Banana Splitz XXL - V1.1
(Banana XXL + FurrySpec | Simple Furry Mix + ReKemonoV9 + Furry Dreams + a few LoRAs )
Minor change in prompt adherence
style update - slightly more painterly and lite 2.5D aesthetic.
Prompt adherence slightly LESS sensitive (allows for more advanced prompting)
Camera angles and poses are about the same.
Banana Splitz XXL - V1.0
(Banana V9 + FurrySpec | Simple Furry Mix + DeepNavyV3+ ReKemonoV7)
Minor update to coloring + lighting
Major style update - smoother, cleaner, less painterly more solid.
Prompt adherence slightly more sensitive.
Offers a high quality 2D - 2.5D look with expressive posing and faces.
Camera angles are more creative when propmting for dynamic angles.
Banana Splitz - V9.0
(Banana V8 + V6)+DeepNavyV3+Luno V6[Renamed ReKemono])
Major update to coloring, lighting, poses, camera angles.
LoRa compatibility is still great
Minor clip difference from 8.0 (adherence is roughly the same)
Banana Splitz - V8.0
(Banana V6+V7a+V3)+DeepNavyV3+Hyphoria+More)
Major clip update. Expanded clip with Hyphoria
which is a model that cleverly merged 10 different models clips
Rule 34 Data set added as usual.
Prompt adherence is crazy strong.
LoRA compatability is also very high.
Backgrounds are wildly beautiful.
Can do humans/Anime with no problem.
Extremely large spectrum of styles can be prompted.
You can combine artist tags easier like in the SD1.4 days
_______________________________________________________________________
Banana Splitz - V7.0 Side B
This is basically V7 but with a more westernized toon style
(Banana V6+V7+V5(ReMerge)+DeepNavyV3+Oops All Toons)
LoRA Used:
KSM - Gives the model a more bit of western toon appearance. Helps with artist tags
Dirtysoda - Another LoRA for helping westernize this model more
Rule 34 Dataset LoRA - Smallest ammount used for activation. This is for better LoRA compatibility and higher usage causes noise. To finetune the R34 Lora within this model, please use actual Rule34 Lora. (READ INSTURCTIONS)
https://civarchive.com/models/1191424/pornmaster-rule-34-prompts-illustrious-and-noob
__________________________________________________________________________________________________
Banana Splitz - V7.0 Side A
(Banana Mix+MolKun-Mashup+Luno Mix)
Banana Splitz XL 6.0 + Banana Splitz 5.0
Other models in recipe
Hyphoria
Nlxl:
Nlxl v0.9 is a checkpoint that produces more stable and detailed results through enhanced Model Merged checkpoints created through model merging, combining the natural image generation capabilities of Illustrious, the extensive dataset of Noob, and the unique strengths of different models depending on the version
_______________________________________________________________________
Merged LoRA info, so you know how much extra of the same lora to use, or if it lora's wont be compatable. Also lets you know whathey do.
KSM - Gives the model a more bit of western toon appearance. Helps with artist tags
Rule 34 Dataset (LoRA)
The developer speaks Chinese only and doesnt really reply correctly to anyone.
but this LoRA adds a large amount of tags that do not exist on danbooru or e621
This creates an effect that allow an awesome solid mix of tags/natural langauge.
It also makes NSFW content so much better. I think it also re-enforces the knowledge of already know tags for pose, expression and anatomy wise.
