OneTrainer User Guide

Open the SETUP tab and click Install All. It happens once.

1. 1
   Check Prerequisites — looks at what is already installed and at your card, and ticks what is there. It installs nothing.
2. 2
   Install All — five steps in order: a portable Git, a portable Python, the card check, OneTrainer itself (a version validated by ExpSoft, so an update elsewhere never breaks your install) and its Python libraries with the PyTorch build that fits your card. A step already done is skipped.
3. 3
   The steps — a green tick means done.
4. 4
   The log — every command and its result, also saved in logs\.

On our test PC, with a fast connection, the whole setup took about 3 minutes and uses about 5 GB.

Open SETTINGS. Some models are gated: you accept their licence on their Hugging Face page, then download them with your own token.

1. 1
   HuggingFace Token — paste a token with Read access (huggingface.co › Settings › Access Tokens). The field hides it, so it never shows on screen or in a screenshot.
2. 2
   Save Settings — keeps it in the app's config.json, on your PC. It is only ever sent to Hugging Face.

1. 1
   HF Key — a gated model: open its page on Hugging Face, accept its licence with your account, then come back here.
2. 2
   Download — fetches the model into models\. When it is done the row turns green and shows the size on disk.

The list holds 31 models, from Stable Diffusion 1.5 to Flux.2 and Qwen-Image. The ExpSoft profiles are built for three of them: Flux.2 Klein (9B and 4B), Krea 2 Raw — the version of Krea 2 made for training — and Ideogram 4.

Put your images in a folder under datasets\, each with a .txt file of the same name that describes it: photo-01.png + photo-01.txt. The captions decide what the LoRA learns, so they are worth ten minutes.

Below, the two captions of one of the 20 photos of our test set: a close-up taken outdoors, in front of a teal wall and green leaves.

• A trigger word and a class — a rare word for your subject, followed by what it is: wodk, a man. You write the same words in your prompts later to call the LoRA.
• Describe what changes, never what stays — framing, pose, expression, clothes, background, light. Do not describe the face: what you never describe sticks to the trigger word.
• Variety over quantity — our test used twenty photos taken on different days, in different clothes, places and angles.

For Krea 2, write plain sentences — it reads long, natural descriptions:

A close-up photograph of wodk, a man, his face filling the frame with the top of his head cropped. He tilts his head slightly and gives a soft closed-mouth smile, looking straight into the camera. His hair is messy and pushed back. A white shirt shows at the bottom right. Behind him, a teal painted wall on the left and dense green oleander leaves on the right, softly out of focus. Natural daylight outdoors, casual snapshot.

For Ideogram 4, the caption is a JSON description — the only form this model was trained on. It gives an overall line, the style, and each element of the picture with its box (top, left, bottom, right, from 0 to 1000) and its main colours:

{"high_level_description":"A close-up outdoor portrait of wodk, a man, smiling softly at the camera with his head slightly tilted.","style_description":{"aesthetics":"casual, warm, candid","lighting":"soft natural daylight, even and diffuse","photo":"extreme close-up, eye-level, shallow depth of field, smartphone snapshot","medium":"photograph","color_palette":["#8D6656","#AD9388","#A77F71","#554533","#392C1F"]},"compositional_deconstruction":{"background":"A teal painted wall on the left and dense green oleander foliage on the right, blurred.","elements":[{"type":"obj","bbox":[0,40,1000,960],"desc":"wodk, a man, his face filling most of the frame, head slightly tilted, soft closed-mouth smile, looking into the camera, messy pushed-back hair, a white shirt collar at the bottom right.","color_palette":["#4A392B","#9C7A6C","#B19286"]}]}}

In RUN, click Launch OneTrainer. The OneTrainer window opens next to the app; closing it brings you back.

1. 1
   Launch OneTrainer (on the right of this card, behind the OneTrainer window here) — refreshes the ExpSoft profiles with the paths of your folder, then opens OneTrainer; the card then says it is running.
2. 2
   The log — what OneTrainer prints. A few notes at start-up are normal (see Troubleshooting).
3. 3
   The OneTrainer window — the model and training type at the top right, the settings in its tabs, Start Training at the bottom right.

Then, in the OneTrainer window:

1. 1
   Load Preset › ExpSoft — pick the profile of your model and card (see the profiles). It sets the model, the memory settings and the folders in one click.

• concepts — the ready-made concept reads every folder under datasets\; point it at one folder to train on that set only. The captions are read from the .txt files next to the images.
• sampling — add two or three prompts with your trigger word, for example A head-and-shoulders portrait photograph of wodk, a man, wearing a navy blue suit…, at the size you train at. Set a fixed seed to compare the rounds.
• training — the number of Epochs. The profiles keep OneTrainer's 100; our test used 60 epochs for its 20 images (1,200 steps).

1. 1
   Epochs — how many times the training goes through your images.
2. 2
   Local Batch Size and Resolution — images per step and their size. Both weigh on the video memory.
3. 3
   Layer Offload Fraction and Offload Activations — the share of the model kept in system memory instead of the card (Memory settings).

Click Start Training. The step and epoch bars at the bottom of the OneTrainer window show the progress.

Every image OneTrainer renders during a training lands in the app's SAMPLES tab, newest first — no need to dig through folders.

1. 1
   Refresh / Open Folder — reload the list, or open the samples folder of the latest training.
2. 2
   The images — labelled with the training and the prompt. Click one to see it full size.

At the end of a training, the LoRA is saved in output\ — output\krea2_lora_1024.safetensors with the Krea 2 LoRA 1024 profile. The in-between checkpoints stay in workspace\<profile>\save\: if the last one looks overcooked, an earlier one is often the better pick.

Load the LoRA in the image tool you use with that model, write your trigger word in the prompt, and start around full strength.