Furry Model Merging Resources
All models are available to use with the links provided below, if you are unsure about a model then you can try them before downloading. They're also now available for use here. I would recommend reading the usage guide below or look at the gallery and the prompts if you are new so you can get an idea for prompting and make the most out of your buzz.
Future Releases
They'll come as they go. These things take time.
BreakPoint
Sat on this one for seven months, here you go.
Realism - Enhanced.
Improvement to TheTerribleTimmy
Bravo Zulu (V-Pred Model)
Sat on this one for three months, here you go.
Realism/2D Model, finetune of 3wolfmondaiSDG_v10
Thanks to qek for their help!
I would strongly recommend zsnr off.
*zsnr off*
zsnr throws all the weights off and this model was tuned without zsnr active, you'll need the node above to turn it off.
Usage Guide
Here's a quick sampler and scheduler guide:
List of working Samplers for zsnr off:
All of Them, except sa_solver and sa_solver_pece.
List of working Samplers for zsnr on:
DDIM | seeds_2 | er_sde | gradient_estimation_cfg_pp (CFG: 2) | gradient_estimation | res_multistep_cfg_pp | res_multistep | deis | ipndm_v | dpmpp_2m_sde_heun_gpu (CFG: 2.5 - 3) | dpmpp_2m_sde_heun (CFG: 2.5 - 3) | dpmpp_2m_sde_gpu (CFG: 2.5 - 3) | dpmpp_2m_sde | dpmpp_2m | dpmpp_sde_gpu | dpmpp_sde | dpmpp_2s_ancestral_cfg_pp | dpmpp_2s_ancestral | dpm_adaptive | dpm_fast | lms | dpm_2 | exp_heun_2_x0_sde | exp_heun_2_x0 | heunpp2 (CFG: 4 - 6) | heun (CFG: 3 - 5) | euler ancestral | euler_cfg_pp (CFG: 2 - 2.5) | euler
Optimal Samplers for zsnr on: dpmpp_2m, dpmpp_2m_sde, euler, euler_ancestral
Schedulers - Don't use linear_quadratic and exponential, the rest are good.
CFG: 3 - 6
Steps: ≥25
I can't guarantee functionality, but they're posted on the below sites for use.
PixAI:
https://pixai.art/model/1966105706285623265
This one works for on-site generation.
TensorArt:
https://tensor.art/models/951908180273684193
This one works for on-site generation.
SeaArt.AI:
This one does not work for on-site generation.
TheTerribleTimmy! - (Illustrious Realism Model)
Realism Illustrious.
Usage Guide
Use the same settings as MyPreciousJimmy!, guide is in that section.
If you lack the hardware to use this model locally, you can use them on the sites linked below (and here):
PixAI:
https://pixai.art/model/1851155741065773330
TensorArt:
https://tensor.art/models/834232328764066788
SeaArt.AI:
https://www.seaart.ai/models/detail/e5abb38332068b8216b2c9a11bc0ca97
MyPreciousJimmy! - (Illustrious Realism Model)
Realism Illustrious.
Usage Guide (for Illustrious model):
Sampler: Euler A, or DPM++ 2M if you want much more detail. I run Euler A because the detail can sometimes be too much. I have a guide further below.
CFG: 2 - 7.5, 2 - 4 is a good range. I run 3, images get very cooked past 7.5, don't move past 4 when prompting for 'realism'. Adjust this to also increase detail.
Steps: 17 - 50, I go with 25. You may want more or less depending on what you're going for and your hardware.
Scheduler: Automatic, or Karras. I run Karras. Feel free to experiment with this because I haven't experimented with this thoroughly, but I have ran some tests so I can offer some guidance, it follows below.
Samplers & Schedulers
Samplers Tested:
DPM Samplers (DPM++ 2M, DPM++ SDE, DPM++ 2M SDE, DPM++ 2S a, DPM++ 3M SDE, DPM2 a)
Euler Samplers (Euler a, Euler)
DDIM Samplers (DDIM, DDIM CFG++)
Schedulers Tested:
Uniform
Karras
Exponential
Polyexponential
SGM Uniform
KL Optimal
Align Your Steps
Simple
Normal
DDIM
Beta
I tested side-by-side each of the mentioned Samplers and Schedulers at what I consider optimal settings for this model when it comes to general content (CFG 3, Steps 25) and I tested for noise and detail. I did not test for artifacting or generation errors. These are my results while prompting for realistic or photorealistic images (this isn't comprehensive, this is simply my observations after limited testing, I'd post the grid tests but each file is 250 MB).
