A local training workbench for multiple models and engines
Prepare datasets, train models, and follow your runs in one place.
Download · Quick start · Documentation · 中文
A 100-step style LoRA test by KohakuBlueleaf, trained with their KohakuMossTrainer. The top row shows the base model; the bottom row shows the LoRA results. Shared with the author's permission as a MOSS-Image training example; these results were produced with KohakuMossTrainer (also available as an optional engine in the current source version).
- Multi-engine workbench: Kohya, Anima Fast, Musubi, AI Toolkit, DiffSynth-Studio, and MOSS (
anima_lora) are presented through one engine workflow, with an updated model/engine matrix and tutorial index. - Independent Kohya runtime: Kohya uses its own environment under
extensions/kohya/; trainer source updates are separated from engine installation, repair, and auditing. - Kohya survives updates: source and portable-package updates preserve user-owned training runtimes; the portable updater excludes
Next-Trainer/extensions/so updating the main program does not uninstall Kohya. - MOSS-Image: Kohya remains the recommended/default engine. The current source also integrates optional KohakuMossTrainer and
anima_lorapaths; availability in portable packages depends on the package version. - AI Toolkit and Ideogram 4: AI Toolkit training and an Ideogram 4 LoRA entry are available as optional engines and are not force-installed.
- Plugin marketplace: the currently empty marketplace is hidden from users and skipped by default, so its unused dependencies are not downloaded.
- 3.3.1-dev portable package: a sub-2 GB lite package includes the base runtime, frontend, WD tagger model, and offline tokenizer assets without preinstalling Kohya.
Windows users: download the current Next-Trainer-v3.3.1-dev-lite.7z, extract it, and use the included launcher. It contains the base runtime, frontend, WD tagger model, and offline tokenizer assets. Kohya is not preinstalled, and the empty plugin marketplace is not downloaded; install a training engine from Settings → Training engines when needed. An NVIDIA GPU is required; see setup and package selection.
Release status: the recommended portable package is 3.3.1-dev. The larger preinstalled-Kohya package is distributed separately from GitHub Releases; the empty plugin marketplace is hidden from users and skipped by default during builds.
For the latest source version or Linux setup, follow run from source.
Source installation:
- Close the WebUI and stop active training jobs.
- Run
git statusin the project directory and commit or back up any local changes. - For the stable version, switch to
mainand rungit pull --ff-only. To follow the next development version, rungit fetch originand then switch todev. - Start again with
run_gui.baton Windows orbash run_gui.shon Linux. If the launcher reports changed dependencies, reruninstall-cn.ps1/setup_environment.pyas instructed; do not delete the project directory's user data.
Source updates change repository code only. The installed Kohya environment lives under extensions/kohya/; do not delete it manually. If the UI reports a broken Kohya environment after an upgrade, use Settings → Training engines → Kohya → Repair/Install so the current version can audit or repair it.
Portable package:
- Prefer
Update-Next-Trainer-Release.batfrom the package root. It downloads the latest Release and merges the program files. If GitHub Release downloads are unavailable, useUpdate-Next-Trainer.batfor a Git update. - Close the WebUI before updating. The updater preserves
Next-Trainer\extensions\, models, outputs, logs, training data, and configuration.Next-Trainer\extensions\kohya\is the independent Kohya environment and is excluded from the update merge. - Start again with
run_gui.bat. If the launcher reports changed dependencies, run the includedupdate\update_dependencies.bat; do not deletepython_embededorextensions\kohyafirst.
Old portable packages without an updater: do not extract a new 7z directly over the old directory. Close the application, copy the old Next-Trainer\extensions\kohya\ directory to a safe location, extract the new package into a new directory, then restore that directory to Next-Trainer\extensions\kohya\. Also migrate sd-models, output, logs, train, config, huggingface, and tagger-models. Launch from the new package. This keeps the program update separate from the Kohya environment and avoids losing the engine through an old overwrite workflow.
The lite package does not include Kohya. Install it from Settings → Training engines when needed. After installation, use the updater above and keep Next-Trainer\extensions\kohya\; updating the main program will not uninstall Kohya. See the portable update contract for the detailed rules.
Kohya · Anima Fast · Musubi · AI Toolkit · DiffSynth-Studio · KohakuMossTrainer · MOSS (anima_lora)
Install engines from Settings → Training engines, then select the appropriate engine on the training page. After upgrading from 3.1.1 or older, install Kohya once before running Kohya training. Available engines depend on your trainer version; not every model or feature supported upstream is integrated here.
For Qwen-Image-2.1, select DiffSynth-Studio → LoRA, then choose Text-to-image or Edit. Edit currently requires batch size 1 and can use gradient accumulation for a larger effective batch.
For MOSS-Image LoRA, use Kohya (recommended and selected by default); KohakuMossTrainer (high-performance LyCORIS trainer by KohakuBlueleaf) and MOSS (anima_lora) are also available as optional engines.
MOSS-Image · Qwen-Image-2.1 · Anima · SD 1.5 · SDXL · Flux · FLUX.2 Klein · Krea 2 · Ideogram 4
Training targets, hardware requirements, and optional engine installation vary by model. See the training tutorials below.
✅ = an integrated LoRA entry; — = no integrated entry. This table follows this branch's training modules, not the full capabilities of upstream engines. Model variants and hardware requirements still depend on the selected engine.
| Model | Kohya | Anima Fast | Musubi | AI Toolkit | DiffSynth-Studio | KohakuMossTrainer | MOSS (anima_lora) |
|---|---|---|---|---|---|---|---|
| MOSS-Image | ✅ | — | — | — | — | ✅ | ✅ |
| Qwen-Image-2.1 | — | — | — | ✅ | ✅ | — | — |
| Krea 2 | — | — | ✅ | ✅ | — | — | — |
| Anima | ✅ | ✅ | — | ✅ | — | — | — |
| FLUX.2 Klein | — | — | — | ✅ | — | — | — |
| Ideogram 4 | — | — | ✅ | ✅ | — | — | — |
| Flux | ✅ | — | — | ✅ | — | — | — |
| SDXL | ✅ | — | — | ✅ | — | — | — |
| SD 1.5 | ✅ | — | — | ✅ | — | — | — |
Kohya also provides full fine-tuning entries for Anima, SDXL, and SD 1.5. MOSS-Image defaults to Kohya, with KohakuMossTrainer and MOSS (anima_lora) as optional engines. Lumina has a Kohya LoRA entry accessible via ?model=lumina, but is hidden from the normal model picker.
The illustrations above are conceptual; their artwork is not a benchmark or a sample generated by each labelled model.
- MOSS-Image
- Qwen-Image-2.1
- Anima
- Anima Fast
- Krea 2 (Linux / multi-GPU)
- Ideogram 4 (Musubi)
- TOML / CLI
Interface tour · Tagger model setup · Training monitor · All documentation



