LoRA step counts can differ across AI-Toolkit and OneTrainer

BelowSubway · reddit · 2026-07-22

LoRA training steps are not directly comparable across tools

The poster asks whether a 1,500-step LoRA trained in AI-Toolkit is equivalent to a 1,500-step LoRA trained in OneTrainer.

Their experience suggests the answer is no. In AI-Toolkit, they usually find good LoRAs around 1,250–2,000 steps, but running OneTrainer for 100 epochs produced a checkpoint labeled 4,500 steps. Even so, the 3,000+ step versions performed better than expected.

The practical takeaway is that “steps” can be counted differently across training setups, so you can’t compare LoRA checkpoints by step count alone; you also need to account for epochs, dataset size, and how each tool logs progress.

Original post →

More from Multimodal

Multimodal channel →