AI Toolkit is Ostris's open tool for training LoRAs on current models: FLUX, Wan, Qwen-Image and SDXL. Its web UI queues your jobs and generates sample images from your prompts during training, so you can judge the result before the run ends. On gpu.nz it starts from the template: rent a card, open the interface and create a job.
In short
AI Toolkit runs on cards with 24 GB or more. The most affordable suitable card right now is the Tesla V100 32 GB from 23 ₽ per hour, billed by the minute.
Updated 13 October 2026
24 GB of video memory is the minimum: that is enough for a FLUX LoRA with quantized weights. An RTX 4090 24 GB is a comfortable fit for FLUX and SDXL, and an RTX 5090 32 GB adds headroom. Wan and Qwen-Image are heavier: move to 48 GB, such as an L40S or RTX A6000, if 24 GB forces constant offloading. Sample images need memory too, so leave some room.
Prices and availability come from the live catalog and update automatically.
For a character use 10–30 photos, and for a style 40–100 images. Each image needs a text file with a caption of the same name. Put a trigger word that is not in the dictionary, such as ohwx woman, into the caption. To choose a card for the job, see the GPU guide for LoRA.
dataset/
01.jpg
01.txt # ohwx woman, close-up portrait, golden hour
02.jpg
02.txt # ohwx woman, full body, city street at nightA job sets the step count, how often samples are made and how often checkpoints are saved. For a character on FLUX, 1,500–3,000 steps is a common starting point; a style needs more. Turn samples on with 3–4 prompts that test the trigger word in different poses, and generate them every 250 steps.
| Task | Photos | Steps (starting point) |
|---|---|---|
| A character or a face, FLUX | 10–30 | 1,500–3,000 |
| A style, FLUX or SDXL | 40–100 | 2,000–4,000 |
One FLUX training run of 2,000 steps on an RTX 4090 usually takes between one and two hours. The first steps show the training speed: multiply the time of one step by the total step count to get a rough estimate. For how FLUX works and what it needs, see the guide FLUX on a rented GPU.
By default the LoRA is saved to the output folder, inside a subfolder named after the job, as a .safetensors file. Intermediate versions follow the save setting, and samples go to a separate subfolder. Download the files through JupyterLab (right-click and Download) or with this command:
scp -P <port> <user>@<address>:<path to output>/<job name>/*.safetensors ./