A tested MOSS-Image workflow: exact models, LoRAs and settings
AI_Enhancer · reddit · 2026-11-12
A Reddit user who initially found MOSS-Image lackluster shares a full trial-and-error workflow (pastebin link), arguing the base model delivers huge variety, coherence and prompt understanding once optimized.
Models & LoRAs: kromoss fine-tune, official qwen3.5 9B text encoder (do NOT use abliterated encoders — no extra uncensoring, just quality loss), flux2-vae small, the lodestones 8-step DMD2 LoRA (best by far), plus lenovo/samsung realism LoRAs and a careful-strength nicegirls LoRA.
Settings: ER-SDE sampler, powershift scheduler, 11 steps (8–25), CFG 1 (1–4 by step count), 2–4MP resolution — 10–20s per image on an RTX 3090.
More from Multimodal
- AI-made 30s Black Friday ad "The Last Shift" built with Topview T1 — azed_ai · 2026-11-13
- Gemini Decisions API demo answers audio questions near real-time across all modalities — AI_Andrew · 2026-11-13
- Seedance 2.5 animated retelling of the Ramayana squirrel legend shows strong narrative chops — CurieuxExplorer · 2026-11-13
- Krea launches Bots: AI agents that take on entire creative video projects, beta offers $1,000 credits — angrypenguinPNG · 2026-11-13
- Midjourney sref combo 548542432 + 2511950545 yields cohesive sportswear illustration style — sergeantsref · 2026-11-13
- MiniMax H3-Image tested: 16MP editorial covers in 6 steps, ~20s per image — Arn777333 · 2026-11-13