AI Toolkit vs OneTrainer: which LoRA trainer should you use?

Pick the LoRA trainer that lists the base model you train on. On 4 Oct 2026, AI Toolkit’s README listed Krea 2 and Z-Image Turbo, OneTrainer’s listed Z-Image but not Krea 2, and Kohya’s GUI listed neither. We use AI Toolkit because the course trains a Krea 2 character LoRA, and AI Toolkit runs as a web UI on a rented GPU. It’s the trainer the course uses to make a consistent AI character.

Affiliate note: RunPod and Fanvue links on this page are referral links. If you sign up through them, AI Empire may earn credits or a commission. You pay the same. How this works.

On this page
  1. Which base models does each trainer support?
  2. Why do we use AI Toolkit?
  3. When does OneTrainer make more sense?
  4. What about Kohya and musubi-tuner?
  5. Which trainer is fastest or gives the best faces?
  6. What can go wrong?

Which base models does each trainer support?

Each trainer only trains the base models its authors added, and new models land at different times. So start with the list. Below is what each project’s own README said on 4 Oct 2026, plus its interface. “Not listed” means the README didn’t mention it that day.

AI Toolkit OneTrainer Kohya’s GUI musubi-tuner
Made by Ostris Nerogar bmaltais, on kohya-ss sd-scripts kohya-ss
Interface Web UI (port 8675) and CLI Desktop GUI and CLI Web GUI (Gradio) and CLI Command line
Krea 2 Yes (Raw and Turbo) Not listed Not listed Yes (own docs page, marked experimental)
Z-Image / Z-Image Turbo Both Z-Image Not listed Z-Image
Qwen-Image Yes Yes Not listed Yes
FLUX.1 / SDXL Both Both Both FLUX.1 Kontext, FLUX.2
Wan video Wan 2.1 and 2.2 Not listed (lists Hunyuan Video) Not listed Wan 2.1 and 2.2

Sources: the READMEs of AI Toolkit, OneTrainer, Kohya’s GUI, sd-scripts and musubi-tuner, checked 4 Oct 2026. FluxGym, another name you’ll see, is a simple web UI for FLUX LoRAs that runs Kohya’s scripts underneath.

Why do we use AI Toolkit?

Three reasons. It lists Krea 2, the base model the course trains on. It runs as a web UI, so a rented pod and a browser tab are all you need. And it can be locked with a password (AI_TOOLKIT_AUTH), which matters on a pod anyone could reach by its URL.

On top of the official AI Toolkit image, our LoRA trainer template adds a few helpers: no Hugging Face account or token, a dataset .zip dropped into datasets/ unzips by itself, and everything lives on /workspace/aitk so a restart loses nothing. JupyterLab on port 8888 handles uploads and downloads. You set up the job yourself in AI Toolkit’s job screen.

Our current template job, for this base model only: Krea 2 (Raw), rank 64, 3,000 steps (2 to 3 hours), a save every 250 steps with 12 kept, on an RTX PRO 6000 with quantization and Low VRAM off. It’s ours, not a universal recipe.

Video 3 has every field in a table. We rent the card on RunPod, at $2.09 an hour on secure cloud (RunPod pricing, checked 4 Oct 2026).

The demo AI character generated with her LoRA at a beach club, white bikini.
A photo made with her LoRA, trained in AI Toolkit on Krea 2 and generated with Krea 2 Turbo.

When does OneTrainer make more sense?

When you train on your own PC with a desktop app, or need features its README lists: masked training, image augmentation, EMA and built-in captioning with BLIP or WD-1.4. Its code also has a RunPod cloud module. If your base model is Z-Image or Qwen Image, it lists both. For Krea 2, it listed nothing on 4 Oct 2026.

One warning on the built-in captioning. Automatic captions describe what they see, and that includes her face and hair. For a character LoRA, training captions should never describe her face or hair, so the model ties both to the trigger word. Our Dataset Maker writes captions that follow this rule; if you caption another way, check every file. The rule is explained in LoRA captioning.

What about Kohya and musubi-tuner?

