Z Image Turbo Versions
Edit: This is possibly going to be the last of this model range, and is very much ready for use as is. Making this has been a great learning experience, and a lot of what's been learned here will be carried forward into the next model.
I hate asking, but if you can spare some buzz for training, or add anything you generate with it to the gallery here, it'll be appreciated.
The current optimum (for my taste) generation parameters are 4-6 Steps, CFG 1.0, using DPM++ 2s a RF with Bong Tangent, Beta, Simple. This works well with Forge Neo. If you use Comfy, YMMV. It does seem to do well with other schedulers/samplers, but there, you may find yourself using 6-8 steps, with Euler, for example.
As always, feedback welcome.
Have fun using it! :)
The new LexiVision-Z has been trained on top of the base Z-Image-Turbo, an experiment with generation of synthetic data using SDXL LexiVision AIO VNSFW and training locally. Beta 2 has been trained for an additional 4000 steps. I'm still trying to tweak it stylistically to match a more analog aesthetic, but it's taking a lot longer than I thought it would. It can work with DPM++/Simple, Res Multistep/Simple or Beta, but I'd suggest playing around with your favourite combo.
SDXL Versions
This is a new merge of AIO-NSFW with the fantastic Photo1 by SubtleShader.
This is another merge of two private models that are being used to generate synthetic training data for other models. There are various 'flavours' of model, the first of which will be one focused (pardon the pun) on photo-realistic pictures of more mature models.
Prompts can be generic or more descriptive, and prompt adherence is ok. The model's fairly NSFW, so be warned. It can generate SFW or NSFW content equally well.
It's DMD2 based, and the VAE is baked in, so few steps are required for generation.
As always, my personal favourite generation parameters are in the showcase images.
Description
This is LexiVision-Z-RC-4Steps with a couple of new LoRA's baked in to refine skin textures and some anatomical issues.
FAQ
Comments (6)
I can't believe how good this is. It is the only checkpoint to be able to generate some of my most complex prompts where I meticulously crafted to manipulate and trick stable diffusion to generate. If it's this good, will a fp16 be better? I don't have enough knowledge to answer that question. But if it does improve over this even more I want it.
Thank you, I appreciate the feedback. I’ve deliberately used FP8 to keep the size down so it’s easier for people to run on less powerful machines. I am working on a new model with the same dataset, using the same base. I’ll be publishing that here too once it’s finished, and will include an FP16.
@TrueToLife_Fauxto Sounds awesome, waiting for fp16 :)
I tested v1.0-4Step and I have to say, it's incredibly good, everyone should try it :) The color is now perfect for me. It has incredibly good prompt adherence. I'm using er_sde/sgm_uniform with 8 steps again and I'm very happy with it. I also had almost no errors in my tests. I'm looking forward to your future projects; until then, I'm very happy with v1.0-4Step. Thanks for it
@TrueToLife_Fauxto It is also one of the very rare models that can create a matching metal collar :) very nice
Glad you like it, and thank you very much for the kind words, @Gorean