Models with "filters" don't just block lewd content—they also nerf normal expressions.

Picking up from yesterday’s post—lots of people said it was useful, so I went ahead and did a detailed comparison between the original Krea2 Turbo and three uncensored LoRAs: two tiny LoRAs that are just a couple hundred bytes, tweaking a few weights purely for uncensoring, plus one properly trained LoRA.

Let me be clear upfront—this comparison isn’t about NSFW content. What I want to show is that while the model’s censorship blocks certain things, it also nerfs a ton of normal use cases, like facial expressions, bruises and injuries, body types, and emotions. For each set, I’m posting four images: the uncensored base without any LoRA, Krea2FilterBypass V2 (weight 10), V3 (weight 10), and SNOFS, the real trained LoRA (weight 1). One heads-up—uncensoring does make it easier to generate NSFW stuff, so use your own judgment. Also, comparing the real trained LoRA directly to the bare base isn’t totally fair, since its training introduced a bunch of other changes—can’t credit all of that to uncensoring alone.

A lot of people don’t realize that the censorship system is falsely flagging emojis.

A 200-byte LoRA? That’s insane.

Yeah, the bruising and scarring being smoothed out really messes with serious work.

Compared to those who obsess over control variables, OP at least has some sense.

Sure, it’s unbanned now, but keep an eye on the platform rules.

Yeah, training a LoRA with too many variables mixed in is a mess, totally agree.

Bookmarked

even legit needs like drawing scars or bruises get blanket-blocked, this over-censorship collateral damage has never really been fixed