Training MOSS-Image LoRAs locally on 8GB VRAM: 500 steps beats 8000, full settings inside
AcademiaSD · reddit · 2026-10-20
The author shares a complete recipe for training character LoRAs on MOSS-Image (OpenMOSS's Apache-2.0 5B single-stream model) locally on an 8GB card using the open-source AcademiaSD LoRAlab Trainer Studio:
- Recommended settings: Base model, 500 steps (1200–1500 only for large/hard datasets), LR 0.0004 with cosine schedule and 100 warmup steps, rank/alpha 8/8, batch 1 with grad accumulation 4, 512² resolution, save every 100 steps. With NF4 quantization, training runs under 6GB VRAM (1s/step on an RTX 5080).
- Why 8000 steps fails: with grad accumulation 4, 500 steps = 2,000 image views vs 32,000 at 8000 steps — long past overfitting; more steps ≠ more quality.
- Checkpoint selection: judge in ComfyUI, not trainer previews; test 300/400/500 before extending.
- Captions: built-in Qwen3-VL-8B NF4 captioner follows OpenMOSS's recaptioning guide with trigger word first; clean, varied captions matter more than dataset size.
- LoRAs export with original layer names for native ComfyUI loading; a RunPod template exists for GPU-less users.
More from Multimodal
- ACEdit Stabilizes 15-Turn Image Editing With Asymmetric Conditioning — tau · 2026-10-20
- AI-generated 4K action film demo suggests action movies will never be the same — JeffSynthesized · 2026-10-20
- Neon cyberpunk AI character art showcase in green and purple — OVolosin82152 · 2026-10-20
- Motion2Code: H3 Max designs motion, Claude Opus 5.5 rebuilds it as open-sourced GSAP HTML — OdinLovis · 2026-10-20
- Polya AI launches Dialog-RSN-1: realtime voice model with sub-300ms median first-token latency — matthen2 · 2026-10-20
- CueForge v3.4.0 brings ComfyUI 0.42.1 parity to mobile with 3D previews and low-resource mode — galactic_lobster · 2026-10-20