Controlled Krea 2 character LoRA test: 0.5 MP is the sweet spot, 1.0 MP overbinds
Cold-Worldliness5392 · reddit · 2026-09-28
A controlled character-training experiment on Krea 2: the same 35-image Guy Fieri dataset trained as four variants (0.25/0.5/1.0 MP LoRA plus 0.5 MP LoKR with only LR bumped 5e-5→1e-4), all at 20 epochs with identical seeds/prompts/workflow on an RTX5090 via Fizgig on runpod.
Training times: 10 min (0.25 MP), 18 min (0.5 MP), 34 min (1.0 MP), 23 min (LoKR) — all far faster than older long-epoch workflows.
Key findings:
- 0.25 MP was already very strong, defying claims it's too low for Krea 2.
- 0.5 MP is the sweet spot: better facial structure, more natural detail, best likeness/flexibility balance.
- 1.0 MP adds micro-detail but overbinds dataset style cues: headshots went from no necklace (0.25) → subtle chain (0.5) → large diamond rope (1.0). The model learned the full "Guy Fieri visual package", becoming caricature-like. More detail ≠ better identity learning.
- LoKR preserved detail/structure while avoiding overbinding — a promising "premium" option.
Recommendation: 0.5 MP + 20 epochs + rank 8 LoRA as production baseline; 0.5 MP LoKR worth further testing. A second-character comparison is underway.
More from Multimodal
- Seedance 2.5 Flash Generates Intense Fight Scene From Just 2 Random Images — aziz4ai · 2026-11-16
- Runway CEO Predicts Fully AI-Generated Cultural Phenomenon by 2027 — c_valenzuelab · 2026-11-16
- Stunning AI-generated motion-blur portraits with a film-photography feel — azed_ai · 2026-11-16
- A Bauhaus functionalist prompt template for geometric poster-style AI art — LudovicCreator · 2026-11-16
- Vivix W1 video model: 100 ten-second ads for about $5 at $0.005 per second — Aiden_Tech_Ai · 2026-11-16
- User Shares Coral Reef Creation Labeled 'Qwen 4 Max (xhigh)', an Unannounced Model Name — intellectronica · 2026-11-16