Brand new model trained using synthetically-generated datasets from publicly-available and privately-trained models, combined with tagging using multimodal LLMs. I kinda love this thing. I hope you love it too.
This model has been tuned for excellent composition, aesthetics, and "soul." Characters come alive, scenery is dynamic, and there are a /ton/ of well-supported art styles. If ZBase can't make it, no one can ;-)
Please use LoRAs as low ratios (0.4-0.6) to accomplish softcore and hardcore NSFW, as the model has been made "flaccid" for general audience use. Erotica and nudes are fully doable (and quite good) with the base model however, so enjoy :)
Description
Fix some issues with artifacting and "image burn" due to improper balancing of LoRAs
Support for softcore and hardcore content /without/ LoRAs
Improved hair texture and body
Improved eye rendering
Brought back better styling for non-realistic/hyperrealistic prompts
This marks the start of the Beta line of models. There will be minimal changes to overall capability moving forward, the Beta series will be focused solely on fixing rendering problems.
If you find styles or content types that you cannot render /without/ introducing LoRAs, and you would like a LoRA for it, please leave a comment. Popularly requested LoRAs will be my next project.
FAQ
Comments (18)
Amazing model. I fell in love with it immediately, seriously!
I'm so glad you like it!!! Don't forget to leave a rating, and tell your friends <3
What is the "completely new method"? :)
Complicated and still under development, but basically I'm creating a ton of synthetic datasets and curating+tagging the best quality images using several multi-modal LLMs, then using a combination of relora and qlora to do the training, finally merging the results together using hand-tuned weights.
@dustydecapod interesting, how are you liking qlora?
the model gets results, though there is a lack of detail about what merges were used and whatnot, i cant really complain. Good work
This is not a merge.
Also, glad you like it! Please leave a review some time :)
For Alpha5, there is an error :
Base Model is SD 1.5
Whoops! Ty!
This model is absolutely incredible. It reminds me of many high quality late 1.5 ones. Most SDXL models get rough and muddy results, unless you put a lot of crap in the negative prompt. Not needed with this one. 10/10 from me.
Really glad you're enjoying it! Next release will be close to final, then I'll be releasing a series of LoRAs to go with.
This is one of the few SDXL models that don't get insanely grainy with depth ControlNets, excellent work!
Great to know, I didn't even do any testing with ControlNet.. xD
@dustydecapod Btw, I tested a bit and this model also works exceptionally well with the new Lightning LORAs, allowing for really quick generation at 8 steps
@banji Nice. Probably works well since I already trained it to be decent at 12-18 steps, 8 steps isn't a huge leap from there :)
@banji Civitai seems to have eaten my reply for some reason, but that's pretty awesome! The stock model already does very well at 12-18 steps for most prompts, doesn't surprise me that it's very compatible with lightning!
Really enjoying your model, seems to go well with all the loras I added as well ;)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.