System Requirements:
This guide requires at least 12GB Vram or Higher. I have not tested it with 8GB VRAM configurations.
Tested Hardware:
RTX 3060 12GB - Training works but is noticeably slower
RTX 5070 Ti 16GB - Smooth training experience with faster processing times
The performance difference between 12GB and 16GB VRAM is expected and primarily affects training speed rather than capability.
Note1: This guide does not cover LoRA training for Wan 2.2 While it’s technically possible using AIToolkit, it demands high VRAM. Since you’ll be using RunPod for Wan 2.2, it’s far better to follow the video tutorial—it’s simpler, clearer, and built for that setup.
Wan 2.2 Training Guide: Click Here
Note2: If you wanna use Ai-toolkit in Runpod you can use my colab notebook: ClickHere
This article may be a bit lengthy, but rest assured, it is well worth your time. I will also assume that you are already familiar with the process of installing Ai-ToolKit locally.
If you don't know how to install:
Install the prerequisites: Git, Python 3.11+, and Node.js v22+
Clone the repo:
git clone https://github.com/ostris/ai-toolkit.gitInstall Python dependencies:
pip install -r requirements.txtNavigate to the UI folder and install Node dependencies:
cd ui→npm installStart the UI:
npm run dev
STEP 1: Dataset Preparation
Images: 200+ recommended, 50 minimum — clear, sharp, varied angles/lighting
Part A — JoyCaption Setup (optimized version by me) for Image Caption
git clone https://github.com/official-imvoiid/Joycaptioncd Joycaption
Run in order: GetConda.bat → SetEnv.bat → InstallRequirements.bat → StartTextCaptioner.bat
Bulk Tab, select all images
Type/Select what you want the model to tag/focus on
Run captioning
Download the caption ZIP at the end
Unzip it → move all
.txtfiles into the same folder as your images
Part B — Trigger Word Adder
https://github.com/official-imvoiid/Random/blob/main/TriggerWordAdder.py
No pip install needed — pure Python. Just run TriggerWordAdder.py and paste the path of you folder having images and captions/tags ⚠️ Critical: Type trigger word as voiid, (with comma+space)
✅
voiid, other-tags❌
voiidother-tags
STEP 2: Pre-Requisite
Make Sure to move this folder containing images & caption to <Path-To-Your-Ai-Toolkit>\datasets And also to apply following settings
STEP 3: Actual Training
From Here the real training process begins, see these settings [also given as pdf in attachments]
Parameters [New Job]:
1. JOB Section
Training Name = Your LoRA name
Trigger Word = Same word as used in your captions
2. TARGET Section (CHANGE) Set as shown: Linear Rank: 40, Conv Rank: 20
3. SAVE Section (CHANGE) Data Type: BF16, Save Every: 500, Max Step Saves: 20
4. TRAINING Section
Batch Size: 1/2 (use 2+ only if VRAM > 16GB)
Gradient Accumulation: Leave at 1 (no need to touch)
Steps: Calculate Shown below ↓
Example:
5. DATASET Section
Click Dataset → select your character's dataset
Num Repeats:
~50 images → set to 2 or 3
100+ or 200+ images → set to 1
6. SAMPLE Section (CHANGE) Fill in a test prompt under Sample Prompts so you can monitor training output visually.
STEP 4: Start Training / Pause / Resume / Edit
Note: If you remove the character output folder from \Ai-Toolkit\Output, or delete important files, you will not be able to resume or edit training — it will simply restart from 0. You can, however, edit .safetensors files directly (e.g., kurosaki-500.safetensors or kurosaki.safetensors).
Training Locally and Enjoy Lora
-> Q/A
Q1: What tools can I use if I don't have enough VRAM?
A: If your GPU lacks sufficient VRAM, you have two excellent cloud-based alternatives. Use RunPod for LoRA and Checkpoint training, which provides powerful GPU instances on-demand. For running ComfyUI workflows in the cloud, use GoogleColab instead. You can get the Jupyter notebook for GoogleColab at my Github. Both options let you train models and generate images without investing in expensive hardware.