De-censor LoRAs Show Krea2 Turbo’s Built-in Filters Degrade SFW Image Quality for Expressions, Body Types, and More
English summary
A follow-up comparison tested Krea2 Turbo model with and without three different de-censor LoRAs. Two tiny LoRAs (about 200 bytes, changing only 2–3 weights) solely bypass the model's censorship filters without adding new styles or concepts. Results demonstrate that the built-in filters significantly nerf SFW image generation for facial expressions, bruises, body builds, emotions, and other natural features. The bypass LoRAs restore these lost capabilities by freeing concepts the model already possesses but underutilizes due to filtering. A trained SNOFS LoRA also removes filters but introduces additional styles, making it a less controlled comparison. The study uses fixed prompts and standardized parameters for fair evaluation.
Chinese summary
一项后续对比测试了Krea2 Turbo模型在未加及加上三种不同去审查LoRA时的表现。两个微型LoRA(约200字节,仅改变2-3个权重)仅禁用审查过滤器,不引入新风格或概念。结果表明,内置过滤器显著削弱了模型在面部表情、瘀伤、体型、情绪等SFW内容上的生成质量。去审查LoRA释放了模型已有但被过滤限制的概念,恢复了这些能力。训练型SNOFS LoRA虽也能去除过滤,但会额外引入风格,非公平比较。测试使用固定提示词和标准参数确保公平。
Key points
Krea2 Turbo's built-in safety filters severely degrade SFW image quality, affecting facial expressions, bruises, body builds, and emotions.
Krea2 Turbo内置的安全过滤器严重损害SFW图像质量,影响面部表情、瘀伤、体型和情绪。
Two ultra-small LoRAs (about 200 bytes, modifying only 2–3 weights) can bypass these filters without altering the model's style or concepts.
两个超小型LoRA(约200字节,仅修改2-3个权重)可绕过这些过滤器,而不改变模型风格或概念。
Test comparisons used fixed prompts and seeds, showing that filtered outputs lose realistic details like tiredness, bruises, and rage, which are restored by the bypass LoRAs.
测试使用固定提示词和种子,过滤后的输出丢失了疲惫、瘀伤、愤怒等真实细节,而去审查LoRA将其恢复。
The trained SNOFS LoRA also removes filters but introduces new textures and behaviors, making it a less fair comparison for isolating filter effects.
训练型SNOFS LoRA也能去除过滤,但引入了新的纹理和行为,因此对于单独评估过滤效果而言不够公平。