MOSS image model LoRA recipe: 50-100 images, done under 5k steps
Any_Tea_3499 · reddit · 2026-10-25
The author shares a working LoRA training recipe for the MOSS image model, praising its diversity and candid realism but noting LoRA training was tricky. Key settings:
- Datasets of 50-100 images with natural-language captions; LoKr 4, trained at 1024
- Automagic3 optimizer with Sigmoid scheduler on High; start LR at 0.0005 with a minimum LR floor since Automagic drives LR down so fast training stalls around 1k steps
- Don't exclude block 22 — noticeably better quality
- Most LoRAs converge before 5k steps; slower to train than Qwen or Krea but better results
Config file plus 20-step fast and 50-step generation workflows are shared for download; Lodestone's turbo LoRA for MOSS base speeds up inference.
More from Multimodal
- Moss Image demos turn heads on Reddit: a 5B single-stream DiT trained with Muon — lovelycontributer · 2026-10-25
- Ant-affiliated inclusionAI's 0.5B open-source TTS model trends on Hugging Face — inclusionAI · 2026-10-25
- One-take AI video: jet ski rescue through a drowning city with no cuts — umesh_ai · 2026-10-25
- Ribbed-glass + neon lighting: a stylized Clayface Midjourney prompt — michaelrabone · 2026-10-25
- Midjourney explorations in a 1930s black-and-white cinematic style — hewarsaber · 2026-10-25
- FIBO 2 vs FLUX 3 Klein side-by-side: photorealistic mug shelf render-off — RobbaW · 2026-10-25