Hey everyone, ran into a frustrating issue with my Leonardo training last night and hoping someone here has cracked this. I'm trying to create a model for a personal project—some custom character art—and my source images look crisp and clean on my machine. But once I upload them to Leonardo and start the training, the preview looks terrible: pixelated, weird artifacts, almost like they've been heavily compressed.
I'm coming from a cloud/observability background, so my first thought was to check the specs like I would with a cloud workload. I've been using PNGs, 512x512 resolution, trying to follow the guidelines. The images aren't small files, either.
Here's a quick rundown of what I've tried:
- Exported from my graphics tool at exactly 512x512.
- Tried both PNG and JPG (high quality, low compression).
- Verified there's no embedded color profile weirdness (sRGB).
- Even upscaled a set to 1024x1024 before uploading, just to see.
The result is always the same degraded look in the training preview. It's killing the potential output quality before I even start.
Is there a hidden pre-processing step I'm missing? A specific format they prefer that isn't clearly documented? Maybe something about the background? I've seen great models from others, so I know it's possible. Feels like I'm getting hit with a "lossy" transformation somewhere in their pipeline.
Any insights or workflows that have worked for you would be a huge help. This feels like trying to debug a metric ingestion problem where your high-res logs come out as blurry aggregates on the other side.
cost first, then scale
Oh man, I feel this. I had the exact same heart-sinking moment during my first training run. My images looked perfect locally, then like a blurry mess in the preview. I was sure I'd messed up the whole dataset.
What finally worked for me was a bit counterintuitive. I read somewhere that the platform might be doing its own resizing if it detects odd dimensions, even if you're at 512x512. I started making sure my images had canvas dimensions that were perfectly square - like, literally checking the pixel grid to see if I had any stray transparent pixels pushing the bounds. After I did a really strict crop on everything, the preview looked a lot closer to my originals.
Have you double-checked the actual canvas size versus just the exported resolution? That tiny difference drove me crazy for a day.