Krea 2 LoRA fails at 6,000 steps: Tala Madani's style resists every model tried
Inevitable_Pen9043 · reddit · 2026-11-06
Reddit user InevitablePen9043 reports that LoRA training on Krea 2 to learn painter Tala Madani's style failed even at 6,000 steps with learning rates between 1e-4 and 3e-4 — outputs stay generic paintings and never transform characters into her signature abstract creatures.
Oddly, Krea 2 successfully learned several other abstract, difficult painting styles. The poster suspects it might need to be trained as a character rather than a style, and asks the community for input — a notable edge case for style learning in image models.
More from Multimodal
- Opus 5.5 Motion Design Videos Are "Extremely Mid" Despite Hype — beechinour · 2026-11-07
- Vork: agent-first 3D/VFX engine claims 24x faster rendering than Blender — ravisparikh · 2026-11-07
- Mistral's Voxtral 3 adds end-of-turn detection for responsive voice agents — MistralAI · 2026-11-07
- ChatGPT Images 2.6 vs Nano Banana 2.1: 4B weekly images vs a 50% price cut — Aiden_Tech_Ai · 2026-11-07
- Magnific shares 6 motion experiments with Magnific One + Move, prompts included — charis_ai · 2026-11-07
- AI-generated kitten wrestling video fools viewers — IcyPea3192 · 2026-11-07