Merged Krea 2 turbo models with carefully weighted selections of over a dozen LoRAs, all of which are openly available on this site.
Steps: 5-8
Samples were generated with 5 steps, using the same seed for each pair of matching prompts between the models.
Due to hardware limitations I can only work with FP8 format. If someone wants an INT8 or another quant, I can try to make a conversion but won't be able to test it.
Adding some images generated with additional LoRAs on top of Nulparity. It looks like with most LoRAs it needs a couple more steps, these examples were made with 8 steps.
Description
Mostly realistic but a kind of polished look on average; not very gritty, grainy or grimy by default. Prompt adherence is quite good and doesn't make everything NSFW. Works fairly well with additional LoRAs for various concepts and anatomy adjustments.
FAQ
Comments (5)
English is not my native language, so I used a translator. I hope this won't cause any reading difficulties for the author. This model runs well on my low-memory graphics card, and the cue words are quite compliant, but it seems to simplify the background (unless it's directly described in the cue words). Thank you for everything you've done, and I look forward to your future works.
No problem. You are probably right about the backgrounds, the model is more focused on features on the foreground so it makes sense that the level of detail is decreased in other areas because attention is finite.
Sourparity is amazing! It's fast, stable, versatile and completely uncensored. Thank you for this checkpoint / model! I recommend using ER-SDE Simple and 8 steps.
Thanks, glad you like it! The cobra you posted looks really crisp, nice to see the model can still do other than human subjects too. I hadn't even tried other samplers but er_sde does seem to work pretty well.
See, I think nulparity is where it's at. Good models both.