Fifth Mix: a furry-anime fusion model series

♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥

Welcome to Fifth Mix, what mix would you like to start with tonight?

[Screwdriver]: NAI (v-pred)
[Vieux Carré]: NAI (v-pred)
[Vieux Carré](Beta): NAI (v-pred)
[Banana daiquiri]: Illustrious (EPS)
[Absinthe]: Illustrious (EPS)
[Paradise]: Illustrious (EPS)

Now that you've ordered, please read the instructions below.

When using the recommended tools and settings, all versions of Fifth Mix should be usable within the same recommended values without alteration.

You are however welcome to approach Fifth Mix in any way you're most comfortable with. Below are my recommendations based on my personal experiences and usage.

There are other UI choices, but A1111 is seemingly abandoned, and I don't know enough about anything outside of ReForge/Forge. Comfy has a large and supportive community and I know someone will be able to help you if you chose that.

Stable Diffusion UI:

Forge
ReForge
Comfy.

Extensions:

ADetailer (ESSENTIAL) link

Sweet and simple, you know what it dose and I couldn't live without it now. Here's the three ADetailer models I use most, and consider necessities for your library.

Kohya HRFix (RECOMMENDED) built-in with ReForge/Forge

Kohya HRfix is generally required for 1.5x raw resolution and non-standard resolution generation stability.

Sometimes when using both K-HRFix and Forge Couple, they might break. In most cases, this is related to the K-HRFix option for Downscale Factor just keep that at the default (2) and it should be fixed.

Fifth Mix [Screwdriver] and [Vieux Carré] have been thoroughly tested with this tool and I can't begin to recommend it enough.

Forge Couple (QOL) link

(Or any alternative to it, are) One of the next most essential tools you should have available. Opening up many more options for generation, such as for couples, groups, or certain difficult image layouts that wouldn't be possible otherwise.

Tag autocomplete (QOL) link

An absolute necessity everyone should have. Don't forget to change the .csv in your UI options to use the e621 tags.

The first row of resolutions represent the base supported resolutions that all SDXL and all fine-tunes built out of it's lineage should be capable of.

But, the further away from square images, both in terms of aspect ratio and resolution you move, most models become increasingly unstable without the support of tools like Kohya HRFix.

However, many newer models are slowly adding more support for higher base resolutions making more of these resolutions attainable without additional tools. I still highly recommend having access to Kohya HRFix, because raw 1.5x without Kohya HRFix can be very temperamental, resulting in very strange, impossible, grotesque and duplicate anatomy.

UI Settings:

Sampling steps: 25-50
CFG: 3-6

Sampling method | Schedule type

RECOMMENDED: Kohaku_LoNyu_Yog | KL Optimal
FALLBACK: Euler A | Automatic

Hires. fix settings:

Upscale by: 2-2.5
Denoising strength: 0.15-0.4
Upscaler: 4x_foolhardy_Remacri (link)

(Use lower denoising with higher upscale percentages.)

universalized prompting:

Below is a non-exhaustive list of some fairly core prompt tags I've pooled into my prompting framework.

Some of these prompts have more or less impact on the final result, and many prompts will also lose some weight with larger prompts, but they all do something.

Everything has been organized as best as possible and made as easy to copy-paste as I could make it for general use. If you keep notes, I would suggest saving this section in a notepad or somewhere so you can easily add to or modify it for your own needs.

Base positive prompt:

very awa, masterpiece, best quality, absurd res, hi res, newest, new,

Lighting:

hdr, 4k, 8k, 
chromatic aberration,
multiple lighting sources, volumetric lighting,
cool lighting, cool colors,
warm colors, warm lighting,
soft shading, soft lighting, 
dark, high contrast, low light, low lighting, dimly lit, dim lighting

Colorized-Lighting:

These tags will assist with helping to limit or focus the lighting spectrum. Great for establishing setting and atmosphere in your images.

colored lighting, muted colors, restricted palette

And this absolute nightmare will help with lighting the scene in the color(s) of choice.

