-21
- Kroma/Krea 2 Raw ultrareal 2x upscale workflow: ~60s per image on RTX 5060 Ti — gabrielxdesign · 2026-10-21