Train Custom AI Models with LoRA
LoRA (Low-Rank Adaptation) is a lightweight fine-tuning method that lets you teach AI models new concepts, characters, or styles using just a handful of reference images. Instead of retraining an entire model from scratch, LoRA adapts specific layers, making it faster, cheaper, and accessible to individual creators.
Try LoRA Training NowWhat is LoRA Training?
LoRA training works by injecting small trainable matrices into the frozen weights of a base model like FLUX or Stable Diffusion. This means you can add your own face, a consistent character, a product design, or an artistic style without affecting the model's core capabilities. Training typically takes 15 to 45 minutes depending on dataset size and settings. The result is a compact file (usually 10-200 MB) that can be loaded on top of any compatible base model to reproduce your trained concept on demand. Apatero handles the entire training pipeline in the cloud, so you do not need a powerful GPU or any technical setup.
Step-by-Step LoRA Training Workflow
Follow these steps to get the best results with lora training on Apatero
Prepare Your Dataset
Collect 15-30 high-quality images of your subject. Use consistent lighting, varied angles, and clean backgrounds. Crop images to focus on the subject and remove distractions. For character LoRAs, include close-ups, mid-shots, and full-body images.
Configure Training Parameters
Set your training steps (800-1500 for most subjects), learning rate (1e-4 is a solid default), and rank (16-32 for characters, 64+ for complex styles). Choose a trigger word that will activate your LoRA during generation.
Start the Training Run
Upload your dataset and launch the training job. Apatero runs training on cloud GPUs so you can close your browser and come back when it finishes. Training typically completes in 15-45 minutes depending on your settings.
Test and Iterate
Generate test images using your trigger word at different strengths (0.6 to 1.0). Check for likeness accuracy, style consistency, and any artifacts. If results need improvement, adjust your dataset or training parameters and retrain.
Tips for Better LoRA Training Results
Use these tips to get the most out of lora training on Apatero
Use consistent lighting across all training images for best results
Include 5-8 close-up shots and 5-8 full-body shots for character LoRAs
Remove watermarks, text overlays, and other distractions from training images
Start with a learning rate of 1e-4 and adjust down if you see overfitting
Use a unique, uncommon trigger word to avoid conflicts with existing model vocabulary
Train for 1000-1200 steps as a starting point, then increase if results are underfit
LoRA Training FAQ
Common questions about lora training on Apatero
How many images do I need to train a LoRA?
For most subjects, 15-30 high-quality images produce strong results. Character LoRAs work well with 20-25 images that show varied poses and angles. Style LoRAs may need 30-50 images to capture the full range of the aesthetic.
How long does LoRA training take on Apatero?
Training typically takes 15-45 minutes depending on your dataset size and training steps. Apatero runs training on cloud GPUs, so you do not need your own hardware.
Can I use my LoRA with different base models?
LoRAs are compatible with models in the same family. A LoRA trained on FLUX works with all FLUX variants. A Stable Diffusion LoRA works across SD 1.5, SDXL, and similar checkpoints.
What is the difference between LoRA rank 16 and rank 64?
Higher rank means the LoRA can learn more complex details but requires more VRAM and training time. Rank 16-32 works well for characters and simple concepts. Rank 64+ is better for complex artistic styles or multi-concept training.
Do I keep ownership of my trained LoRA?
Yes. Any LoRA you train on Apatero is yours. You can use it for personal or commercial projects, and your training data is never shared with other users.