Fifth Mix

Fifth Mix is a furry focused, anime capable merge model that comes in several different flavors for you to enjoy.

I want to extend a personal thankyou to LatteLeopard for the amazing BANANA SPLITZ model. Please check it out, I think you'll enjoy it.

Version names? Why not v1, v2, ect?

Fifth Mix is a merge compilation of models with traits I like, mixed to reach differing goals of styles and capability. Each mix isn't necessarily a progression of quality, but simply mixed for different preferential uses.

• Fifth Mix [Vieux Carré]: This mix's main focus was on enhancing color depth, saturation, and contrast using v-prediction. The style has been mostly stabilized and cleaned up over (betamix), able to create character focused images in a modern, clean 2D anime-like style.

• Fifth Mix [Vieux Carré](Betamix): 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.

• Fifth Mix [Banana daiquiri]: this mix is an excellent 2.5D model for close character focused creations, with a focus on enhancing color depth, richness, and expressive lighting.

• Fifth Mix [Absinthe]: This mix is a solid foundation for most furry needs, it was created specifically to enhance 2D/2.5D Anthro/Furry/Kemono related prompts and concepts while retaining a fairly neutral style. Absinthe for me has shown to be very capable at handling large, complex prompts, and able to load up several LoRAs at once without breaking easily.

• Fifth Mix [Paradise]: Is the little sister to Absinthe being virtually identical in mix composition, and having very similar outputs. The goal for Paradise was to have something with a strong foundation, but present a soft, kawii aesthetic. (Sic. Better tuned for Kemono.)

Your prompts seem different than most models?

I use currently use ReForge with tag-autocomplete, and prefer using e621 tags where possible. Every model has it's own unique ways being best prompted and it's why my prompting changes so much between models. I encourage you to experiment, be creative, use your imagination! try prompting in whatever way you're most comfortable with. I use the prompts and structure I've found to give me the results I'm looking for on each model, though that may not always be ideal for you.

Artist tags?

Yes, many popular artist tags on Danbooru and e621 should work across Illustrious models, though I can not adequately comment on if a specific artist tag works with any Fifth Mix mix.

-------------------------------------------------------------------------------------------------------------

Settings for [Vieux Carré]

Recommended minimum positive prompt:

I personally recommend putting this at or near the end of your prompts.

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

Extra Positive prompt(s):

digital media \(artwork\), digital drawing \(artwork\), 2025, 2024, 2023, 2022, 2021, 2020, hdr, 4k, 8k,

Recommended minimum negative prompt:

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

Extra Negative prompt(s):

1980, 1990, 2000, 2010, traditional media \(artwork\), painting \(artwork\),

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

recommended settings:

• CFG: 4-5

• Sampler & Schedule: Euler A Comfy - Align Your Steps 32

• Sampling Steps: 20 - 60

Supported resolutions + raw upscaled resolutions.

Note: Raw upscaling is unstable, and may cause significant errors.

[Vieux Carré] is more stable at 1.25x than (Betamix), and I personally only gen at 1.25x resolution now.

-------------------------------------------------------------------------------------------------------------

LoRA compatibility with [Vieux Carré](Betamix)?

Fifth Mix [Vieux Carré](Betamix) is my first [NoobAI] (v-pred) model, however I've been using it since April with all sorts of LoRAs thrown at it for compatibility testing without much of any issues.

While I can't promise every LoRA you love will work here, I can state that I've seen broad compatibility with [Illustrious], [NoobAI], (v-pred) or (eps). Generally it'll just require changing the weight of the LoRA to get the desired results. Don't be shy, give it a try. 💖

Settings for [Vieux Carré](Betamix)

Note: Depending on the length and complexity of your prompts, moving the recommended prompt to the end might enhance it. It's been situational for me, but I've noticed better results with all the detail enhancing and quality tags at the end.

Prompt Base:

Pos+

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

Neg-

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

Neg- (Extra; these aren't always needed, but can help if you're getting signatures/watermarks in some images)

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

Other tags I like: (Pos or Neg)

digital media \(artwork\), digital drawing \(artwork\)

traditional media \(artwork\), painting \(artwork\)

recommended settings:

• CFG: 3-6

• Sampler & Schedule: Eular A - Automatic

• Sampler & Schedule: Kohaku_LoNyu_Yog - KL Optimal

• Sampling Steps: 20 - 60

Supported resolutions + raw upscaled resolutions.

Note: Raw upscaling is unstable, and may cause significant errors.

Some known issues/limitations present in (Betamix):

Feral in gen: While fairly rare, there's very little I could do to correct this, it's a small issue baked into some of the models used for mixing. While it only tends to show up in limited situations, it can be irritating having an otherwise great gen ruined by it.

• Possible fix(s): Use the base resolution, or longer prompts that lower emphasis on the species, this seems to help, but not totally eliminate the issue.

• You may also wish to try some of the fallowing negative prompts if it's enough of an issue

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

• Barring that, sadly, inpainting might be the only real fix.

Symbols of emotions: A more common issue, but it's been easy for me to filter it out with prompts.

