Licenses
Seek Art license, Dreamlike license
Introduction
So, this model has been in work for a very long time. This is not the same as my Aesthetic Mix, nor is it the same as my Wild Mix. It is a combination of the two with something extra thrown on top. It brought along a lot of the things that make the Aesthetic Mix and Wild Mix what they are, but it has some newer bits too. There are a lot of things this mix can do, and I feel like I've only scratched the surface so far. I look forward to seeing what you all create with it. I think version 1 still needs some work, but it is very good where it is at, and super easy to work with. So I figured I might as well release it in this state and let people use it while I continue searching for further improvements. As it stands right now, you get something that works well with a variety of prompts. It obeys prompts well and can produce a range of content. It is very consistent with the images it pumps out and has pretty decent anatomy.
Prompting
This mix is generally pretty easy to prompt. It works very well with any standard prompt you would use with my Aesthetic Mix or my Wild Mix. It can handle any prompt you'd typically throw at Protogen and probably a wide range of other prompts as well. It also supports some danbooru tags. Some of them get lost in the mix (it doesn't do any hardcore content well) but it is very capable with NSFW (softcore) or SFW content. It doesn't seem to lean heavily toward NSFW, but it does tend to lean heavily toward 3d render style generations. Put '3d' in your negative prompts to push it further towards realism if that's your intent. I recommend always putting 'masterpiece, best quality' in your prompt and 'worst quality, low quality' in your negative prompts. If you're looking for photorealism, I also recommend that you put 'photorealistic' in your prompt text.
Recipe
Below is a very long-winded explanation of what brought me to this release. TL;DR: This sh*t was a lot of f*cking work that almost never saw the light of day on multiple occasions. I pushed a lot of buttons, they pushed back, then I merged a lot of things, and came out with something I liked.
Initially this started off as an attempt to upgrade my Aesthetic Mix to use RPG v4 instead of v3. When I was done with that, I was not happy with the results. It didn't really bring much improvement to the Aesthetic mix. There was a little there, but nothing to write home about. So the recipe for that is as follows (this is important later on as it forms part of the base of the MagnumOpus mix but this recipe was never released.)
Aesthetic_v1.1: f222_v1 + elldrethSImagination_v10 @ 0.85
Aesthetic_v1.2: hassanBlend1512And_hassanBlend1512_1 + Aesthetic_v1.1 @ 0.85
Aesthetic_v2.7: Aesthetic_v1.2 + (rpg_v4 - v1.5-pruned-emaonly) @ 0.20
Aesthetic_v2.8: Aesthetic_v2.7 + (seek_art_mega_v1 - v1.5-pruned-emaonly) @ 0.25
Aesthetic_v2.9: Aesthetic_v2.8 + (modelshoot-1.0 - v1.5-pruned-emaonly) @ 0.15
Aesthetic_v3.0: Aesthetic_v2.9 + openjourney-v2-unpruned @ 0.05
Aesthetic_v3.1: Aesthetic_v3.0 + analog-diffusion-1.0 @ 0.05
Aesthetic_v3.2: Aesthetic_v3.1 + dreamlikePhotoRealV2 @ 0.10
Aesthetic_v3.3: Aesthetic_v3.2 + hasdx_hasedSdx @ 0.10
Aesthetic_v3.4: Aesthetic_v3.3 + (hassanBlend1512And_hassanBlend1512_1 - Aesthetic_v3.3) @ 0.15
Aesthetic_v3.5: Loaded into toolkit, pruned to fp16, and fixed CLIP.
After I rebuilt the Aesthetic Mix I was a little discouraged, but I figured why the hell not, let's merge it with the WildMix and see what happens. The result was actually something quite magical. It had all the aesthetics of the Aesthetic mix, and all the vivid colors of the wildmix. It was quite amazing and I liked it a lot, so I considered releasing it as an update to the Aesthetic mix, but then I discovered that all the NSFW content of the Aesthetic mix was neutered. It wasn't fully destroyed, just didn't work as well as it had before. It was super difficult to get an image of the subject with spread legs, the vaginas often came out with 1 or more extra lines, and nipple textures were mangled. I liked it, but... I didn't like it. For reference, the recipe for my Wild Mix is this:
ProtoDucHaitenMix_v5: Protogen 5.3 + (DucHaitenAIArt 2.0 - v1.5-pruned-emaonly5) @ 0.5
WildMix_v2.1: ProtoDucHaitenMix_v5 + (kenshi_10 - Anything-V3.0) @ 0.5
WildMix_v3: WildMix_v2.1 + SynthPunkSearch_raw @ 0.30
WildMix_v3.1: WildMix_v3 + vintedois-v1 @ 0.15
WildMix_v3.2: WildMix_v3.1 + (rpg_v4 - v1-5-pruned-emaonly) @ 0.20
WildMix_v3.3: WildMix_v3.2 + rfktrsDarkdream_v16 @ 0.1
WildMix_v3.4: WildMix_v3.3 + dreamlikePhotoRealV2 @ 0.1
WildMix_v3.5: Loaded into Toolkit pruned to FP16
WildMix_v3.5.1: Fixed CLIP positions
Baked in Standard 1.5 VAE
Then the subsequent mix, which I called Stably Diffused's Experimental Mix was this:
StablyDiffusedExperimentalMix_v1: WildMix_v3.5.1 + AestheticMix_v3.5 @ 0.50
So after that, I went on a journey. I was searching for a way to put better NSFW content into the new Experimental mix that I created. In the end I actually gave up... temporarily. Everything I tried either messed up the anatomy of the Experimental mix, or it didn't solve the issue at all.
