Does uncensored mean it's abliteration heretic?
I'm currently facing the issue of prompt generation for Krea2 and my other "instruct heretic" model simply refusing basic NSFW terms.
Does this uncensored model have the same capabilities as an abliteration heretic one would have?
Or in other words, has this model also (some) removed constraints for text prompting without image input? Thanks
As a standalone model the results are largely uncensored, the instruct VL models have far less safety constraints then the new 3.5 or 3.6 models with think baked in. - In regards to your question you would need to test seed to seed to see what changes the LLM prediction has on the model itself. To my knowledge no DIT model has relied only solely on the safety filters built into the VL or LLM to guide the saftey filter of the diffusion or DIT model.