[ Use case · LoRA & fine-tuning ]
Best AI tools for LoRA training and fine-tuning
Training a LoRA is the most demanding thing you can do on a consumer GPU. The trainer matters less than the workflow around it — dataset curation, captioning, and inference testing all live in the same pipeline.
VRAM FLOOR FOR THIS WORKFLOW
[ WORKFLOW PLAN ]
Decide in this order
- Define the output: resolution, duration or context, batch/concurrency, and how often you will run it.
- Set hard constraints: platform, privacy, license, VRAM/RAM, and whether slow CPU offload is acceptable.
- Choose the workflow: eliminate tools that fail a hard constraint, then compare ecosystem, reproducibility, and switching cost.
The catalogue picks below are a shortlist, not proof that every default configuration fits. Open each datasheet and verify the exact model or extension you intend to use.
Evidence companion
LoRA training hardware
See the assumptions, official sources, and memory trade-offs behind this workflow.
Our picks
06 MATCHED
AI-Toolkit (Ostris)
Modern training framework — Flux, SDXL, SD3 LoRAs in YAML.
OPEN SOURCE16–24 GB VRAM
VRAM fit16–24 GB
sd-scripts (Kohya)
The underlying scripts powering Kohya & most LoRA trainers.
OPEN SOURCE8–12 GB VRAM
VRAM fit8–12 GB
✓ WHAT TO LOOK FOR
- Block-swap support for tight-VRAM training
- Built-in captioning (Florence, JoyTag, BLIP)
- Sample generation during training so you can spot collapse
- Resumable checkpoints
! HONEST TRADE-OFFS
- AI Toolkit documents a specific Flux training path around a 24 GB GPU; other recipes may differ
- SDXL fit depends on resolution, batch, optimizer, checkpointing, and which components are trained
- Full fine-tunes add substantial gradient and optimizer-state memory beyond LoRA training