IffyMix V6
A true continuation from V 5.0 (not the HD versions or Re:mixed)
Better posing
About the same prompt adherence
Ever so slightly more detailed cartoonish look. But not by much
Lora compatibility high (as no LoRA's that interact with visuals are implemented)Notice your prompt used in the cover images? DM me and ill send you buzz
IffyMix Re:Mixed v2.0
A continuation of polishing the Re:Mixed version.
smoother, cleaner look.
wider style shifting.
artist tags work even better.
prompt adherence is low end of top tier.
natural language works when guided with tags.
Lora compatability is very high.IffyMix Re:Mixed v1.0
A complete remerge of the base models.
updated to to be merged with more modern models
Heavier toon look
Hi-Res Fix - No more orange tint
IffyMix V5.0
Its finally here. Iffymix with a base made of V2,3 and HD1
This has a nicer, smoother pastel like effect.
I feel that this mix gives surprisingly creative results.
Euler A, LCM, and UniPC samplers only!
IffyMix HD Remix V2 - EULER / EULER A ONLY
Various EULER A methods work! ( Automatic / SGM / Beta )
The "HD" like remix using IffyMix v2 + Iffmix v3 as the base.
!!!NEW RULE 34 V2 DATASET INCLUDED!!!
Contrast + Brightness precisely fixed.
Solid prompt adherence
This version continues with the vibrant cartoonish aesthetic.
I would say this version is slightly less 2.5D by just a smidge.
Background detail is improved mildlyIffyMix HD Remix V1 - EULER A IS PREFERED.
An "HD" like remake of iffymix (contains v1-v4)
Produces a clean, vivid, and a wonderfully polished cartoon aesthetic with a mild painterly effect.IffyMix V4 - STRICTLY NON DPM++ 2M KARRAS
Launching iffyMix...
Update found. Apply? [y/n]
V4 Released! (Online starting 6/26/25 @ 19:00)!!!! Currently Online !!!!V4 Update + Preview + Suggestions + unreleased goodies
https://civarchive.com/articles/15470/iffymix-v4-suggestions
Click to help shape and polish off Version 4!
V3 released
Some visual refinements made, added some lora's to increase anatomy, pose and dynamic angles. Increased R34 Lora strength. More predictable results, which sadly means some random creativity has been lost. Background detail has improved. Illustrious v2.0 has been merged in so higher quality gens all the way up to 1536x1536 can be done. Mileage may vary as this is experimental.
V2 Released!
This version adds some visual refinements as well as a huge update in tokens. I've merged in R34 tags/database, which was curated by iamddlta, creator of Pornmaster and it's variants.
This not only has wildly increased how good it gens explicit content, but the accuracy of anatomy and prompt adherence have also skyrocketed.
IffyMix - A Revamp of the best version of YiffyMix.........V51
Whats Different about this version besides its V51 + NoobAI?
Unbaked horrendously noisy lightning step before merging.
Completely prunes noisy, signature filled model, ChromaMix and its variants.
Merged with high quality models.
Restored its natural style's cartoonish look.
Parameters:
CFG - 4.5, seems to be the sweet spot.
Steps: 20 -30
Euler A
Tons of artist tags that sucked on NoobAI, do awesome on here. Totally unintentional, but its something I came across.
Quick Start Prompt
POS+
masterpiece,best quality,amazing quality, absurdres, anthroNEG-
bad quality,worst quality,worst detail,sketch,censor, text, ui, signature, patreon,Description
FAQ
Comments (27)
I decided to revisit some guy's fairly impressive SDXL checkpoint, and found a nice little nugget that he dropped alongside it:
https://github.com/DataCTE/SDXL-Training-Improvements
Would this be of any use, I wonder?
That's really interesting. But thats more for if I had 10,000 images that I wanted to train ON TOP of a model
Like noobAI is a checkpoint trained model. But Nova furry is a model merge like mine.
Model merges mix the unet weights of different models together.
There's no real training that goes into it.
V3 and V2 have both been pretty dang impressive. They have certainly become my favorites, especially when using higher than usual resolutions like 920x1360, often not needing any fixing for facial features/hands. Keep up the good work!
Thank you! I forgot to mention the newest version of Illustrious is mixed in here. So it can do up to 1536x1536 or something like that.
I gotta double check 😂
@LatteLeopard Yep that should be the correct max, I'd test how well it works in all its glory but amd has some issues and I can get a single image made before my system locks up from OOM
Very cool model, been using it the last few days and it works incredibly well to get a lot of non-2.5D slop styles / different furry types. Definitely recommend.
What do you mean by "2.5D slop styles"?
@Wallace_Tanaka Lol, when I read this I immedately knew what he was reffering to.
Ill give you a hint its a model that begins with the letter N
While v3 looks very nice, it seems that its somehow a downgrade in terms of creativity in comparison to v2. I once made an image in v2 and it looked visually spectacular, and when I used the exact same prompt in v3 multiple times, it gave me lackluster results. I dont know how or why, but it did.
Sorry that V3 is having that sort of result for you. V2 was a bit noisy and blotchy at times. In terms of style, the only change to V3 is noise control.
Do you have any recommendations for what you would like to see in V4?
I tested some of your posted images with V2 and you are absolutely correct in the difference.
I will restore some of the aspects of V2 when I release V4
@LatteLeopard I dont have any recommendations in particular. You've seen the clear difference in creativity for the two model versions, so you get where I'm coming from. I cant wait for v4 though!
