// tools/lora-training-calculator

LoRA Training Cost Calculator

Calculate the exact cost to train a LoRA before you start. Pick your model, steps, and GPU — see cost and training time across every major provider. Flux Dev · SDXL · SD 3.5 · prices verified June 2026

Best quality. Needs A100/H100 or RTX 4090 with gradient checkpointing. Standard for production LoRAs.

1,500

Most LoRAs need 3–5 runs before the result is production-ready.

Select GPU— only GPUs with enough VRAM for Flux Dev (20GB required)

Training estimate — A100 80GB · 1,500 steps · 3× runs
Training time
5m 30s
per run
Cost per run
$0.082
cheapest provider
Total budget (3× runs)
$0.247
before you get a good LoRA

Note: Larger batches possible. Ideal for dataset > 100 images.

Provider comparison — A100 80GB · 1,500 steps

#ProviderType$/hr$/runTotal (3×)
1Vast.aiMarketplace$0.900$0.082$0.247
2RunPodCommunity$1.19$0.109$0.327
3ModalServerless$2.10$0.193$0.578
4LambdaDatacenter$2.79$0.256$0.767
Sorted by cost per run · prices verified June 2026 · full GPU cost breakdown →
Budget for failure, not just success

The most expensive part of LoRA training is not the GPU — it is failed experiments. Budget for 3–5 training runs before you get a good LoRA, not just 1. The first run tells you if your dataset is clean. The second validates your hyperparameters. The third (or fifth) is usually the one that actually ships.

// after training

Deploy your LoRA to production

Once trained, deploy your LoRA to Runflow for production serving. Runflow handles LoRA routing, hot-swapping, and scaling automatically — no DevOps overhead.

Training times estimated from community benchmarks at 1024×1024, batch size 1, bf16, gradient checkpointing. Actual times vary ±20% by dataset and configuration. · More tools → · Updated June 2026