So I moved on, thinking that I'll just work on something else for a little while and see where I get. I started working on an Animated mix that was going to be based off of grapefruit_v4 but pushing it more towards a semi-realistic 2.5d style without destroying the grapefruitiness of it. I got some results that were pretty good, but every time I made some success it still didn't keep the more hardcore aspects of grapefruit intact. It destroyed all of that. Again, on a whim I decided to merge one of these iterations with my Experimental mix just to see what happened, and again, I was pleasantly surprised. It still didn't retain the more hardcore aspects of grapefruit, but it did improve the NSFW abilities of the Experimental mix, and kept all the coolness that I saw in Experimental. The downside was that it also removed a lot of the realism that was in Experimental and it created something that was only about 2.75d. It was cool, but not cool enough. So then that started me off onto a new journey. I figured out something that would be able to fix the NSFW capability in Experimental mix, but I also wanted it to retain photorealism at its core. I had now learned how to use the Merge Block Weighted extension in Auto1111, and was playing around with it to find different ways to merge models. So that led to the following recipe to create a 2.5d Semi realistic Anime/NSFW base that had good anatomy, good eyes, and brought good aesthetics to the mix. You will probably notice that there are a lot of skipped numbers in the versions below. Those were the failures, we won't talk about those.
AnimeNSFWBase:
AnimatedNSFWBase_test_3: MergeBlockWeightedMerge::
Model A: Corneos 7th Heaven V2
Model B: Grapefruit_v4
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
AnimatedNSFWBase_test_4: MergeBlockWeightedMerge::
Model A: dantiondrawmix_20
Model B: AnimatedNSFWBase_test_3
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
AnimatedNSFWBase_test_5: MergeBlockWeightedMerge::
Model A: AnimatedNSFWBase_test_4
Model B: anidosmixV2_anidosmixV2
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
SR_AnimeNSFW:
SR_AnimeNSFW_v0.6: MergeBlockWeightedMerge::
Model A: perfectWorld_perfectWorldBakedVAE
Model B: AnimatedNSFWBase_test_5
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
SR_AnimeNSFW_v1.2: MergeBlockWeightedMerge::
Model A: chilloutmix_NiPrunedFp16
Model B: SR_AnimeNSFW_v0.6
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
SR_AnimeNSFW_v1.5: MergeBlockWeightedMerge::
Model A: duchaitendarkside_v30
Model B: SR_AnimeNSFW_v1.2
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
SR_AnimeNSFW_v1.6: SR_AnimeNSFW_v1.5 + modelshoot-1.0 @ 0.10
SR_AnimeNSFW_v1.7: SR_AnimeNSFW_v1.6 + dreamlikePhotoRealV2 @ 0.15
SR_AnimeNSFW_v1.8 Prune to fp16 (fix broken clip)
Which leads us... finally... to what created the MagnumOpus mix v1:
MagnumOpus_v1.0: MergeBlockWeightedMerge::
Model A: StablyDiffusedExperimentalMix_v1
Model B: SR_AnimeNSFW_v1.8
Base Alpha: 0
Preset Weights: SMOOTHSTEP*2
Credits
Here's a list, with links, to all the models that were used in this merge.
Description
Initial Release
It is a pruned fp16 model with the Standard SD VAE baked in.
ALL EXAMPLE IMAGES WERE GENERATED WITH XFORMERS ON! (All x/y/z grids were generated in Auto1111, the single images were generated in InvokeAI. The Grids were created to show the comparison between this model and my previous releases. They also include the Experimental Mix which was not released.) No Lora's or Embeddings were used to generate any of the example images.
Settings in Auto1111:
Clip skip: 1
Restore Faces: Checked (Using CodeFormer at default settings)
Hires Fix enabled
Hires Steps: 15
Upscale by: 1
Upscaler: Latent
Denoising Strength: 0.7
Default Width: 768
Default Height:: 1024
Sampling Method: DPM2 Karras
CFG Scale: 8
Steps: 24
XFORMERS ENABLED
Settings in InvokeAI:
Steps: 24
CFG Scale: 8
Default Width: 768
Default Height: 1024
Sampler: k_dpm_2
Face Restoration: enabled
Type: gfpgan
Strength: 0.8
Upscaling: disabled
Other Options:
Seamless Tiling: disabled
High Res Optimization: enabled
High Res Strength: 0.75
XFORMERS ENABLED