Note: The prompt metadata on Civitai is broken due to using the Dynamic Prompt extension for WebUI. It includes my positive in the negative prompt. Careful :)
Mix IV - What!?
This is the current pinnacle of my merging "career"; a noise-offset enabled, realism-focused merge.
We started with TheAlly's Mix - SD1.4 + Waifu 1.3 + NovelAI. Hundreds of test mixes, and three release mixes later, we've arrived at the Holy Grail of TheAlly's Mix Merges (until version V).
This mix contains 75% TheAlly's Mix III, with the other 25% being a huge, crazy, block-merge mashup of realistic-focused models such as the excellent;
Uber Realistic Porn Merge (URPM) v1.1
QGO-10b
Realistic Vision V2.0
I tried my best to fix hands using some of the common block-merge hand-repair techniques, editing the IN00, IN01, IN02 blocks, and they've turned out not bad, have seen worse, still not great.
What's it good at?
Everything! Wild results. Awesome portraits. Check out the example prompts and try them for yourself. I will say, if you want the best from it, you're going to have to work at your prompts.
Recreating my sample images?!
Some of my sample images are created with ControlNet, to get poses and compositions from my previous generations - you'll never be able to reproduce those. Others, it's possible, but some do use Patron exclusive embeddings. The sample images are meant to show what's possible with a model. Use your imagination!
Description
FAQ
Comments (14)
hello great job, how could one make a library like this?
Does this include a baked VAE?
No baked VAE - not a big fan of that concept! :)
@theally good to know thanks. It might be worth mentioning that a VAE needs to be used in the description.
@abracadoomer hmm, a lot of these images were created with no VAE at all - I went through a no-VAE phase, lol. Really depends on user preference.
This is my go-to model for portrait generation! It's absolutely incredible, really love this model! Have been using it since release - almost non-stop =)
Landscapes also work well, if you hit the right (and precise) prompt tokens.
Seems to work best (in my workflow and experience) @ ~ 70-100 iterations, can't say why. Very nice results with hiresfix engaged (while using "4x_NMKD-Siax_200k" as upscaler & the kl-f8-anime VAE - additionally)
Why is The Ally's Mix models all on different pages, for example version III? https://civitai.com/models/10752/the-allys-mix-iii-revolutions
No particular reason - just the way I posted them. They are new mixes, technically. I just use the I, II, III, IV naming convention. They're continuations of previous merges, true, but... different enough to warrant a new model card.
@theally Oh, I just though it was odd cause so many models just list different versions on the same page and explain how they vary, etc...thanks for answering.
Models share the same model card when they obsolete previous ones and you're supposed to use the most recent one, here, none of the previous models have become obsolete.
@theally will their be an update for 1.5 soon or are you fully moving to xl
Hi!
It seems like I found good recipe to fix the hands for my model.
https://civitai.com/models/116814?modelVersionId=142151
I used the closeup photos with hands for training, but it was not enough.
You also need to use SuperMerger with MBW checked, but not with IN00-IN001.
See explanation https://huggingface.co/WarriorMama777/OrangeMixs/discussions/66
According to the discussion above, hands and fingers should be somewhere in a middle.
IN03..IN09 and OUT07..OUT02 (from top to the bottom)
I see there are custom trigger words in this model ("Style-Micro, Style-NebMagic" are in a couple of the example images). Is there a list of these available?
They're not triggers, they are Textual Inversion embeddings which I don't think I've released 👀