Kohya’s GUI is a Gradio web front end for kohya-ss’s sd-scripts, with a long list of older and mid-generation models: SD 1.5, SDXL, SD3, Flux.1, Lumina Image 2.0, Anima. musubi-tuner is the same author’s newer, command-line trainer, and its documentation covers Krea 2, Z-Image, Qwen-Image and Wan. Neither is a beginner’s first click.

From their READMEs, checked 4 Oct 2026:

  • Kohya’s GUI can run locally or in the cloud (it documents Colab and RunPod; on free Colab, a web UI breaks Colab’s rules: see where to train a LoRA) and trains LoRA, LoHa, LoKr, DreamBooth and full fine-tunes. Its latest release was v26.0.0 on 9 July 2026.
  • musubi-tuner recommends 12 GB of VRAM or more for image training, 24 GB or more for video, and 64 GB of system RAM. It has no GUI in its README: you run scripts. Its Krea 2 page marks the support as experimental and, like the course, trains on Raw and generates with Turbo.

If you already know Kohya’s scripts and your model is on its list, there’s no reason to switch. If you’re comfortable with command-line scripts and want Krea 2 outside AI Toolkit, musubi-tuner documents it. If you’re starting from zero on Krea 2, a web UI saves you an evening.

Which trainer is fastest or gives the best faces?

We don’t know, and nobody can tell you without a fair test: the same dataset, base model, steps, GPU and seed on every trainer. We’ve only trained the course’s LoRAs in AI Toolkit, so we won’t put speed or quality numbers on the others. Pages that do usually show no method.

Whichever trainer you use, these decide her face:

  1. The dataset. One image that isn’t clearly her teaches a second face. Delete it before training.
  2. The captions. Scene only, never face or hair.
  3. The checkpoint you keep. Download the final file plus the 2,500 and 2,000-step saves, run the same prompt and seed on each, and keep the one that looks most like her. In video 3 that was the 2,500 file. How to test a LoRA has the full method.

What can go wrong?

  • The trainer doesn’t list your base model. Check the README before you set anything up. A LoRA only works with the base model it was trained on.
  • Out of memory on a 32 to 48 GB card. Our Krea 2 settings only fit with fp8 quantization and Low VRAM on, and the LoRA comes out worse. Train on 80 to 96 GB if you can.
  • The early checkpoints are gone. AI Toolkit keeps 4 saves by default. Raise Max Step Saves to Keep to 12 before you start a 3,000-step run.
  • Anyone with the pod URL can open the trainer. Set AI_TOOLKIT_AUTH and JUPYTER_PASSWORD before you deploy.
  • Auto-captions describe her face. Then the face sticks to the captions instead of the trigger word. Rewrite them as scene-only.
  • You compare trainers with different datasets. Then you’re comparing datasets. Change one thing at a time.
  • The bill keeps running. Download your LoRA, then stop and terminate the pod.

Questions people ask

Is OneTrainer better than AI Toolkit?
Neither is better in general. OneTrainer has features AI Toolkit's README doesn't list, like masked training and built-in image augmentation. AI Toolkit lists more of the newest base models, including Krea 2, and runs as a web UI. Pick by the base model you train on and where you train.
Can OneTrainer train a Krea 2 LoRA?
Its README didn't list Krea 2 when we checked on 4 Oct 2026. It listed Z-Image, Qwen Image, FLUX.1, Flux.2 and others. AI Toolkit and musubi-tuner both covered Krea 2 that day.
Is Kohya still worth using?
For the models it lists, yes: Kohya's GUI covers SD 1.5, SDXL, SD3, Flux.1, Lumina Image 2.0 and a few newer ones, and its last release was v26.0.0 on 9 July 2026. It didn't list Krea 2, Z-Image or Qwen-Image.
Do the trainers cost money?
No, all four are free to download from GitHub. What costs money is the GPU: your own card, or a rented one by the hour. We train on a rented RTX PRO 6000 at $2.09 an hour (secure cloud, checked 4 Oct 2026).
Which trainer is fastest?
We don't know, and we won't guess. A fair answer needs the same dataset, base model, steps, GPU and seed on every trainer. We've only trained the course's LoRAs in AI Toolkit.

Read next