Spent last weekend downloading every "perfect portrait" LoRA I could find. Most were garbage. Overfitted, weird artifacts, or just trained on three images of the same influencer.
The only one that consistently gives me usable, varied, natural-looking faces is **epi_noiseoffset2**. Don't let the name fool you.
Why it works:
- It's not a person-specific LoRA. It's trained on a noise offset technique, so it improves skin texture and lighting across the board.
- Works with multiple base models (SD 1.5, SDXL) without breaking them.
- No more plastic-y skin or that uncanny valley "LoRA look" you get from over-trained checkpoints.
Weight around 0.4-0.6 is the sweet spot. Higher and you might get too much grain.
Everything else felt like a gimmick. This one actually improves the fundamentals. Go find it on Civitai. The rest are just clutter 😒
CRM is a means, not an end.
Nice find. I've been down the same rabbit hole, and you're right - most portrait LoRAs are way too specific or just bake in the same flaws from their training set.
I'd add that the noise offset approach works because it's essentially a form of fine-tuning the denoising process itself, rather than injecting a new concept. That's why it doesn't break composition and plays well with different models. It's more like a universal "better rendering" patch.
The weight range you mentioned is key. I've seen people slam it up to 0.8 and then complain about the grain, missing the point entirely. It's a subtlety tweak, not a sledgehammer.
Latency is the enemy, but consistency is the goal.