Virtual product
{{ variable.name }}
Select products to buy together. Choose each variant and quantity.
Quantities are the totals for this purchase. Saving combines the items into one bundle and recalculates the discount.
{{ bundleConfig.name }} Main product
Variant and quantity follow your selection above · × {{ quantity }}
{{ bundleStockLabel(product) }}
{{ bundleStockError(product, quantity) }}
{{ bundleLimitError(product, quantity) || bundleLimitLabel(product) }}
{{ item.name }}
{{ bundleStockLabel(bundleSku(item)) }}
{{ bundleStockError(bundleSku(item), item.quantity) }}
Check the selected variants, quantities and stock.
{{ (item.checked && bundleLimitError(bundleSku(item), item.quantity)) || bundleLimitLabel(bundleSku(item)) }}
The main product is excluded. Multiple units of the same product count as one style.
All complementary products have been deselected. Saving will change this to a separate purchase of the main product, without the bundle discount.
{{ bundleSelectedLabel }} Highest eligible discount applied
{{ bundleSelectionError || bundleError }}
Unable to calculate. Resolve the message above first.
Based on your current selection. Final discounts are confirmed at checkout.
Quantity: {{ quantity * bundleSetQuantity }}
Sets: {{ bundleSetQuantity }}
Customize the main product in your cart or at checkout. Customization fees are charged separately.
{{ bundleError }}
Loading…
A complete step-by-step guide to training custom LoRA models for Z-Image Turbo — the ultra-fast 6B parameter model that generates images in just 9 seconds. Learn the entire workflow from dataset preparation to the critical model conversion step that ensures ComfyUI compatibility.
Product Type: Digital Tutorial (HTML + Markdown)
Why Z-Image Turbo?
- Ultra-fast inference: 9-second image generation (vs 30+ seconds for Flux)
- Lower VRAM requirements: Runs on 12GB GPUs with optimization
- Photorealistic quality: Excellent results for portraits and realistic scenes
- LoRA compatible: Train custom characters, styles, and concepts
Tutorial Structure (6 Chapters):
【Chapter 1: Environment & Hardware】
- GPU requirements: NVIDIA 12GB+ VRAM recommended
- Storage: 50GB+ SSD space for models and cache
- AI Toolkit installation with path requirements (no Chinese/spaces!)
【Chapter 2: Dataset Preparation】
- Image specifications: 20-50 images, 1024x1024 resolution
- Captioning workflow with TT Tool (Image Tagging Tool)
- Trigger word setup: Critical for LoRA activation
- File structure: Matching .png/.jpg with .txt caption files
【Chapter 3: Configuration (YAML)】
- Complete parameter breakdown with explanations
- LoRA rank settings: 16 or 32 for beginners
- Low VRAM mode: Essential settings for 12GB/16GB GPUs
- Training steps: 2000-4000 recommended range
【Chapter 4: Training Process】
- Step-by-step execution guide
- Latent caching explanation (why first steps seem slow)
- Time estimates: ~4.5 hours for 3000 steps on 4070Ti
- Progress monitoring and Loss value interpretation
【Chapter 5: Model Conversion (CRITICAL!)】
- Why raw AI Toolkit outputs don't work in ComfyUI
- Using the included converter tool
- Format compatibility fixes for standard safetensors
- File size optimization after conversion
【Chapter 6: Testing & Application】
- ComfyUI workflow setup for Z-Image + LoRA
- Weight recommendations: Start at 0.8
- Trigger word usage in prompts
- Quality assessment and iteration
Bonus: FAQ & Troubleshooting
- FileNotFoundError: Path issues with Chinese characters
- CUDA Out of Memory: VRAM optimization checklist
- Noise/garbled output: Model conversion reminder
- Overfitting vs underfitting: Step count guidance
Technical Requirements:
- NVIDIA GPU with 12GB+ VRAM (16GB+ recommended)
- Windows with AI Toolkit integration package
- 50GB+ free storage space
- ComfyUI for testing trained models
Package Contents:
- Z_Image Basic Alchemy Explanation.html (formatted tutorial)
- Z_Image Basic Alchemy Explanation.md (markdown source)
Who This Is For:
- Artists wanting to train custom character LoRAs
- Developers building personalized image generation models
- Commercial users needing brand-specific AI outputs
- Anyone seeking faster alternatives to Flux for LoRA training