BMegaMiXL

Note: Inpainting model was slapped together after googling "how to make your own inpainting model" and following some Reddit tutorials and editing a ComfyUI Workflow to suit my purposes.

I tried to carry over the inpainting properties while keeping the model otherwise unchanged. (m:1, model-A:inpainting, model-B:v2, model-C:v2) - I have not tested this well enough to know how well it works, but...

It didn't give me any error codes, it did some cool stuff, I figured I would upload it.

V2 notes:

I got angry at the SD3 release, so I decided to double down and focus my efforts on SDXL again. I like this merge a lot. So I'm throwing it out there for the rest of you to use as well.

I noticed that there's been a lot of new NSFW models for SDXL. So, over time, I've started to figure, "why the hell not?" and I started to throw them them in. I was astounded at how much a few of them managed to shift the style, even for SFW outputs, assuming that the NSFW models didn't break the SFW model first due to some obscure structural incompatibilities or their wildly different focus in terms of training data and keywords. This model is absolutely NSFW-capable, but my focus has still been scifi and... "Overall Coherence," as I've started to phrase it. (The results are not particularly cherry-picked, for example.) The "wow" factor with this model is the variety of styles compared to my previous merges/SDXL experiences. Compared to v1/etc, it fills the gap of "what I expected to see from SD3"

I have to admit that I rarely individually test the models that I am merging with, so I'm not entirely sure 'which model made it so much cooler.' I simply do not have time: I make my own merges to do my own things, so "the best one I've made" is always the one I go to first. When I download a new model to merge in, I look at the preview pictures, and then at the results on the model's page for the images that other people have created with them and uploaded... and I judge from that alone, in many cases. I've tried to make it a bit of a Turbo model in the past, but I sort of gave up on that. Coherence Is Key! Variety in style is second. I just want a model that does cool sh*t, man.

So I test and re-test my own merges, and if the result is bad or worse, I throw it out. If the result is cool, I keep it. If the result is 'cooler in some ways, but lacking Coherence, I might try to merge it with a previous version, 'kneading the model like dough' (and doing a little ritual to summon the Gods of Chaos to guide me through this chaoswarp with a sound mind) until I have something good.

That's what I did with this one. I recently merged Pony DPO with SnowPony, merged that with Yamer-NihilMania, merged that with my own cthulu-abomination of a merge... but that's just this week's version. and before that... a whole lot of others. The result was 'cool, but incoherent', so I merged it with a coherent merge I made about a month ago... And I very much enjoy the result.

Below is what I said about v1. The edits take v2 into account. I couldn't think of any other ways to improve upon my rant.

(Has been merged with a few NSFW models, [but they are extremely diluted by now. This model is mainly SFW, with scifi+quality as my goals and primary criteria.] edit: I have unironically found that merging with NSFW models seems to somehow improve the merged-model's overall creativity and image variety/composition, when done sparingly. It can also break a model.)

I've been merging models ad nauseam ever since SD1.5 first came out, and after all of this time spent lurking CivitAI, I can't help but feel like I should give something back - so, consider this to be my (perhaps sole? idk) contribution.

I was testing this one out tonight and, it ... felt ..... worthy.

This is the latest in a long line of merges where I have been trying to blend my own personal all-purpose model (on my own hard drive, this "v1" was originally saved as "v8.9.20r" [and "v2" was "1.8.3rrrrr(x)Plus1.7.4.0-vaeBaked" for a little while]). The main goal is sci-fi, but it's good at everything I've tested it with.

Long prompts, short prompts - give it a shot. It's not quite as photorealistic as I would like it to be (it leans a bit toward anime/artistic), but hey, sometimes you gotta make some sacrifices for overall quality/diversity; if it struggles with perfect realism, I'll settle for artistic. (One of its ancestor versions was literally focused on being 50/50, capable of doing artistic or realism to an equal degree, and it's doable! but difficult to sustain.)

My nagging superstition is that there's two main tricks to merging: adding a boatload of concepts, and then stabilizing them (merging, checking results to avoid excessive distortion, merging with previous versions that were more stable, rinse, repeat). I think of it almost like breeding plants for some particular desirable characteristic.

Long story short, I basically knead these models together like dough until the results please me. It isn't scientific, but it works.

It has become such an extensive "chaos-merge" that I've mainly lost track of its contents, so I think it's probably fair to say that this model contains "at least a little bit of everything," especially accounting for other people's merges being thrown into my mix. But I can say that (for v1) "SDXL Unstable Diffusers ☛ YamerMIX" and "NewRealityXL" are were two of the most recent additions, and they definitely shook things up a bit.

edit: For v2, I most recently added Pony v6 DPO, SnowPony, Yamer's NihilMania... but that's just this week's merge.

There are no LORAs merged into this model (or at least, I didn't put them there), as I simply haven't figured that part of the process out yet. Please send help.

I don't really have any specifically "suggested settings," since I haven't found anything that it does particularly poorly with, except maybe Heun. But, my typical settings are:

• Extras tab, 'resize seed from height/width'=maxed

• Some base resolution that equals a total of 1024x1024 (896x1152 etc)

• Sampler: I just found out about "DPM++ 3M SDE Karras" (as opposed to "2M") today, and the results are fantastic. I don't change the Sampler very often.

• Hi-Res Fix: upscale by 1.4 to 2.0

  • in settings, enable 'Save images Before Hi-Res Fix' because sometimes the smaller version is legit better, or img2img-worthy (same for all models) (i don't really do this much anymore)

• aDetailer (auto1111 webui extension) can help a lot (especially with faces), but isn't typically necessary

• I generally don't bother with LORAs.

  • SDXL_Black_Sa_May (and the corresponding white one) are pretty handy, though, for shifting the overall color palette to a darker/lighter tone

    • v2 uses Pony's VAE (baked in) and is really good at contrast already

    • i use 'lastpiece' or 'flatpiece' VAEs to brighten/darken images now instead, when necessary. pony's vae seems to be the middle-ground between those two, and is pretty much ideal on its own

• CFG = 1.5 - 6, optimally 2

• Upscaler: 4x-UltraSharp

• 30 steps + 30 hires steps

  • (as low as 15 works pretty well. i have done as low as 8 with hires enabled [total 16 steps], using "dpm++ 2m sde Turbo" scheduler, but i still usually use "3m sde karras".)

• hires denoise strength: 0.5

• i don't bother with "score_8 score_9" etc but booru trigger words are absolutely present in this merge. i haven't really tested it for anime styles yet