I don't recommend using more than 2 colors with these prompt blocks, and some colors will overpower others, for example: Pink > Purple

[color] light, [color] lighting, [color] glow, [color] rimlight, [color] shine, [color] refraction, [color] backlight,

red light, red lighting, red glow, red rimlight, red shine, red refraction, red backlight,

green light, green lighting, green glow, green rimlight, green shine, green refraction, green backlight,

blue light, blue lighting, blue glow, blue rimlight, blue shine, blue refraction, blue backlight,

pink light, pink lighting, pink glow, pink rimlight, pink shine, pink refraction, pink backlight,

purple light, purple lighting, purple glow, purple rimlight, purple shine, purple refraction, purple backlight,

yellow light, yellow lighting, yellow glow, yellow rimlight, yellow shine, yellow refraction, yellow backlight,

orange light, orange lighting, orange glow, orange rimlight, orange shine, orange refraction, orange backlight,

white light, white lighting, white glow, white rimlight, white shine, white refraction, white backlight,

Enhancements:

detailed, 
detailed lighting, detailed shading,
detailed hair, 
detailed eyes, 
detailed mouth,
detailed fur,
detailed scales,
detailed genitals, 
detailed pussy,
detailed penis,
detailed nipples,
detailed [ect...]

Base Negative prompt:

worst quality, low quality, low res, compression artifacts, old, oldest, early, ai generated,

extra negative prompts:

signature, watermark, artist name, character name, patreon username, patreon logo, url, blurry,

Corrective negatives:

Long/unusual:

long body, long torso, long limbs, long legs, long arms, long fingers, long toes, unusual anatomy, unusual pupils, unusual tail, unusual eyes,

merged/fusing/mutated:

merged, merging, fused, fusion, fused legs, fused fingers, fused arms, fused toes, fused eyes, fused mouth, fused shadow, mutated, mutant, mutation,

unwanted creature/feral:

micro feral, feral, holding feral, holding animal, inset, cutaway, mascot, chibi, smaller feral,

Unwanted characters in background:

group, crowd, crowded, crowd scene, background character, silhouette, shadow creature, ambiguous character, duplicate, cloned, 2girls, 3girls, 4girls, 5girls, 6+girls, multiple girls, 2boys, 3boys, 4boys, 5boys, 6+boys, multiple boys, everyone, absolutely everyone, lineup, queue, harem, male harem, mixed harem,

Medium prompts:

Reminder that Fifth Mix is weighted to a clean, modern, 2D-2.5D stylized look, and while these prompts will alter the base styling, use your best judgement about what, where, and how much you use these.

digital media \(artwork\), digital drawing \(artwork\), 
anime style, official art,
traditional media \(artwork\),
painting \(artwork\), oil painting \(artwork\), acrylic painting \(artwork\),
watercolor \(artwork\),
marker \(artwork\),
pen \(artwork\),
graphite \(artwork\),
colored pencil \(artwork\),
3d \(artwork\), source filmmaker \(artwork\), blender \(artwork\),
pixel \(artwork\),
photography \(artwork\),
mixed media,
microsoft paint \(artwork\), 
sketch, colored sketch, sketch page, guide lines, monochrome, greyscale,

Symbols of emotion:

^^^, +++, puff of air, Anger vein, Spoken anger vein, Blush, Spoken blush, Heart, Arrow through heart, Broken heart, Spoken heart, Notice lines, Shout lines, Smiley face, Squiggle, Spoken squiggle, Sweatdrop, Comedic sweatdrop, Flying sweatdrops, Spoken sweatdrop, Squeans

This prompt typically dose the most heavy lifting out of that block of prompts if you're trying to stop any floating emotion symbols from showing up.

Notice lines

Character related prompts:

Eyes:

Use one of these prompts if you want more fine control over a characters base eye shape/design.

jitome,
sanpaku,
tareme,
tsurime,

Parts of the eyes you can modify with colors:

Eyes (Iris)
Sclera
Pupils (inconsistent)

Below is roughly how I block out my prompts:

Use roughly this prompt structure if you want to best emulate my prompts.

Positive:

[Character (if applicable)], [physical appearance],
[Clothing (if applicable)], 
[Setting (location, scenery, objects, background)], 
[action/pose/position],
[Lighting], 
very awa, masterpiece, best quality, absurd res, hi res, newest, new,
[enhancements]

negative:

Reminder, you don't need to use this whole block of negatives. But, you can.

