Tips and tricks

Here is a practical, up-to-date tutorial (March 2026) for training a LoRA using AUTOMATIC1111 WebUI + Kohya_ss (currently the most popular and reliable combination).

This guide focuses on Character LoRA training (face + body consistency) using SDXL base models such as Juggernaut XL or RealVisXL, which are widely used in 2026 for photorealistic virtual models.

Two Main Paths (Choose One)

Kohya_ss is generally better in 2026 (more parameters, better regularization, faster, more stable results), so this tutorial focuses on it.

Step 1 – Prepare Your Dataset (Most Important Part)

Goal: 20–80 high-quality images of one character (face + body visible in most shots).

Rules for a Good Dataset (Critical for Consistency)

• Variety is key: front face, ¾ view, profile, looking up/down, different expressions
• Full body, upper body, close-up face
• Different poses (standing, sitting, lying, looking back)
• Different lighting (soft, hard, indoor, golden hour)
• Different backgrounds (not too distracting)
• Same character in every image
• No other people in images
• Resolution: at least 512×512 (prefer 768×768 or 1024×1024)
• Mix portrait, square, landscape ratios
• Avoid heavy filters or extreme angles in all images

Sweet spot: 30–50 images for SDXL character LoRA.

Fast Dataset Creation Workflow

• Use WebUI + IP-Adapter FaceID + OpenPose
• Generate 40–60 images with seed variations + pose changes
• Select the best 30–50 (sharp, correct anatomy, consistent face/body)

Naming & Captioning (2026 Best Practice)

Style A – Filename Training (Simple & Effective)

zkw woman, front view, smiling.jpg
zkw woman, rear view, looking back.jpg
zkw woman, sitting, crop top.jpg

Trigger word example: zkw woman (rare 3–5 letters + class like woman/girl/person)

Style B – .txt Caption Files (More Control)

001.jpg
001.txt → zkw woman, front view, detailed face, photorealistic

Step 2 – Install Kohya_ss GUI

• Download latest release from GitHub (bmaltais/kohya_ss)
• Choose Windows installer (.exe) or portable version
• Extract / install
• Run kohya_gui.exe or kohya_gui.bat

Step 3 – Basic Training Settings (SDXL Character LoRA)

Folders Tab

• Image folder → dataset folder
• Output folder → e.g. F:\ai\my_loras\zkw_woman_v1
• Model folder → SDXL checkpoint (Juggernaut XL etc.)
• Resolution → 1024×1024 (or 768×1152 for portraits)
• Enable buckets → checked

Parameters Tab (Most Important)

• Preset → LoRA (SDXL) / Character LoRA
• Network Rank → 24–64 (start 32)
• Network Alpha → half or same as rank
• Conv Dim / Alpha → 4–16 (start 8)
• Learning Rate → 5e-5 (range 3e-5 to 1e-4)
• Text Encoder LR → 1e-5
• Unet LR → 5e-5
• Epochs → 15 (range 10–30)
• Batch Size → 2–3 (16GB VRAM safe)
• Save every N epochs → 2–5
• Trigger Word → must match filenames
• Max Token Length → 225
• Mixed Precision → fp16 or bf16
• Optimizer → Prodigy (recommended) or AdamW8bit

Advanced / Regularization

• Regularization images → optional
• Noise offset → 0.035–0.1

Training

• Click Start Training
• Estimated time: 30–120 minutes

Step 4 – After Training

• Best LoRA often around epoch 8–20
• File example: zkw_woman-000010.safetensors
• Move file to: models/Lora/
• Use in prompt: <lora:zkw_woman:0.7-0.9>, zkw woman, …

Quick Troubleshooting

This workflow (Kohya_ss + 30–50 images + IP-Adapter reference) produces excellent face + body consistency for virtual model training.