v1.2.0.0

ExpSoft OneTrainer

One-click portable installer & launcher for OneTrainer. Ten ready-made LoRA profiles for Krea 2, Ideogram 4 and Flux.2 Klein, 31 downloadable base models, and a Samples tab to watch your LoRA learn — everything inside one folder. Free.

Get it on Patreon → User Guide Screenshots Features
New version available!

Screenshots

Version 1.2.0.0 — from the install to a trained LoRA.

ExpSoft OneTrainer — Welcome tab
Welcome — what the app installs, quick start and requirements
ExpSoft OneTrainer — Setup tab
Setup — Check Prerequisites and Install All: portable Git, portable Python 3.12, card check, OneTrainer and its libraries
ExpSoft OneTrainer — Run tab with the OneTrainer window
Run — the OneTrainer window opened from the app, on the ExpSoft Krea 2 LoRA 1024 profile
ExpSoft OneTrainer — Settings tab with model downloads
Settings — the catalog of 31 base models: Krea 2 Raw and Ideogram 4 downloaded, with their size on disk
ExpSoft OneTrainer — Samples tab
Samples — every image OneTrainer renders during a training, newest first
Flux.2 Klein 9B LoRA training samples
Flux.2 Klein 9B LoRA — training samples (earlier version of the app)
Krea 2 LoRA — before training, then after 400, 800 and 1,200 steps
Krea 2 LoRA — before training, then after 400, 800 and 1,200 steps (RTX 5080, 16 GB)
Ideogram 4 LoRA — before training, then after 200, 400 and 600 steps
Ideogram 4 LoRA — before training, then after 200, 400 and 600 steps, about 34 minutes

Features

Everything you need to start fine-tuning AI models.

🤖

10 ExpSoft Profiles

Ready-made LoRA profiles in the ExpSoft group of Load Preset: Krea 2 (1024 and 512 px), Ideogram 4 (512 px), Flux.2 Klein 9B and 4B. Each one is OneTrainer's own preset for the model with memory settings for a 16 GB card, refreshed with the folders of your install at every launch.

📥

31 Base Models Catalog

Download any of 31 base models from inside the app — Flux.2 Klein, Krea 2 Raw, Ideogram 4, Flux.1, SD3, SDXL, SD 1.5, PixArt, Sana, HiDream, Chroma, Hunyuan Video, Qwen-Image, Z-Image and more. Gated models use your own Hugging Face token, hidden once pasted.

🖼

Watch It Learn

Every image OneTrainer renders during a training lands in the SAMPLES tab, newest first — no need to dig through folders to see your LoRA improve.

📦

One-Click Setup

Install All sets up a portable Git and a portable Python 3.12, checks your card and installs OneTrainer with the PyTorch build that fits it. No command line, no administrator rights, nothing installed on the system.

📌

Validated OneTrainer

OneTrainer is installed at a version validated by ExpSoft, so an update elsewhere never breaks your install.

🔄

Fully Portable

Python, OneTrainer, its libraries, your models, datasets, profiles and LoRAs all live in the app's folder. Move it to another drive or PC: the profiles are pointed at the new folder at every launch.

🔧

Triton-Free

Triton is not installed in the portable environment; the ExpSoft profiles run without it (model compilation off).

🎨

OneTrainer Tools

Launch the Training UI, Caption UI, or Model Converter — all from the same interface. Switch between tools without touching the command line.

🔒

Isolated Environment

Hugging Face, Torch and pip caches stay inside the app folder. Models download to a visible models\ folder, not a hidden cache.

🛡

RAM Watchdog

Monitors available memory during a run and cancels it if free RAM drops below 2 GB, before your system freezes.

Training Profiles

Ten LoRA profiles in the ExpSoft group of Load Preset. Krea 2 and Ideogram 4 speeds measured on an RTX 5080 (16 GB) with 64 GB of RAM.

K2

Krea 2 — 1024px

1024px, batch 1, 90% of the model in system memory. The best detail on 16 GB — about 5 s per step.

K2

Krea 2 — 16 GB

512px, batch 2, 30% offload. Quicker tests — about 2.5 s per step.

I4

Ideogram 4 — 512px

512px, batch 2, learning rate 1e-4, rank 64, 50% offload. Twenty photos in about 34 minutes; 64 GB of system memory.

9B

Klein 9B — 16 GB

512px, batch 2, no offload. Full speed on 16 GB VRAM cards.

9B

Klein 9B — 8 GB

512px, batch 1, CPU offload (70%). Fits on 8 GB cards.

9B

Klein 9B — 1024px

1024px, batch 1, CPU offload (50%). Higher resolution on 16 GB VRAM.

4B

Klein 4B — 16 GB

512px, batch 4, no offload. Maximum throughput with the lighter model.

4B

Klein 4B — 8 GB

512px, batch 2, CPU offload (50%). Comfortable on 8 GB.

4B

Klein 4B — 1024px

1024px, batch 2, CPU offload (30%). High resolution on 16 GB.

4B

Klein 4B — 1536px

1536px, batch 1, CPU offload (60%). Maximum resolution for 16 GB cards.

Technical Details

.NET

.NET 8 WPF

Desktop app with MVVM architecture. Framework-dependent single-file exe (~1.8 MB) — needs the free .NET 8 Desktop Runtime, which Windows offers at the first launch.