bruhhhh Its such a huge an potent lora. Its 3.4gb and is merged is sometimes merged in SUB 0.1 weights. like 0.08, 0.06. Its gonna be doing something powerful
__________________________________________________________________________________________________
Banana Splitz - V6.0
(Banana+Oat+Sudachi Mix)
-Banana Splitz XL 2.0 & 4.0 mix
-Sudachi XL
-Plant Milk - OAT
-Xavier-Shift (Trunk of Furry Junk)
-Mol_Kuen DeepNavy
_______________________________________________________________________
Banana Splitz - V6.0 BETA
(Banana-Hassaku+Walnut Mix)
***NOTE V3.0 AND UP IS FOR EULER A SAMPLER ONLY**
New base merge recipe:
-Banana Splitz XL 3.0 & 5.0 mix
-Sudachi XL
-Plant Milk - Walnut
-Hassaku 3.0
-Mol_Kuen DeepNavy
-Hyphoria
-Rule34 Prompt Lora updated to V2.0
__________________________________________________________________________________________________
Banana Splitz - V5.0 [ READ ME! ]
(Banana-Hemp+Walnut Mix)
***NOTE V3.0 AND UP IS FOR EULER A SAMPLER ONLY***
Try using a CFG Rescale plugin for ComfyUI or A1111 if want to attempt another sampler
New base merge recipe:
-Banana Splitz XL 2.0 & 4.0 mix
-Sudachi XL
-Plant Milk - Hemp ll
-Mol_Kuen DeepNavy
-Rule34 Prompt Lora updated to V2.0
TEXT ENCODER MODIFICATION:
This version will properly acknowledge prompting from those who write longer and wildly detailed prompts. BUT! it will also help/support simple prompts into creative works of art!
Great for beginners and fantastic for pros to push their creative limits!
Positives for the Pros!
highly specific prompting will now have a much higher chance to affect small details.
I suggest even putting weather or time of day (morning, golden hour, night).
Even objects in background like a swing on a tree or a bookshelf indoors.
Style/Artist blending can have surgical precision.
detailed negative prompting can help shape result moderately better too
Positives for everyone!
Simple prompts will have a boost too:
Short or vague prompting will give creative wildcard like behavior.
Can give good solid and creative prompts with simplistic prompting.
Good for beginners to help build upon their prompts making them more verbose and gaining the skill to write more detailed prompts.
Great for beginners to see the changes by adding, changing how they prompt.
Remember! Everyone prompts differently and this model can certainly help you tone up on your own prompting skills!
_______________________________________________________________________
Banana Splitz - V4.0
(Banana-Walnut Version)
New base merge recipe:
-Banana Splitz XL 2.0
-Plant Milk - Walnut
-Sudachi XL
-Rule 34 Dataset V1.0
-Style is mostly the same, but higher quality.
-More detailed backgrounds
-Prompt adherence is average. Needs work.
-Extremely LoRA compatible
-Artist styles absolutely SHINE on here.
-Can one-shot generate a good image 3/4th of the time
_______________________________________________________________________
Banana Splitz - V3.0
(Banana Flax Daquiri Mix)
A 'high-quality' Western-Kemono hybrid with crisp linework, soft but dimensional shading, and a balanced 2.5D aesthetic.
Better and easier prompt adherence
Clean, stylish, thick lines with glossy 2.5D textures
Contains Rule 34 Dataset.
Great eyes! Great Fingers, Awesome poses.
Check out this amazing finetune of Banana Splitz by LS582.
Fifth Mix - Banana Daquiri does exactly what its told, clean, crisp style. This is what ChromaXL should have been IMO.
If you like this model try it out!
_______________________________________________________________________
Banana Splitz - v2.0
(Banana Daquiri Mix)Updates:
-Moderate Style Change
Model now is has moderately higher saturation, slightly lower contrast, more colorful.
Less flat and washed out (if you like that look from V1 I may consider making a V1.X, Plz Comment)
When no style is specified you will get random styles beetween 2.5D cartoon and 2.5D semi-photorealstic
-Massively Improved prompt adherence.
Has improved natural language understanding(needs to be mixed still with tags)
Prompt this model like a combo of Fifth Mix - Banana Daiquiri and NoobAI
-Refined quality and body comp.
Much less random zoom ins and cropping when using "close up" tags
You can get good eyes on raw gens 60% of the time with full body/multi character
You can get good eyes on raw gens 95% of the time with portrait/close up/solo
Different male and female body types
better knowledge of animal dicks and pussies(specifically canine)
-Pure NoobAI cyberfix, train difference injection
Use e621 tags for artist associated with NoobAI
vast knowledge of furry artists and can nail the style very closely 80% of the time. (check last 3 images in example images)
Natural Langauge is better understood. but not perfected.
-Making Lora/Lyrcoris models use the following.
