Which LoRA trainer should I use? Do I need the command line?

LoRA Dataset Studio (LDS) trains with ai-toolkit and writes its training config for you: once ai-toolkit is installed and set in Settings → Local tools, you pick a recipe and start the run from the browser, with the queue and logs on the Runs page. Datasets also export as standard ZIPs of images with same-name .txt captions, so the same set can feed kohya_ss, OneTrainer or another trainer. The command line is needed once, to install ai-toolkit, which LDS does not install for you.

ai-toolkit is the training engine. LDS runs it as a command-line process, from its own folder and its own Python environment (the one with PyTorch), and starts and stops it for you. The web interface that ships inside ai-toolkit is unrelated: LDS never launches it or reads it.

From the dataset's training panel, LDS:

• builds the ai-toolkit config from the family recipe, the dataset kind and your Advanced options;
• checks images, captions, duplicates, free disk, VRAM, the base model and compatibility before the launch, then again when a queued run actually starts;
• queues local runs and lists every run, local or cloud, on the Runs page with progress, logs, samples, stop, retry, continue and download actions, plus a paste-safe Share config summary.

1. Install ai-toolkit in its own folder. The getting started guide recommends Python 3.11, or ai-toolkit's own installer, python -m manager install, which provisions a supported Python and the CUDA build of PyTorch that matches your driver. A plain pip install torch on Windows installs a CPU-only build, and training then runs on the CPU for hours without an error.
2. In Settings → Local tools → ai-toolkit, set the folder that contains run.py. Fill Python interpreter only when there is no venv or .venv folder next to it, for example with conda, uv or a portable Python.
3. Press Test. It checks that the interpreter can import PyTorch; a passing test unlocks training and JoyCaption captioning.

Export ZIP writes one 10_<trigger> folder of PNG images with same-name .txt captions. Captions carry the trigger word, and a reference photo, when the dataset has one, is included as an extra image captioned with the trigger alone; Style exports contain content captions only, with no trigger and no reference. The 10_ prefix follows the kohya repeat-count folder convention, and ai-toolkit ignores the folder name. Write .txt files puts the same kohya-style captions next to each kept image in the dataset folder instead.

For an architecture LDS does not support, or for experimental config keys, use ai-toolkit directly with the exported set.

• Local training needs an NVIDIA GPU with enough memory; see how much VRAM a LoRA needs.
• LDS drives ai-toolkit and no other trainer.
• The Docker images and the RunPod image do not include local ai-toolkit training: connect an ai-toolkit on the host, or train in the cloud.
• Without a suitable GPU, the Cloud training plugin rents one per run on vast.ai with your own API key, for Z-Image, Krea 2, FLUX.2 Klein and Qwen-Image 2.1. SDXL, FLUX.1 and Anima train locally only.
• If ai-toolkit uses conda, uv or the system Python and is not detected, fill the Python interpreter field; see ai-toolkit isn't detected.

• What steps, rank and learning rate should I use?
• How much VRAM do I need to train a LoRA?
• Train, guided and advanced (workflow guide)
• ai-toolkit settings