Training Minimax H3 character LoRAs on 16GB VRAM: settings and pitfalls
No_Date4828 · reddit · 2026-08-18
A user successfully trained character LoRAs for Minimax H3 on an RTX 4090 mobile (16GB VRAM) with 64GB RAM plus a 60GB pagefile (the model uses nearly 110GB in practice), sharing config links and lessons:
- Dataset: 20-70 images suffice; caption only the trigger word and attributes you don't want baked into the character (clothing, scenery); keep captions to 2-3 short sentences.
- Learning rate: 0.0003 with AdamW fully overcooked by step 1000 (good likeness by 250); 0.0002 trains slower with best likeness at 1000-1200 steps.
- Distance consistency: include distant face shots in the dataset; 512x training limits far-away facial detail, while 1024x with a good dataset should do better.
Only two LoRAs trained so far — the author invites the community to find better settings.
More from Multimodal
- pagedMark: Open-Source Tool Strips Invisible AI Watermarks from Images and Video — d0ofz · 2026-08-18
- As AI Makes Imagery Worthless, Embodied Beauty Grows More Valuable — tedmitew · 2026-08-18
- Testing Song-dynasty aesthetics in image models: local hits, unstable holistic judgment — sujingshen · 2026-08-18
- ComfyUI tool uses model hashes to auto-fix red missing-model nodes — demongatanjieu · 2026-08-18
- Seeking Workflow to Render Fortnite Screenshots as Photorealistic Images — drmarkway · 2026-08-18
- 4DGS enables hyper-realistic 3D videos in web browsers with WebXR support — Scobleizer · 2026-08-18