• Possible fix(s): Here's the Danbooru list of emotion symbols if you want to prompt for/remove them

  ^^^, +++, 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

• Notice lines - shows up the most, and also keeping this in the negative can do some heavy lifting to remove most unwanted floating emotion symbols

Backgrounds: if no background prompt is given, it will default to white or some kind of simple letterbox. Overall, backgrounds are hit or miss on quality.

• common interior settings, such as bedrooms, or other domestic interior settings tend to be higher quality, sharp, clean, and well detailed.

• less common backgrounds can be blurry or have strange artificing, best to seek out some background LoRA's or use other tools.

burry eyes/details: This is sadly more common than I'd like, and the best solutions I've got for it are, sadly just as common...

• Use ADetailer

• Use inpainting

-------------------------------------------------------------------------------------------------------------

General settings for [Banana daiquiri]

Minimum recommended prompt:

As a minimum, this is what I would recommend.

Prompt:

masterpiece, best quality, absurd res, hi res,

Negative:

worst quality, low quality, low res, compression artifacts, signature, watermark, artist name, character name, patreon username, patreon logo, url, blurry, ai generated,

recommended settings:

• CFG: 2-5

• Sampler: Eular A

• Schedule type: Any (I personally use polyexponential)

• Sampling Steps: 20 - 60

Resolution:

• Fifth Mix is capable of all resolutions supported by SDXL models.

• 832 x 1216, 1024 x 1024 tend to be more stable.

Additional tools: (or your UI's equivalent)

• Hires. fix

• ADetailer

• Regional Prompter

• tag-autocomplete

Enhanced settings for [Banana daiquiri]

For more advanced guidance using Banana daiquiri, I recommend reading the fallowing.

Prompt structure:

This is just a general guideline if you want to replicate my prompting structure, you can also move each block around to change how much strength they'll have on the final output.

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

But wait! Some of your images have the characters near the end after a BREAK?

short answer, that's Regional Prompter for A1111 and is how I make images with multiple characters that don't bleed over. I really don't ever need to use the BREAK command outside of that, so odds are, if you see BREAK in my prompt, it's because I used Regional Prompter

Light and contrast assist prompts:

Below is just a small example of some tags you might consider using to spice up the lighting in your prompts.

hdr, 4k, 8k, 

chromatic aberration,

multiple lighting sources,

volumetric lighting,

high contrast, low light, low lighting, dimly lit,

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

sharp focus,

soft focus, soft shading, soft lighting,

cool colors, 

warm colors,

warm lighting,

Enhancement prompts:

I would advise caution when using more than a few enhancement prompts, these prompts can cause instability and loss of quality when overused for small details, but are powerful when used in moderation.

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...]

Light negative:

You can use similar lighting prompts in the negative to (try to) remove certain lighting from a prompt, using the [color] prompts below as white can significantly help if you're trying to make a dark and/or specific color saturated image.

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

cool colors, 

warm colors,

-------------------------------------------------------------------------------------------------------------

General settings for [Absinthe] [Paradise]

Minimum recommended prompt:

Prompt:

masterpiece, best quality, rendered, illustration

Negative:

worst quality, low quality, low res, compression artifacts, signature, watermark, artist name, character name, patreon username, patreon logo, url, blurry, ai generated,

recommended settings:

• CFG: 2-5

• Sampler: Eular A

• Schedule type: Any (I personally use polyexponential)

• Sampling Steps: 20 - 60

Resolution:

• Fifth Mix is capable of all resolutions supported by SDXL models.

• 832 x 1216, 1024 x 1024 tend to be more stable.

Additional tools: (or your UI's equivalent)

• Hires. fix

• ADetailer

• Regional Prompter

• tag-autocomplete

-------------------------------------------------------------------------------------------------------------

Notes:

• 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(Betamix): 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(Betamix): In a similar story to Banana daiquiri My prompting has changed since uploading the last mix, namely that a GPU upgrade forced me into a new UI (ReForge UI), and change gave me access to v-pred models. Because of the move from eps to vpred with Vieux Carré(Betamix) and access to a newer and more comprehensive toolset my prompting has changed with the tools.

• Banana daiquiri: My prompting and general workflow has improved since first uploading Fifth Mix, so with the introduction of Banana daiquiri, I've taken the chance to recreate the display images using updated prompting with new seeds. All display images for Banana daiquiri are passed through HiRes with 4x Foolhardy Remacri at 2.5x/2.75x upscaling, and then hit with various ADetailer models as needed.

• All models prior to [Vieux Carré](Betamix) use only Epsilon-Prediction (EPS) and should be highly compatible with anything else listed for [Illustrious] and some [NoobAI] (eps) LoRA and tools.

• It is possible in the future, I might attempt to remix older models with very slight changes to the mix ratio and/or order, using updated base models from all the wonderful model makers available; in such a case, I will mark them with a version number. But the naming scheme will remain the thematic differentiator between divergent mixes.

• On the note of other versions; I want to head off an expected question with an answer. "Will you make a Fifth Mix for Pony v6?" and the answer to that is. No. When Pony v7 comes out, that will be more likely yes.