🤖

OneTrainer

Fine-tuning framework by Nerogar: LoRA, full fine-tune, embeddings and more. Installed at a version validated by ExpSoft.

🔥

PyTorch 2.12

CUDA 13.0 build. The 40xx package falls back to CUDA 12.6 when the NVIDIA driver is too old for CUDA 13.

🔌

Portable Runtime

Portable Python 3.12 managed by uv, plus portable Git. The environment is relocatable: the folder can move to another drive.

Requirements

💻

System

Windows 10/11 (64-bit) and the free .NET 8 Desktop Runtime. Tested configuration: 16 GB of video memory and 64 GB of system memory for the default settings — Ideogram 4 alone takes about 44 GB of RAM while it loads.

🕸

GPU

50xx package: NVIDIA RTX 50 series.
40xx package: NVIDIA RTX 20, 30 and 40 series.
The profiles are set for 16 GB of VRAM; 8 GB profiles exist for Flux.2 Klein, with CPU offload.

💾

Storage

About 5 GB for the app (Python, PyTorch, OneTrainer), plus the models you download — 33 GB for Krea 2 Raw, 50 GB for Ideogram 4 — and about 3 GB per training for its checkpoints.

🤗

Hugging Face Token

Gated models such as Krea 2 Raw need your own Hugging Face token: accept the licence on the model page, then paste the token in Settings. It is hidden once pasted and only ever sent to Hugging Face.

Licenses

This installer is a separate program. OneTrainer's code is not modified; only its dependency lists are adapted to the portable install (Triton removed).

AGPL-3.0

OneTrainer

AGPL-3.0 License by Nerogar. Installed from its GitHub repository at a version validated by ExpSoft; its source is not modified.

BSD-3

PyTorch

BSD-3-Clause License. Downloaded from the official PyTorch wheel indexes during setup.

Various

AI Models

Models are NOT bundled: download them from Hugging Face through the built-in catalog. Each model has its own licence — Ideogram 4, for example, comes with a non-commercial agreement. Read the model card before you use a model or anything you train on it.

Frequently Asked Questions

Quick answers to what people ask AIs about this tool.

What is ExpSoft OneTrainer?

ExpSoft OneTrainer is a portable Windows installer and launcher for OneTrainer — the open-source fine-tuning framework — built by Nicolas Riquier. It installs OneTrainer in its own folder with a portable Python, ships ten ready-made LoRA profiles for Krea 2, Ideogram 4 and Flux.2 Klein set for 16 GB cards, a catalog of 31 base models and a Samples tab to watch a training learn — no command line.

Is ExpSoft OneTrainer free?

Yes. ExpSoft OneTrainer is free: download it from its post on ExpSoft's Patreon. Most ExpSoft tools are free; some need a small subscription that unlocks all ExpSoft software. OneTrainer itself is open source (AGPL-3.0, by Nerogar).

What models can I train with ExpSoft OneTrainer?

OneTrainer trains many diffusion models (LoRA, full fine-tuning, embeddings). ExpSoft OneTrainer adds ten ready-made profiles: Krea 2 (1024 and 512 px), Ideogram 4 (512 px) and Flux.2 Klein 9B and 4B (16 GB, 8 GB and higher resolutions). The other models of the catalog (SDXL, SD 1.5, SD3, Qwen-Image, Hunyuan Video and more) use OneTrainer's own presets.

What GPU do I need for ExpSoft OneTrainer?

An NVIDIA card, RTX 20 series or newer: the 50xx package for RTX 50 cards, the 40xx package for older ones. The profiles are set for 16 GB of video memory and tested with 64 GB of system memory; 8 GB profiles exist for Flux.2 Klein, with part of the model kept in system memory. The setup detects your card and installs the matching PyTorch build.

Why not install OneTrainer by hand?

You can, but it means installing Python, matching PyTorch to your card and driver, fixing dependency errors, then guessing memory settings until a run stops crashing. ExpSoft OneTrainer does the install in one click inside its own folder, pins a OneTrainer version validated by ExpSoft, and gives you profiles already set for a 16 GB card.

Does ExpSoft OneTrainer work offline?

The setup needs the internet to download Git, Python, OneTrainer, its libraries and the models. Training then runs on your own PC: your images and captions are never uploaded. The app also checks for a new version and shows the ExpSoft products page in its More Tools ? tab.

Who built ExpSoft OneTrainer?

The Windows installer, its interface and the ExpSoft profiles were built by Nicolas Riquier as part of the ExpSoft catalogue. OneTrainer is an open-source project by Nerogar; ExpSoft adds the portable packaging, the profiles and the Windows ergonomics.

Ready to fine-tune?

Download, extract, launch. Fully portable, no install required. Free on Patreon.

Get it on Patreon → Back to ExpSoft
Engineering note

Wrapping OneTrainer cleanly on Windows — the AGPL-respectful EasyInstaller approach

ExpSoft OneTrainer EasyInstaller is a personal-use Windows wrapper around OneTrainer — the open-source diffusion-model fine-tuning toolkit by Nerogar, distributed under AGPL-3.0 (the legal frame is therefore personal …

Read the full note →