NoobAI EPS1.1
Illustrious V0.1
WAI-NSFW-Illustrious V11 (Smooth looking with mild risk of sucking)
Fifth Mix Or Banana Splitz (Highly creative / Lower risk of sucking, but makes it useable for these models only)
__________________________________________________________________________________________________
Banana Splitz - v1.0
(The Original Banana Splitz Mix)
ORIGINAL VERSION 1 MERGE DETAILS:
Models used:
Fifth Mix
FurryToonMix
MolKun Mix
NoobAI CyberFix EPS V1.1
Custom Oops! All Furries
(Re-Merge/Pruned All trace of models that use ChromaXL Or ChromaMorgy
Custom Indigo NoobAI Mix(I might release this)
Custom Nova Furry Mix
(Custom Add Diff/Train Diff Mix of 3 versions - Might release this too)
Hassaku V2.1
Interested In any of the custom models used?
Let me know!
Description
I know its been a while since I updated, so take a sneak peak at its current state
Full version released soon
[Model not finished]
Increased visual clarity
Increased fine details in characters
Increased fine details in background
Slight increase in prompt adherence.
Moderate increase in creativity.
FAQ
Comments (25)
Wow! What happened to the backgrounds? They shifted to some weird, diffuse, blurry realism. The characters are better, but they just don't fit the background anymore. I went from V4 to V6 to try it, and while I find the characters more defined, the backgrounds ruin the image every time.
Thank you for letting me know, a new version of v6 will be released during the week before the auction ends
Wow, That was fast
what is the visual difference from 5.0 to 6.0? im very new to generating but i was just using 5.0 a bunch recently and like 2 days after i started you have updated it to 6.0 lol. I have to install A1111 to use it, so i dont know what the visual difference would be yet.
btw is comfyUI worth installing? sorry if thats too off topic
Never used A1111 before, but can say that ComfyUI its amazing for generating, its my actual setup and its pretty powerful and versatile, if you dont know how to make a workflow to just open and start generating, you can use mine by just draw and drop one of my last images (like the renamon examples of the yourdigimongirl lora i've uploaded) on ComfyUI
If you have any question with ComfyUI, just send me a DM
Enjoy! ❤️
ahahah oh no, this is v6 beta. This is for people to test before I release the fnished product.
I decided to let users who can use the models locally test it and help me fix errors before the final release.
@anomedra i have an AMD gpu and i installed reForge with ZLUDA to skip the CUDA errors but it seems to take 4+ minutes to load any checkpoints and LORAs and when i try to generate something it just sticks to 0% and never starts.
can i still install another UI like comfy or regular Forge without it causing issues with this reForge installation?
wait it just started generating images suddenly after like 10 minutes of it being at 0%... is that normal ?
when i change prompt or settings it takes several minutes to generate again, except sometimes it doesnt and it generates within 5 seconds, i dont think its supposed to be like this... but i guess it is working.
does comfyui or forgeui have a really long pause like that or is that just an amd thing in general (im using rx 7900 xt so i have decent amount of VRAM)
edit: it seems like it mostly lags a ton when i change a LORA or checkpoint but once it initializes (spending many minutes to generate an image or two) it works fine ? i guess reForge isnt broken it just has an insane load time for LORA and checkpoints?
@shibemorten It's because you have an AMD GPU. The model (checkpoint or LORA) is designed to be run on devices that can use CUDA code/cores. ZLUDA is converting all the CUDA stuff in the model over to something AMD GPU's can run. It'll happen every time you try to run/use a new model, but once it's done it's done, you shouldn't see any extra delays or slowness if you try to use that model again.
I recommend generating a sample image with any new models you download when you're about to go do something else. Gotta do the dishes/run to the store/mow the lawn/sleep? Convert some models while you're away. Then they'll be ready when you're back at your PC.
@HavingFunWithAI okay i see, sorry i didnt know about that beforehand lol, i guess thats not too bad i was concerned at first that it was going to generate every image for that long but it does seem to consistently only do that really long startup with a load a new model and lora. thanks for the info
@shibemorten
Dont worry I dont get locally either, Becuase I have other applications running that use GPU.
Check out vast.ai you can rent servers with A1111, Forged, Comfyui with a RTX4090 for about $0.03 - $0.45 cents an hour.
That means you can use about 10 - 12 hours of generating images for $5.00
https://cloud.vast.ai/?ref_id=115535
Seriously, check it out. I think they really offer a great service. I've been using them for years.
Ive used Runpod and some other service I cant think of the name. But Vast is the best priced and works the best in my opinion.
