FluxGym is a simple way to train a FLUX LoRA: you upload photos, Florence-2 writes a caption for each one, and a single button starts training. You can train a model of your face, a character or an object without configuring any scripts. The gpu.nz template comes with everything ready, and you pay by the minute while the server runs.
In short
FluxGym runs on cards with 16 GB or more. The most affordable suitable card right now is the RTX A4000 16 GB from 13 ₽ per hour, billed by the minute.
Updated 13 October 2026
The minimum is 16 GB of video memory, as in the template. For a faster run take an RTX 4090 24 GB, where training finishes noticeably sooner. An RTX 4080 16 GB also works, just more slowly. A 16 GB card is fine for a trial run, while a full training run is easier on 24 GB.
Use 10–20 photos of one person or one object, with different angles, expressions and lighting. Identical shots and collages do not help, because the model will also learn stray details of the background. Keep the photos sharp, since blur carries over into the model.
It is easier to upload a large archive with scp. To set up access, see the guide SSH access to a GPU server.
Florence-2 writes a caption for each photo automatically, and the captions come out in English, which FLUX understands best. Check them by hand: a wrong caption ends up in the model. Put a rare trigger word that is not in the dictionary, such as ohwx man, at the start of every caption. You later use it in the prompt to call your model.
ohwx man, a photo of a man smiling in a park, natural lightPrices and availability come from the live catalog and update automatically.
For more flexible training, where you need manual settings, there is the Kohya SS template.
Steps are counted as photos × repeats × epochs at batch size 1: 15 photos with 10 repeats and 10 epochs give 1,500 steps. On an RTX 4090 such a run usually takes between half an hour and an hour, and on an RTX 4080 longer. The first run shows the exact time; for how to estimate it in advance, see the guide FLUX on a rented GPU.
The finished LoRA is in the outputs folder, in a subfolder named after the project. Download the .safetensors file through JupyterLab (right-click and Download) and move it to wherever FLUX runs: Forge or ComfyUI.