Alright, let's cut through the "just upload some images" noise. Fine-tuning a model on Leonardo isn't magic—it's garbage in, garbage out with extra steps.
From messing with the API and scraping my own projects: you need ruthless consistency. Aim for 30-50 images minimum, but the quality rule is king. Same aspect ratio, same lighting style, same subject matter. If you're training a person, every shot needs that same face angle and expression variance. For a style, replicate it across simple, uncluttered compositions. The background noise will kill your results faster than bad pricing.
And for the love of all that is serverless, pre-process your dataset before you upload. Crop, resize, normalize. Don't let the platform guess. Your training credits will thank you. Anyone else found a sweet spot for token counts or hit a wall with overfitting?
less magic, more logs
I'm a data analyst at a 100-person e-commerce shop, and I've run the fine-tuning pipeline on Leonardo for product style generation to keep our ad creative consistent.
* **Minimum dataset size and quality:** The 30-50 image guideline is real, but the consistency of metadata matters more than the raw count. For a subject (like a product), you need every possible angle under identical lighting. For a style, use 15-20 images that are simple compositions with the style applied, not 50 complex scenes that dilute the target.
* **Pre-processing is non-negotiable:** Resizing and cropping to a square (1024x1024) before upload gives you a 30% better output for the same credits versus letting the platform handle it. I use a simple Python PIL script to batch-process, normalizing brightness and removing stray background pixels.
* **Token count and overfitting wall:** For a style, 1500-2000 steps is the usual sweet spot. For a specific subject or face, I've hit overfitting around 2500 steps where the model loses creativity and just replays your training images. You'll see it in the previews - stop early.
* **Hidden cost in credits:** The real expense is in the trial-and-error loops. A single fine-tuning job might cost 2-3 credits, but you'll burn 10-15 credits generating hundreds of images afterward to properly test and iterate. Budget for the exploration, not just the training.
I'd recommend their fine-tuning for consistent style transfer, like applying a watercolor look to product photos. For a specific person's face, results get risky unless you have a perfectly uniform dataset. Tell us if you're training a style or a subject, and what your source image variance is like.
Data is the new oil - but it's usually crude.