NSFW edit
I noticed it censors slightly nsfw parts of the image, unprompted. Any way to fix it, in future updates? Or is this model serves strictly as sfw one?
find in google and use: krea 2 textfusion refusal reduction lora
I don't think that's an issue with the LoRA and more of a Krea 2 thing.
I use one that works pretty well for every type of generation, not just nsfw and I haven't noticed any random nsfw stuff when not prompted.
Here's a pastebin with three filter bypass LoRAs I found on civitai: https://pastebin.com/dxtPUqLf
Unsure if posting the actual links would get me in trouble for some reason so this is just safer 😅
Thanks @CeciliaXCIX . To clarify from my side: the LoRA is trained only on SFW data and I have no plans to add NSFW concepts, so it will not actively help there, and any filtering you are seeing is base Krea 2 behavior, not this LoRA. What you combine it with is up to you. I have added a short note on scope and responsible use to the model card.
A way easier way to avoid the nsfw issue is just to use the Krea conditioning revalance custom node. Its easy to use and has the math already in it. You can use it with anything krea related and it doesnt steal ram or vram
I don't know exactly how training a LoRA works yet, but it could be very useful to have the option to attach other NSFW LoRAs, or use some other method, to allow NSFW content.
So by default, it wouldn't contain any NSFW content, but if someone wants to use it, there could be an external way to enable it.
I don't know exactly how training a LoRA works yet, but it could be very useful to have the option to attach other NSFW LoRAs, or use some other method, to allow NSFW content.
Like i said bro, the conditioning rebalance node corrects filtration to allow ALL nsfw content. You literally dont need loras anymore for krea. Krea is already trained on nsfw it just has a safety filtration you bypass with the node.