Decensoring image models does more than generate NSFW

Krea 2 is a new, very capable open-weights image gen model. It runs on commodity hardware, super easy to use, and produces very good-looking results by default.

There are, however, complaints about prompt adherence. You ask for one thing, you get something slightly (but not completely) different. You'll get a beautiful image, but not exactly what you prompted.

For most people, that's good enough, or even preferable - not everyone is a nerd who devotes their life to learning good prompting techniques.

¶There is an easy solution, from an unlikely tool

Word is that Krea 2's guardrails affect your image gen significantly, even when NOT generating NSFW images. To test this hypothesis, we'll use a "bypass LoRA".

LoRA (Low-Rank Adaptation) is a fine-tuning technique that allows AI image models to be customized for specific styles, subjects, or concepts without retraining the entire base model. These things are tiny compared to the model - usually 100-500 MB.

Specifically, we'll use "FilterBypass V3" - a faux-LoRA (just 160 bytes).

Unlike a real LoRA, all it does is tweak a few of Krea 2's internal safety settings instead of teaching it a new style or character. It leaves basically everything else unchanged, so the model behaves like normal Krea 2, just less likely to trigger its built-in safety filter.

¶Trigger warning

If you're a prude and the word "cleavage" offends you, turn back right now and never come back. For you, here's the tl;dr:

• Krea 2 creates beautiful images by default, in every conceivable art style
• The pretty pictures usually come at the expense of prompt adherence - the guardrails that stop you from generating NSFW also negatively affect how the model renders the images
• This can be solved by skipping the built-in guardrails. Though you need to be more thorough and explicit with what you want, or you'll get un-pretty pictures
• However, "decensoring" too much ends up censoring the model (even more than vanilla!) and degrades prompt adherence

The rest of you, read on.

¶The prompt

This is the prompt we'll use throughout the experiment. The key words here are red and black palette, back of hand, flame motifs, spider lilies, and you guessed it, cleavage.

anime-style adult female portrait, dark red and black palette, head slightly tilted backward, prominent nose, hyperdetailed glowing red eyes, threatening smile, dark hair with radiating golden hairpins, chin resting on back of hand, off-shoulder kimono with skull is hanging down from her shoulders, revealing cleavage, visible shoulders, flame motifs, foreground red spider lilies, background maple leaves, dark circular framing, grunge texture

background has deep cobalt blue and black palette, clean black outlines

¶Before we start

Here's the picture of an actual spider lily. Remember this, it will be important later:

¶Strength 0 (no bypass LoRA)

The "vanilla" Krea 2 model (no decensoring) gives us this:

¶Observations

Overall, nice-looking picture. A lot to like:

• Maple leaves look fine, the spider lilies look like spider lilies (again, it'll be important later)
• The skin is red, which is not what I expected, but it makes sense because the prompt mentions the image has a "dark red and black palette"
• The kimono really is off-shoulder and it's an okay-ish amount of cleavage for a stock model. GPT or Nano Banana will kick you out for a ToS violation for a lot less

What isn't great:

• The prompt mentioned "chin resting on back of hand". Her chin is resting on her hand, but not on the back (this will change soon)
• The hairpins are going through her head. Ouch ☠️

¶Strength 1

Now, let's introduce the LoRA at "Strength 1". This has some interesting effects on the image.

Each LoRA has a different meaning of "strength". Some work great at 1.0, some at 4.0, some at even 20+. Others will mess your image up if you use more than 0.5. TL;DR: More isn't always better, less isn't always worse.

¶Observations

The changes are mostly positive:

• Her chin is now actually resting (somewhat) on the back of her hand
• The cleavage is more prominent (big surprise)
• It added a bun to her hair, as it should! The hairpins are not going through her skull anymore 🎉
• Color palette intact, the spider lilies look like spider lilies

Overall, better prompt adherence, and less censorship. Things get better from here.

¶Strength 2

At Strength 2, we're finally getting somewhere:

¶Observations

This one is really different:

• The skin color changed! This is the color I expected the skin to be from the beginning

That said, the red skin wasn't a bad assumption by the vanilla model. We did mention "red and black palette" after all

• The cleavage is a lot more prominent. Objective achieved!

However, there's a very subtle regression right now which will get more obvious later: the number of stamens in the spider lilies reduced! It still looks like a spider lily, but slightly less so.

Overall, this is the best strength level for this kind of generation. Things get worse from here.

¶Strength 2: An aside

If you want the skin to be red, just add "red skin" to the prompt and it'll oblige. I don't, so I didn't bother.

¶Strength 4

Strength 4 is the start of very real degradation. It looks similar to strength 2, BUT, the model has started to drift and "misunderstand" the prompt.

¶Observations

There are three big regressions:

• The "spider lilies" keep losing more of their character - fewer stamens, wider petals. They look more and more like regular lilies
• Another big one: after reading "spider" in spider lilies, the model put an actual spider on her hair. This was not the intention of the prompt!
• It also brought back an issue without the LoRA: the woman lost the bun and the hairpins are going through her skull again ☠️. Ouch

The elephant in the room

• Somehow the model is doing more, not less censorship. The character has started to turn sideways, and the cleavage is less visible than ever

None of these issues are one-offs! As we increase the strength of this LoRA, we get to see less and less of the goods, which, most would argue, destroys the original selling point of the LoRA.

¶Strength 6

At Strength 6, it's officially over. The character has turned around even more and the cleavage is nowhere to be seen.

The woke has taken over and the West has fallen.

¶Observations

It's still a pretty picture, but prompt-wise, nearly everything is a regression:

• The "flame motifs" in the prompt are not just motifs anymore. There's an actual fire that's burning on the skull
• The spider is still on her hair and the hairpins are still going through her head
• The flowers look nothing like spider lilies anymore - they're regular lilies now

¶Verdict

I've done more experiments (that I'll publish soon), mostly without any NSFW elements, and these are my conclusions:

• The vanilla, no-LoRA version is a nice starting point if all you want is good looking pictures
• Overall, depending on the prompt and LoRA strength, you win some, you lose some
• Strength 2 is preferable for most Krea 2 prompts

That said, every art style and prompt has somewhat different results, so experiment and find what works for you. This applies to everything in the AI world.