worst quality, low quality, low res, compression artifacts, old, oldest, early, ai generated, signature, watermark, artist name, character name, patreon username, patreon logo, url, blurry,
white light, white lighting, white glow, white rimlight, white shine, white refraction, white backlight,
micro feral, feral, holding feral, holding animal, inset, cutaway, mascot, chibi, smaller feral, group, crowd, crowded, crowd scene, background character, silhouette, shadow creature, ambiguous character, duplicate, cloned, 2girls, 3girls, 4girls, 5girls, 6+girls, multiple girls, 2boys, 3boys, 4boys, 5boys, 6+boys, multiple boys, everyone, absolutely everyone, lineup, queue, harem, male harem, mixed harem,
[targeted removal of other things in specific images]
long body, long torso, long limbs, long legs, long arms, long fingers, long toes, unusual anatomy, unusual pupils, unusual tail, unusual eyes,merged, merging, fused, fusion, fused legs, fused fingers, fused arms, fused toes, fused eyes, fused mouth, fused shadow, mutated, mutant, mutation,

tips & tricks:

Much of what you can learn here from from all my guidance for using Fifth Mix will work on many other models. Not all, perhaps not even most, but many models work similarly enough to Fifth Mix series models that these directions should server well there too.
Getting a particularly "human" feature in your output? This can usually be solved by weighting prompts related to antho/furry features in your positive prompt up, and/or try adding some human related features in the negative.
Still getting obnoxious background characters that you don't want? Tried the recommended tags and then some? This is where you can use tools like Forge Couple to separate the character you're trying to prompt from the background, (Using advanced or mask mode) this is basically the nuclear option to delete unwanted background characters appearing, because it also tends to be harder to build your prompt and has a drastic impact on the final output. But it works.
Generally bigger prompts are better with Fifth Mix, no need to overly express every little thing, but leaving less ambiguity for the the model to fill in on it's own tends to help.
Opposite to the last tip, sometimes less is more! if you're using smaller prompts, you might want to use less enhancement and medium prompts, they have a more dramatic impact on the final output with smaller prompts.
As much as I prefer and recommend the use of e621 tags, sometimes, you might need to fall-back onto some very common, or specific Danbooru tags to solve something.
I always forget to mention this, but if you're ever trying to generate an oc serval... there's sadly a very heavy prompt overlap with Serval-chan of kemono friends, when using the prompt serval, best advice is to first prompt Serval-chan in the negatives, and better describe your oc serval. Similar occurrences are likely to appear with other species with popular characters that unfortunately use their species name in their official name.

Q&A section:

Addressing the most important question: "May I rehost Fifth Mix somewhere else?"

Tentatively, I'm fairly permissive of re-hosting Fifth Mix. But it'll always be a case-by-case basis for how comfortable I am letting it happen or remain up.

In general, as long as Fifth Mix is kept Free to download, Free to use, provides a link/credit back to this page, and the uploader isn't making money off of it. I'm probably not going to be bothered about it.

Can I use [...] Lora with this model?

That's sadly a very complicated answer. In general, if it's a standard LORA that matches the model type (eps or Vpred), almost certainly yes across every mix.

Across the just v-pred mixes:

Lora, Lycoris = Should always work.

DoRA, LoKr, anything else = probably won't.

Can I merge Fifth Mix into another model?

YES! ABSOLUTELY! No need to ask, I encourage you to merge Fifth Mix.

NSFW?

Fifth Mix is intended to handle most NSFW content well, and I promise you, I thoroughly test this behind the scenes.

However due to CivitAI's absolutely asinine policies, and terrible enforcement, I chose to use mostly safe-to-lewd content to show how models will look and feel.

You're welcome to post anything you create though, I don't hide posts.

Model names why not just v1, v2, ect?

Fifth Mix is a merge compilation of models that possess traits I like, all with differing styles, qualities, and capability. While every mix is intended to be a furry-anime fusion with a clean 2D though 2.5D anime-like style, each version will have it's own slight twist. Each mix is not necessarily a progression of all the same qualities as the previous version. I don't think version numbers convey how I approach merging and sharing Fifth Mix.

Why do you merge so many models like that?

Before I got into merging my own models, I would typically have 50-75 models in my Stable Diffusion folder ready to use, and that's a bit much.

So I started trying to merge down a few that I liked, seeing what would come out. I started like almost anyone else, mixing 2 models and seeing what happened. Then I started mixing 3, 4, 5... and so on, in a chain of merges.

Eventually I started to notice that even a little bit of one model being mixed into a chain of merges still had some fairly strong stylistic impact on the merges further down the line.

At the "Fifth Mix" is where I started to see some of the best outputs, any more than that and the last models would start entirely eclipsing the lower weight merges near the beginning of the chain. That's how I ended up accidentally creating the merge process and name-sake of Fifth Mix.

(Fun fact, the original name for Fifth Mix was planned to be "ani-fur")

Artist tags?