Results follow in the image below:
What is Good?
Noise and quality fall within acceptable ranges.
What is Smooth Details?
Noise is nonexistent, details are generally much smoother/flatter.
What is Noise?
Image have visible noise at the set CFG (3). You should increase CFG if you wish to generate on these Schedulers. (DDIM is good for prompting niche concepts. I felt that I should point this out, as I've been using DDIM since I first began generating).
What is Noisy/Washed Out?
Images have visible noise and is washed out. Do not use for generation (generally).
What is Poor Quality?
Finer details (eyes, vasculature, fingers, etc.) tend to generate at poor quality. Inpainting, upscaling, or refining is recommended.
What is Extreme Noise?
Noise severely degrades the image quality at these settings. Do not use.
This is in no way a comprehensive test of these schedulers and samplers. This is simply a quick guide as to what should generally be used when prompting for realistic or photorealistic imagery at what I consider to be optimal settings for the model.
If you lack the hardware to use this model locally, you can use them on the sites linked below (as well as this site):
PixAI:
https://pixai.art/model/1843720708298681858
TensorArt:
https://tensor.art/models/826612528500065571
SeaArt.AI:
https://www.seaart.ai/models/detail/e5abb38332068b8216b2c9a11bc0ca97
CalypsoRealism - (Realism Model)
Decided to revisit DisgustinglyReal and make some adjusts using some improvements from other models I merged. Better lighting, better detail, better "realism". Wowzers. It's a "realism" model, meaning that it was designed for realistic content, but if you really, really want to try, you can force specific non-realistic styles. But again, this model was designed for "realism". There's two different file versions, CalypsoRealism and CalypsoRealism9M. What are the differences? CalypsoRealism is the base model and does "realism" slightly better, with more nonsensical details and detail density. CalypsoRealism9M is an adjustment to that.
If you're unsure of the differences or which one is the best, I posted a couple comparisons in the gallery. You can decide which version is best for you.
Though not necessary in the slightest, you can read Version Differences for more information.
DisgustinglyReal - (Realism Model)
A quick merge I made out of boredom between several custom models I merged, posted for fun.
It's a realism model, and at times disgustingly so. It has some Pony merged into it so it does respond to score tags, but it isn't primarily Pony as it is based off my other model, Echo Alpha, so E621 tags apply. If you used Echo Alpha and wanted more "realism", this is that.
This does realism, only realism.
Use however you please.
BlackDays - (Environmental Detailer)
What is this model?
Unpublished.
Because no one ever used it. It offered more dynamic lighting and more vibrant/dark environments. It could also do many other unique things compared to other Pony models, but no one cared because I didn't put 'Super Hi-Res Omega Furry Realism' as the name. It's still available on other sites if you care to see it.
Echo Alpha Conceptualizer
An actual SDXL model, cool. Too bad it's unpublished. Bye-bye!
Want to use these models but don't have the hardware to give it a spin?
That's fine! You can use these models at the sites below:
PixAi: DisgustinglyReal - https://pixai.art/model/1780659938859289461
BlackDays - https://pixai.art/model/1803604141625589045
Echo Alpha Conceptualizer - https://pixai.art/model/1803609373522101881
TensorArt:DisgustinglyReal - https://tensor.art/models/762017856750648818
BlackDays - https://tensor.art/models/785758026486397196
Echo Alpha Conceptualizer -
SeaArt:DisgustinglyReal - https://www.seaart.ai/models/detail/cqup0vte878c73fjnrs0
BlackDays - https://www.seaart.ai/models/detail/cs8fudle878c73fuke90
Echo Alpha Conceptualizer -
Usage Tips (for Pony/SDXL models):
You know what to do if you actually bothered to read this far.
Do with these models as you please, crediting unnecessary.
Feel free to post whatever images created in the gallery below.
I'm not responsible for generations produced with these models, generate responsibly.
Description
Realism VPred, finetune of 3wolfmondaiSDG_v10 found here.
Made for use on ComfyUI.
FAQ
Comments (16)
ComfyUI thinks Bravo is an Epsilon model, I have to use ModelSamplingDiscrete
@qek I have the same issue. Since I used SD.Next to merge it and not my usual way, at some point the metadata for auto-reading between EPS and VPred got wiped. I’ve never merged V-Pred before and there’s no documentation on how to fix that issue, so I had to settle on downloading a custom node that let me select between EPS and V-Pred manually, and suggest people use the workflows in the images. Frustrating, I know, but it was the best I could manage. If there is a way to fix that, I will endeavor to fix it in a timely way.
