Train custom LoRAs with ostris AI-Toolkit

AI-Toolkit LoRA Trainer (WAN 2.2 & Z-Image)

Overview

AI-Toolkit by ostris is an MIT-licensed trainer for finetuning diffusion models. It is a standalone trainer with its own web UI, not a ComfyUI custom node. It runs a Node.js UI front end over a Python (run.py) training backend and trains LoRAs for many model families. This skill covers the WAN 2.2 / 2.1 video models and Z-Image (Turbo & Base).

• Repo: https://github.com/ostris/ai-toolkit (cloned by the installers).
• Backend: python run.py config/<job>.yml. UI: a Node.js app under ui/ that schedules and monitors jobs. You do not have to keep the UI open while a job runs.
• Output: a standard .safetensors LoRA you drop into ComfyUI models/loras/ and load with LoraLoaderModelOnly.

Best for:

• WAN LoRAs. A person or character, an art style, or a specific camera or video motion, trained from image or video clip datasets. For using WAN see wan-t2v-video / wan-flf-video.
• Z-Image LoRAs. Fast, very low-VRAM image LoRAs (faces, characters, outfits, styles) on the 6B Z-Image base/turbo. For using Z-Image see z-image-base / z-image-turbo, and the z-image-xy-plot pack to compare trained LoRAs.

For low-VRAM anime image LoRAs on a different stack (kohya sd-scripts), see the sibling anima-lora-trainer.

Two LoRA kinds for WAN. A WAN image LoRA trains on still images; it is cheaper (~24GB-class) and suits identity or style. A WAN video LoRA trains on short clips; it is heavier, best run on cloud, and suits motion. Z-Image is image-only.

Install

The installer comes in two generations. Both clone ostris/ai-toolkit, set up Torch for your GPU, and launch the web UI. Put it in a folder whose full path has no spaces (e.g. C:\AI-Toolkit).

• V1, AI-TOOLKIT_AUTO_INSTALL.bat, expects Git, Python 3.10.x, and Node 18+ already in PATH.
• V2, AI-TOOLKIT_AUTO_INSTALL-V2.bat (recommended), uses an embedded Python 3.10.11, auto-installs Git and Node, builds a clean PATH without your system Python, and adds aggressive pip/curl retries. It has far fewer prerequisites and fails less often. The Z-Image Turbo LoRA training release used it.

Both are CUDA-aware and select the Torch wheel by GPU generation:

Choice	GPU	CUDA	Torch index	Torch packages
1	RTX 50-series (Blackwell)	12.8	https://download.pytorch.org/whl/cu128	torch==2.7.0 torchvision==0.22.0
2	RTX 40 / 30 / 20 and older	12.6	https://download.pytorch.org/whl/cu126	torch==2.7.0 torchvision==0.22.0

Each then clones ostris/ai-toolkit, downloads two launcher scripts (LAUNCHER-TOOLKIT.bat, SECURE_LAUNCHER-TOOLKIT.bat, from https://huggingface.co/Aitrepreneur/FLX/resolve/main/), makes the venv, installs Torch from the chosen index, runs pip install -r requirements.txt, then cd ui && npm run build_and_start.

RunPod / Linux — AI-TOOLKIT_AUTO_INSTALL-RUNPOD.sh (and -V2.sh)

Installs into the persistent volume /workspace/ai-toolkit. It is idempotent; a re-run just relaunches the UI. Use RunPod's PyTorch 2.8.0 template and a 100GB disk. It installs apt deps, clones the repo, makes a venv, installs Torch (torchaudio included), installs nvm + Node 22, then builds and starts the UI.

Choice	GPU	Stream	Torch spec
1	RTX 5000-series (Blackwell)	cu128	torch==2.7.0+cu128 torchvision==0.22.0+cu128 torchaudio==2.7.0+cu128
2	Ada / Hopper / Ampere, older	cu126	torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0

The UI listens on 8675 and Jupyter on 8888. Set AI_TOOLKIT_AUTH (UI password) before launch. Reach it at https://${RUNPOD_POD_ID}-8675.proxy.runpod.net. Use an RTX 4090/5090 for image (WAN t2i/t2v, Z-Image) LoRAs and an RTX 6000 Pro (Blackwell) for heavy WAN video, high-res, or high-rank jobs.

Launching the web UI

• On Windows, run LAUNCHER-TOOLKIT.bat (local) or SECURE_LAUNCHER-TOOLKIT.bat (password-protected) from the ai-toolkit folder.
• On RunPod, rerun the .sh. It detects the install and starts the UI on :8675.

In the UI, create a Job, point it at a dataset folder, pick the model (WAN variant or Z-Image), set params, and start. Jobs run in the Python backend, so you can close the browser. To bypass the UI, copy a config/examples/*.yml, edit it, and run python run.py config/<job>.yml.

Dataset preparation

AI-Toolkit pairs each sample with a same-basename .txt caption and auto-resizes/buckets aspect ratios (no pre-cropping).

Image LoRA (WAN identity/style, or Z-Image)

my_dataset/
  001.png  001.txt
  002.jpg  002.txt

• Captions are natural language. Include a unique trigger word for a person or character.
• Use about 15 to 40 varied images for a person, more for a broad style.

Video LoRA (WAN motion only)

Short clips plus a .txt per clip; caption the motion or camera move. Set per-clip frames via the job's num_frames (e.g. 81). This is markedly heavier, so prefer cloud GPUs.

