I Trained a LoRA on 8GB VRAM and It Looks Like Official Game Art
In this video I walk through my complete process for training Z-Image Turbo character LoRAs using AI Toolkit on an RTX 3070 8GB.
Covers: dataset preparation, captioning, config settings, the uint4 + 0.0002 LR breakthrough, and common mistakes to avoid.
LINKS:
→ AI Toolkit Easy Install: https://github.com/ostris/ai-toolkit
→ Z-Image Turbo model: https://huggingface.co/Tongyi-MAI/Z-I...
→ Training adapter v2: https://huggingface.co/ostris/zimage_...
→ Config file: https://pastebin.com/REPMRkzj
→ All of my LoRAs on Civitai Red are here: https://civitai.red/user/Winnougan/mo...
→ Ko-fi (early access + support): https://ko-fi.com/winnougan
→ Discord: / discord
→Z-Image Turbo ++ paper: https://huggingface.co/papers/2606.12575
→How to one click change JPEG to PNG sequentially with Magick in Powershell: https://pastebin.com/xRccNkM6
TIMESTAMPS:
0:00-4:55 Intro + results showcase
4:55-6:56 What is Z-Image Turbo?
6:56-9:49 Tools you need
9:49-16:55 Dataset preparation
16:55-19:55 Captioning the images in LM Studio
19:55-25:20 The config that changed everything
25:20-32:18 Dataset Configuration and PowerShell code
32:18-33:53 Common mistakes
33:53 Outro