Everyone's buzzing about SD3's "realism," so I ran it head-to-head against Leonardo's Photoreal model on the same prompts. The hype is, predictably, overblown.
SD3 Medium, even with a good prompt, still has that telltale "AI gloss" on skin and a weirdly perfect, almost sterile composition. Leonardo Photoreal nails pores, subtle skin textures, and realistic lighting flaws more consistently. For example, prompting "a tired barista at dawn, slight stubble, under fluorescent lights" gave me:
- **SD3:** A handsome, clean-shaven guy with perfect uniform, studio lighting.
- **Leonardo:** Actual stubble shadows, slightly oily skin, the harsh glow of a real fluorescent tube.
The kicker? SD3 requires meticulous negative prompting to avoid its own artistic quirks. Leonardo's model, being fine-tuned specifically for this, gets there in one or two gens. Here's the kind of negative prompt SD3 needed just to look normal:
```
(worst quality, low quality, normal quality:1.4), (watermark), signature, text, cartoon, 3d, (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy
```
Meanwhile, Leonardo's output was usable with just the base prompt. If you're after actual photorealism for a project and not just playing with the newest shiny model, the specialized tool still wins. SD3 is impressive, but it's a jack-of-all-trades being paraded as a master of one.
prove it to me