Key training params

WAN 2.2

WAN 2.2 14B is a Mixture-of-Experts with a high-noise expert (structure/motion) and a low-noise expert (detail). AI-Toolkit trains both via Multi-stage.

Param	Default	Notes
Linear rank / dim	16	16 simple; 16–32 complex/cinematic
Learning rate	5e-5 (identity)	7e-5–1e-4 style; high LR → plasticky skin
Steps	1500–2500	stop before overbaking
Resolution	512 (or 768)	bucketed; 768 costs more VRAM
num_frames (video)	81	per-clip frame count
Multi-stage	High + Low = ON	trains both experts
Switch Every	10	raise to 20–50 if offload swapping is slow
Optimizer / Quant	AdamW8bit / 4-bit ARA or float8	fits 14B on consumer cards

Z-Image (Turbo & Base)

Z-Image is a ~6B single-stream model with no hi/lo multi-stage. Leave Multi-stage OFF; you train one model. It is the lightest target here. The headline of the Z-Image releases is training on very low VRAM.

Param	Starting point	Notes
Linear rank / dim	16–32	32 for detailed characters/styles
Learning rate	1e-4	lower (5e-5) for tighter identity
Steps	1500–3000	dataset-dependent
Resolution	768 (or 1024)	Z-Image's native range
Multi-stage	OFF	single-stream model, not WAN's MoE
Optimizer / Quant	AdamW8bit / float8	enables sub-12GB training

Train on Base, deploy anywhere. Z-Image Base is the finetuning-friendly model; a LoRA trained on Base generally applies to the Turbo workflow too. Use the z-image-xy-plot pack to grid-compare your trained LoRAs.

The param tables are aggregated starting points from community and training-guide sources, not read from the repo's config/examples/*.yml. Open the actual WAN / Z-Image example config in your clone and tune. See "Unverified".

VRAM / GPU guidance

• Z-Image image LoRA is the lightest. It trains on modest consumer GPUs with quantization (the releases describe very-low-VRAM training); a 4090 is comfortable, and smaller cards work with float8 at 512 to 768 res.
• WAN image LoRA (t2i/t2v) needs 24GB+ locally with quantization. Below that, use RunPod.
• WAN video LoRA, high res, or high rank is heavier. Use cloud (RTX 5090, or RTX 6000 Pro Blackwell / H100).
• Memory savers: quantization, batch size 1, 512 res, and (WAN) raising Switch Every.

Using the trained LoRA in ComfyUI

1. Copy <your_lora>.safetensors into ComfyUI models/loras/.
2. Load with LoraLoaderModelOnly:
   • WAN 2.2 is dual hi/lo. Apply the LoRA to both the HighNoise and LowNoise model branches (like lightning/concept LoRAs in wan-t2v-video). Typical strength 0.5 to 1.0.
   • Z-Image is a single model. Use one LoraLoaderModelOnly on the Z-Image model path (see the z-image-base / z-image-turbo packs). Strength 0.7 to 1.0.

   { "class_type": "LoraLoaderModelOnly",
     "inputs": { "model": ["<base_model>", 0],
                 "lora_name": "<your_lora>.safetensors",
                 "strength_model": 1.0 } }
   

3. Prompt using the trigger word or caption style you trained with. For WAN motion LoRAs, describe the same camera or motion.

Troubleshooting

• No module named 'torchaudio' when starting a job (AI-Toolkit). The venv's Torch stack is mismatched. Activate the AI-Toolkit venv (venv\Scripts\activate), then pip uninstall torch torchaudio torchvision -y and pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 (or your CUDA's index). This only affects the AI-Toolkit install, not ComfyUI.
• self and mat2 must have the same dtype (ComfyUI-WanVideoWrapper, WAN usage). Re-clone ComfyUI-WanVideoWrapper in custom_nodes/ and reinstall its requirements.txt, then restart ComfyUI.
• 5000-series (Blackwell) onnxruntime "QuickGelu" / CUDA error. pip install onnxruntime==1.20.1 in the affected venv.
• Pascal/Maxwell GPUs (GTX 9xx/10xx). Recent Torch (cu128/cu130) dropped them. Reinstall the cu126 Torch build into the venv.
• Path with spaces (Windows). Keep the install path space-free or the build/launch fails.
• OOM during training. Quantization (4-bit ARA / float8), 512 res, batch 1, (WAN) raise Switch Every, or a bigger RunPod GPU.
• RunPod UI won't load / asks for a password. Confirm AI_TOOLKIT_AUTH is set and you're on the 8675 proxy URL.

Unverified / verify before relying

• The param tables (both WAN and Z-Image) are synthesized starting points, not read from the repo's config/examples/*.yml. Open the actual example config in your clone and adjust.
• The release notes describe the Z-Image training VRAM floor only qualitatively ("very low VRAM"). Confirm against your card; quantization plus 512 to 768 res is the lever.
• The Windows UI port is whatever the launcher binds (the installer doesn't print it; check the launcher window). RunPod 8675/8888 are per the template.
• The launcher .bat files are downloaded from a third-party HuggingFace repo (Aitrepreneur/FLX); review before running on a security-sensitive machine.
• Model weights are fetched at job time by AI-Toolkit/HF, not by the installer. Confirm the model selector lists your target WAN variant or Z-Image model before a long run.

Sources

• Official: https://github.com/ostris/ai-toolkit
• Empirical: Windows/RunPod installer steps and VRAM notes from the pack installers, not the vendor's training guide.