Its best for people who are fimilliar with using Jupyter notebook.
Maybe I'll write a guide on how to use it, and navigate the site.
@LatteLeopard I'm not currently interested in spending any money to generate stuff (aside from the initial purchase of my GPU and my power bill I suppose)
That does sound like a good rate/deal though, if I ever feel the need to get some super fast very high quality prompt and seed generations out of something I got locally. Thanks for the info
Gotta go fast! 😆
This is the beta version, I'll be reuploading the completed model during the week.
v6 beta is solid and inline with your version notes. I am really liking the improved clarity and vibrance; though I could see how it may take some getting used to by anyone who prefers the more naturally blurry results produced by v4.
More fun to use with human anime characters than I would have anticipated; can produce a sorta colorized manga effect with linework that pops.
Constructively speaking, you might be reaching a plateau on sharpness - while rare there are instances where the noisier elements of an image can appear distorted. These elements are replicable in v4 and v5, they just are easier for the eye to gloss over.
Great work =D.
Thanks bro! Youre absolutely right about the sharpness and the noise. Also the backgrounds are too realistic for such cartoonish characters. The final version has finetuned it all together in a much better way.
Instead of using Plant Milk's Walnut, Hemp, or Flax. I realized Oat was the best choice for contrast issues and that Sudachi and Xavier on certian UNET blocks added much better color and thick lines that dont appear so dark.
Usually my beta models are like this, I thought it would fun to share how it progresses from version to version.
Its a small gift to my local genners and fans ♥
does this add rule34 tags on top of the danbooru tags that illustrious uses, or does it overwrite danbooru tags with rule34 tags? sorry if thats a dumb question im still new to this stuff lol.
also secondary question: what CFG works "the best" for 6.0, or is there a range between two values that works best? ive noticed with CFG 2 sometimes it mysteriously gets more accurate? but its a rare oddity in my experience so far
To answer this, I should clarify there is no true relationship between where the tags came from, and the model you're using, they are just tokens (e.g. strings/partial strings) the model will recognize.
That said, it sorta does both of what you're asking.
Common tags between boards will sorta blend; so a concept that is on Danbooru but not popular, yet popular on R34, should now respond to such a tag better (and likely more geared toward R34 images used for training).
Additionally, tags that aren't on Danbooru at all, yet are on R34 should be more responsive to the model than a comparable model without that training.
As for your CFG question, can be highly dependent on model, and there is really no best. Use what you prefer. Too low will produce incoherent mush, and too high will produce artifacts or only noise (which will look like a broken GPU outputting something).
For Banana Splitz v6 Beta in particular, I'm having best results between 4-7, with 4 being my go to setting.
@r12x5s 4-7 has been working great for me, but going down to 2 can cause many interesting changes and sometimes it gets more accurate towards either the artists coloring style or the character ive prompted, its strange. for instance, i was prompting "mugendramon" and the difference in accuracy for the characters design increased as i lowered the CFG, im not sure why, but it is a rare occurance, so i test some of the batches on 2 just to see what happens now (especially if im struggling to wrangle it to do what i want it to do)
@shibemorten I only played with 2 a little bit earlier today, but I liked what I saw. Some models seriously suck at anything lower than 3.
My model, XAVIER (which I won't link here out of respect for @LatteLeopard) can make some really fantastic images at 1.5; yet 99% of people would never bother to even try that setting.
In my opinion the best way to improve at generating overall is to do unorthodox things to see what happens, and refine your prompting around what you've learned.
@r12x5s unorthodox definitely describes what im doing lol, currently trying to test every artist and digimon that im interesting in prompting to see if i need a LoRA for them or not, so far every artist ive tried works in some capacity, and most of the digimon ive tried have worked so far. also while testing those ive been trying to think of very dynamic and complicated positions to make it render and figuring out how the prompt tags fight with eachother and work with eachother by doing so and whittling it to a usable state for the complex position/pose im looking for. so far with great success except there is a major complication whenever i try to make a character autofellate while doing something else, but i have gotten it to do so, its just very random currently.
i should also specify im only using banana splitz xl for this testing with no loras and nothing merged, for control data's sake
@r12x5s This is exactly!! the answer I would have gave! Thanks man!
@LatteLeopard Any time my friend!