Training a Krea2 character LoRA with sloppy settings — and it still worked
cradledust · reddit · 2026-09-15
A user trained a Krea2 character LoRA on Runpod using AI-Toolkit, admitting they "did everything wrong," yet the result turned out reasonably good.
The setup: 75 randomly cropped images across 30 aspect-ratio buckets, no captioning, low-to-medium source resolution, all default Krea2 Raw settings — 3000 steps with 10 samples every 250 steps.
The only regret: sample generation was a time sink — nearly 10 wasted minutes per 250 steps, and nothing was visible before step 1500 anyway, so the samples should have been skipped. A useful data point that character LoRA training is more forgiving than assumed.
More from Multimodal
- Synthesia rolls out Avatar Builder to all users for custom AI avatars — synthesiaIO · 2026-09-15
- Runway MCP prompt generates 30s photoreal stealth-game footage, fully shared — tlakomy · 2026-09-15
- AI music videos should be judged by how easy they are to fix, not first-draft quality — Positive-Page1631 · 2026-09-15
- MiniMax H3 T2V generates similar faces across seeds while WAN stays more varied — ThrowAwayBiCall911 · 2026-09-15
- Fixing MiniMax H3 face distortion with a portrait-distance prompt constraint — Devajyoti1231 · 2026-09-15
- GPT Image 2.5 in Action: One Reference Image Keeps Characters Consistent Across Ads — JaynitMakwana · 2026-09-15