I recommend you try out your existing workflow with my negative lora, set between -1 to -1.5.

I'd test it myself, but it looks like you're using comfy and I can only run flux in the cloud/there aren't any prompts.

I think the issue you have has to do with them poisoning nudity. The ablation I do tends to massively enhance the adherence and coherence of loras that fight with the poisoned safety training stacked on the OG model.

Greatful for the effort, but those nipples still look like peeled grapes being popped out the areola 😂. It is not always but it comes out like that most times.

Anyone have a guide or know how to add a LoRa like this to a pod on Runpod.io?

V2 is very very good.

Это база.

i mean, if we can't have an Apache Licensed version, it's kind useless...
Please, build one with the Schnell version.

Was this training run quantised ?

Nudes looks good! I'm seeing a few (3/50) images with the skin removed (except the head) showing muscles and fat.

I generated 500 images with no repeats of the problem. Just a Flux weirdness I think.

if anybody has a suggestion for a good (and I mean GOOD and DETAILED) nsfw capable captioner, please let me know!

I have been using a modified version of the Joy caption tool in a notebook that I copied, where Joy is given a different and contextual prompt for each image or subdirectory. Combining that with abliterated Llama models, followed by a keyword-replacement dictionary has been working well for me. Can share the method if you want.

I don't quite understand why we need to set up unique Trigger Words when we are improving Anatomy.