@run_pumba
Heres a preview of V4 and im taking suggestions so I can deliver the best model to you guys.
https://civitai.com/articles/15470/iffymix-v4-suggestions
Does V3 have Rule 34 dataset, like V2?
Yes every version from V2 and up will have the dataset.
When you see a new merged model updated, the creator usually takes the prior version and adds merged or modifies the block wieghts.
Now the affect of the rule34 dataset can be increased or decreased.
If I increase it past a strength of 0.35 it can cause total chaos.
Is there anything spefic regarding the dataset you would like to see for V4?
This is a nice checkpoint! For V4, I want you to add datasets from DeviantArt, Twitter, Sketchers United, Furbooru, Derpibooru, Newgrounds, and YouTube.
I know there are a plenty of anti-AI artists on Sketchers United, but it's probably best to not train the checkpoint on anti-AI artists' works.
Also, I know YouTube is primarily a place to upload videos, though you can screenshot the videos (which would be time-consuming, so this is optional).
There is a newgrounds model actually. I belive guy90 made it. Also I didnt curate the r34 dataset, the creator of PornMaster model series did.
I could do DeviantARt, Twitter or the other Boorus, but its more redundant than Rule34 and those sites contain larger quantities of "low quality" artwork.
Derpibooru would be a fantastic choice for pony content, but unfortunately im 12 years past my short interest in MLP. Dont really wanna look at all that stuff for hours lmao!
Ill think about it though!
@LatteLeopard Sorry for late response, but:
1. You may want to merge it with NewGrounds Mix if you want, but guy90 said this doesn't work with "4GB LoRAs". Here's proof: https://civitai.com/models/1239699?dialog=hiddenModelComments&dialog=commentThread&modelId=1239699&commentId=705425 /adv /nm /nf
2. I know DeviantArt, Twitter, and other boorus may be redundant and contain larger quantities of bad artworks, but putting "worst quality" and related tags in the negative prompt should prevent low-quality outputs. I'm not trying to bash your opinions. /adv /nm /nf
3. There's a Hugging Face DeviantArt dataset, and deepghs made it. It's only meant to collect images whose like counts are no less than 2k, suggesting most of these images are high quality.
@urbanlegendwiki Excluding low quality images for the dataset is the best way to maintain consistency and quality.
@Wallace_Tanaka I'd agree with that.
I'm digging this checkpoint; it's become one of my favorites. But I was wondering, what image sizes do you all find to be the sweet spot for it? I'm pretty convinced this checkpoint can generate some awesome stuff at resolutions bigger than the usual go-to sizes(1024x1024, 1216x832, etc.), but I'm still trying to nail down the optimal resolution.
Any tips would be awesome!
I think any size based on 832x1216 or 1024x1024 will work best.
1024x1024 (square)
832x1216(portrait)
1216x832(landscape)
In a1111 try this example.
set the resolution to 416x608
In hires fix settings use x2 upscale.
This makes the best 832x1216 images.
You can even push it to 1664x2432.
Works with the other two ratios(square and landscape)
512x832, 832x512 and 640x640 work best for low-resolution compositions (oftentimes less "busy")
for busier higher resolution scenes, it's the same 832x1216/1024x1024/1216x832 for best coherency
otherwise if you wanna toy around with large images, you should probably use selection boxes and krita-ai-diffusion with this model plugged into it
Hi, I've stumbled upon some uncommon shinies.
Someone, somewhere, adapted IterComp somehow:
https://huggingface.co/ohgoshwhatunluck/IterComp-lora-64-comfyui
Used to be this was a monolithic SDXL checkpoint, but this LoRA works to share some of its composition-wrangling juices.
Unusually, it has both UNET and CLIP weights. (unet 0.7 / clip 1.05) seem to do the job, but (unet 0.85 / clip 1.25) also works. These are probably not optimal values, but I'm not particularly motivated to find an optimum.
It ends up helping on tricky compositions, as well as with handling 'natural language' (e.g. "Shantae jumping up and down a forest pathway at sunset") more consistently. Handy!
From the same mysterious place, I decided to give this (unet-only lora) a test:
https://huggingface.co/ohgoshwhatunluck/mapo-beta-lora-64-comfyui
It ended up tidying up compositions a little, generally positive.
What's a little annoying about it, is that it's "best" around 0.75 for simpler requests, but around 1.8 for difficult requests. It still helps at 0.75 though, so probably best to low-ball it.
Why do I post this here?
idk i just find this model the most useful i've found thus far
oh, right
If ever you felt tempted to release a CONTAMINATED (speedy) model variant, then I can recommend this setup:
https://huggingface.co/RED-AIGC/TDD/tree/main
1.40 of sdxl_tdd_lora_weights (yes, 1.4!)
https://civitai.com/models/510261?modelVersionId=567119
2.10 of this SPO thingymajig
--> model cranks out very nice images in 10 steps using LCM / Beta Scheduler / 1.5 cfg
But obviously, that is primarily for people who don't like doing big LoRA chains with these stinkies.
technically then it does even better with 2.0 cfg and skipping cfg on step 1 + 7-10
"Scheduled CFG Guidance (KJNodes), Start Percent 0.10, End Percent 0.6"
but that's rather beyond the scope of simple end-user "i want model print fast" use-cases, hmph
Enjoy!
♥♥♥
Im absolutely going to be DMing you at some point about these when I try using them.
V4 is coming out soon! Come help and make a suggestion!
https://civitai.com/articles/15470/iffymix-v4-suggestions