I've been stress-testing the new XL models (Leonardo Diffusion XL, AlbedoBase XL) on larger canvases (think 1024x1024 and up) and I'm consistently getting bizarre, corrupted artifacts. We're not talking subtle texture glitches—I'm seeing full-on graphic tears, smeared patches that look like a GPU meltdown, and occasionally what I can only describe as "ghost limbs" or extra facial features materializing in the background.
The community hype around these models being "production ready" for high-res work seems... premature. My setup isn't the issue; I've ruled out VRAM limits and I'm using the same meticulous prompts that work flawlessly with the older SD 1.5-based models. The problem appears exacerbated when:
* Using a high "High Resolution" upscaler strength
* Any prompt complexity beyond simple subjects
* Generating in "Alchemy" mode at high resolutions
Before I start digging into whether this is a server-side quantization issue or a fundamental problem with the XL architecture on their platform, has anyone else pushed these models to their advertised limits and seen similar garbage output? I'm particularly interested in:
* The specific resolution and model combo where things break for you
* Whether you think this is a Leonardo implementation problem or inherent to the base XL models
* Any workarounds besides the obvious "downscale and then upscale elsewhere," which defeats the purpose.
The pricing on these tiers isn't trivial, and "experimental" artifacts aren't in the feature list.
trust but verify