Training Tools

Customize AI models with your own data. WaveSpeed provides powerful cloud-based training tools to fine-tune foundation models like FLUX.1 and Stable Diffusion. Whether you need a specific character, art style, or product representation, train high-quality LoRAs and Dreambooth checkpoints in minutes without managing expensive hardware.

Start Training View Training Docs

The Model Training Process

From raw images to a deployable custom model in four simple steps.

Dataset Preparation

Upload 10-50 high-quality images (for LoRA) or 100-1000+ images (for full fine-tuning). Our automated pipeline handles resizing, cropping, and captioning. Smart deduplication removes near-identical images, and quality scoring flags blurry or low-resolution samples before training begins. Supports JPEG, PNG, and WebP formats.

Configuration

Choose your base model (FLUX.1, SDXL, or SD 1.5), training method (LoRA, DreamBooth, or Textual Inversion), and hyperparameters. Presets for common use cases (character, style, object) auto-configure learning rate, batch size, and training steps. Advanced users can fine-tune every parameter manually.

Cloud Training

Training runs on dedicated GPU instances (A100 / H100). LoRA training typically completes in 15-45 minutes; full DreamBooth runs take 1-3 hours. Real-time progress monitoring shows loss curves and sample generations at each checkpoint. Automatic early stopping prevents overfitting. Build on your trained models with LoRA Generation.

Testing & Deployment

Test your trained model instantly in the playground with various prompts and settings. Compare outputs side-by-side with the base model to verify quality. One-click deployment makes your model available via API and the WaveSpeed playground. Version control lets you roll back to any previous checkpoint.

What Can You Train?

Tailor the AI to your specific creative needs.

Brand-Consistent Characters

Train a LoRA on your mascot or brand character to generate unlimited variations in any pose, outfit, or setting. Perfect for marketing campaigns, stickers, and social media content at scale.

Product Photography

Fine-tune on 20-30 product images and generate photorealistic renders in any environment — studio white, lifestyle scenes, or seasonal themes — without physical photo shoots.

Art Style Transfer

Train on a curated set of artworks to replicate any visual style. Generate new illustrations that match the trained aesthetic, from anime to oil painting to pixel art. Use with LoRA Generation.

Face Personalization

Use DreamBooth to train on personal portraits and generate consistent face renderings across diverse styles and contexts. Ideal for avatars, profile images, and personalized greeting cards.

Q & A

Do I need a powerful computer to train models?

No. All training happens on WaveSpeed's cloud infrastructure. You can start a training job from a laptop or even a tablet; our servers handle the heavy lifting.

What is the difference between LoRA and Fine-tuning?

LoRA (Low-Rank Adaptation) is a faster, more efficient method that creates a small "adapter" file (MBs) to work with a base model. Full Fine-tuning updates the entire model (GBs). For most users, LoRA offers the best balance of quality and flexibility.

How many images do I need?

For a LoRA, we recommend 15-20 high-quality images for a face or object, and 30-50 images for a style. Quality matters more than quantity—ensure your images are clear and well-lit.

Can I train on the new FLUX.1 model?

Yes. We support FLUX.1 LoRA training. This allows for incredibly high-fidelity results, especially for photorealistic characters and text rendering.

Is my training data private?

Yes. Your uploaded datasets and the resulting trained models are private to your account by default. We do not use your private data to train our public foundation models.