Catastrophic Z-Image LoRA training results seek config debugging

the3cumfountain · reddit · 2026-08-16

An experienced SDXL and SD 1.5 LoRA trainer faced catastrophic results while training a LoRA for Z-Image (Tongyi-MAI/Z-Image-Turbo). Despite three days of tweaking, the outputs are severely mangled and melted, losing all semantic coherence. The post includes a detailed YAML config asking the community to identify obvious errors in the training setup.

Original post →

More from Multimodal

Multimodal channel →