I think most artist tags you could imagine will work as long as they're well represented on Danbooru and e621.

For perhaps obvious reasons, I haven't tried them all and probably can't tell you if a particular artist tag works, or if it's accurate.

Will you make Fifth Mix for...

No, as of right now, Fifth Mix will only be for Illustrious and NAI. I have no current plans to create any other major branches for the series into other model lineages.

If there's a new model branch in the future I like, I'll probably attempt to branch out into it if it's possible for me.

Pony v6? Hard NO. Not that I expect anyone to actually ask, since it seems Ponyv6 has really fallen off since Illustrious picked up.

Pony v7? Personally, there's a lot of things surrounding this model I don't like, and what it produces so far also isn't looking very appealing to me either. I won't outright say no, but it's not looking likely.

What's next for Fifth Mix?

Two things have been on my mind a lot for the last few weeks while working on [Screwdriver].

A proper re-mix for [Paradise] (to round off my work in NAI v-pred for a while)
And a new Fifth Mix merge for Illustrious. (returning to EPS, as well.)

As is usual with the cadence of my model releases, this is something that will likely be months away from happening, and I wouldn't call these concrete plans, this more just an indication of my want to do these things. Eventually.

Credits!

A massive thank you to all the model creators out there, without whom Fifth Mix in all it's flavors would not be possible.

Model blurbs:

A vestigial leftover from before the [Screwdriver] page overhaul. I'll keep them around, but they're not important, that's why they're way down here.

[Screwdriver]:

Fifth Mix [Screwdriver] is the next important step in my AI journey, this mix was primarily merged for it's heavy prompt stability, and style. Though this did come at the cost of the extremely deep, rich, high contrast and color depth that [Vieux Carré] could achieve, however I believe this was a worthy sacrifice.

[Vieux Carré]:

Fifth Mix [Vieux Carré] represents something I've striven to have since first getting into AI, deep, rich, dark color and lighting. That's why this mix's main focus was on enhancing color depth, saturation, and contrast using v-prediction. The styling has been mostly stabilized and cleaned up over [Vieux Carré](Beta), able to create character focused images in a modern, clean 2D anime-like style.

[Vieux Carré](Beta):

Fifth Mix [Vieux Carré](Beta) This mix is a product of my own internal testing with NoobAI v-pred models, it was never intended to be made public, but I've enjoyed using it enough that I felt I should share it. I would say it's an unstable, inconsistent model with some unique traits that might be worth trying to tame.

[Banana daiquiri]:

Fifth Mix [Banana daiquiri] stands as one of the most capable and viable Illustrious models I've mixed, this mix is an excellent 2.5D model for close character focused creations, with a focus on enhancing color depth, richness, and expressive lighting over [Absinthe].

[Absinthe]:

Fifth Mix [Absinthe] represents my foray into model merging, and stands as the foundation of the direction of the Fifth Mix series, being a furry-anime fusion model.

[Paradise]:

Fifth Mix [Paradise] is a sister model to [Absinthe], sharing a very similar merge linage, just tilting a little more to the kawii kemono styling than the rest of the series.

Hey, if you've been reading all of this, thank you for checking out Fifth Mix. I hope it'll live up to your expectations.

Notes

⦁ [Screwdriver]: All showcase images were created at 1.5x raw resolution using Kohya HRFix; upscaled 2.5x using 4x Foolhardy Remacri and ADetailer. (Half truth, all images were intended to use Kohya HRFix, a few do not, see if you can spot them!)

⦁ [Vieux Carré]: well, I guess you don't get a clean version of the cover image, every single time I've uploaded it, it gets flagged for review and this time Civit's dumb ass bot decided it breaks TOS... I'll let you make that make sense.

⦁ [Vieux Carré]: All showcase images were created at 1.25x raw resolutions and upscaled 2.5x using 4x Foolhardy Remacri and ADetailer. Showcase images with two characters also used Forge Couple to facilitate stability and image layout.

⦁ [Vieux Carré](Beta): All showcase images were created at 1.25x raw resolutions and upscaled 2.5x using 4x Foolhardy Remacri and ADetailer. Showcase images with two characters also used Forge Couple to facilitate stability and image layout.

⦁ [Banana daiquiri]: All display images are passed through HiRes with 4x Foolhardy Remacri at 2.5x/2.75x upscaling, and then hit with various ADetailer models as needed.

Model Type	Checkpoint
Base Model	Illustrious
Published	12/8/2024