@Janloy fyi, ModelSamplingDiscrete isn't a custom node, also ComfyUI saves prediction info in a checkpoint
@qek I’m no expert in Comfy, I only use it because A1111 shits the bed with VPred and I didn’t want to use ReForge because it’s abandoned. You appear to be more knowledgeable than me on Comfy, how would you fix the issue, if you don’t mind me asking?
@Janloy I've returned the key (v-zsnr), no longer need to force it to detect it as VPred
https://huggingface.co/qpqpqpqpqpqp/Bravo_Zulu_Fixed/tree/main
don't have to download, you can check my workflow to resave it as vpred instead
@qek Holy... was it really that simple the entire time? Many thanks!
So i am very upset for missing this, it did not appear in the front page as new model for Jan 8th, i even went all the way back and it DOES NOT appear in the list of models in general. Terrible loss!
@tsiriioannis749 Yeah, it sucks. None of my models show up unless searched and they're all stuck in their original upload dates. And it doesn't push notifications or even say they've been updated except on my profile. Blows, but it is what it is. Regardless of reach, I'll still update them, because at the very least I use them.
I am following you know, i had/was using all of your models but when you went under i thought it was over. Lima 2.5 appears in front page and this is how i saw your back. Thank you Janloy. If you could reupload Quality Embeddings For Furries or even make an update for them it would be greatly appreciated, im pretty sure a lot of people miss them.
@tsiriioannis749 I can reupload them, however I should mention I never made them, Sopenit made them. I would like to update them since I really liked the aesthetic and quality it produced (I made a lora to replicate the style) when combined but I don't have access to the training data. They never seemed to do anything for models beyond SD 1.5, just very slight alterations. I originally unpublished it due to the number of messages and complaints I received about it.
@Janloy Yeah i was aware it was from someone else but then again they went completely missing after you.
People complain really? That sucks to hear, people are out of their minds these days.
A lora huh ? that sounds great to have if you ever feel like posting it or an update for them for modern architectures would be a blessing.
Your models and direction were great. You would say the bravo_zulu is superior to its prior iterations or to each of its own kind of thing?
And last thing the lima 1.5 i consider to be one of the best i have ever used, i understand you moved away from that direction/style to 2,2.5 but consider having a lima alt for the direction of 1.5 imo.
@tsiriioannis749 I made a post using one of the versions of the lora around 10 months ago, used one of the versions for training the Timmy model. I might post the Lora, I trained it off of Lima so it should be good. I'd have to find the correct version first though. I've tried running the old embeddings through a converter a long time ago, but never really did anything with them. They're just sitting around somewhere in my Dropbox.
I prefer BZ, its definitely an improvement over the base model I used to make it. As for sequentially on this model page, depends- one's V-Pred and one isn't. Its a realism model but I use it for 2D and esoteric image generation since V-Pred excels at that. The only issue with BZ is that I primarily used A1111 for image gen, and I tuned the model with znsr disabled, so I use Comfy for BZ and Forge for other models when I'm running model tests. I had other non V-Pred models in the works, I even uploaded one but never published it (BreakPoint)- and I have around 180GB of custom realism illustrious models that are just moth-balled until I decide to make another realism merge. I haven't used TTT in a long while, mainly because I'm always making models and tend to jump to the next thing after getting one thing done. I sat on BZ for three months before posting it, and for Lima V0.1.0, I spent a month testing it before I felt confident in releasing it.
I was actually going to post a more 3D/realism(ish) Lima model initially but decided against it. I used to actually give people choices when it came to model releases, but no one knew how to read the model description and... lots of bullshit. So I just keep it simple and only keep up a couple things, it's why I took down my Pony models, SD 1.5 models and Loras. They're up on other sites, just not here.
If you want, I can give you a link to my model folder so you can test the several dozen mixes and see if one tickles your fancy.
@Janloy Hey, i can definitely do that and come back with feedback.
I saved most of the models you had posted before and/or from other websites even if some could be considered outdated now. Their results were always good imo. I even have the blackdays version around due to how unique it was. Didn't want to have them disappear.
All in all thanks for your effort you definitely had great realism models before anyone else time wise, as for people and comments dunno man people are not able to read door signs these days, we must push for faster ai takeover, generate more models quick!
Absolutely criminal that this checkpoint isn't more popular. I also only noticed it now when I happened to look on your page for any updates to the Timmy model. Wish more people would see it.
BZ is good, your stuff is the only realism I ever end up using because it's got just the right style and detail to it. now that all my stuff is noobv based this slides into my workflow so easily its so nice. its great to see you back and alive.
How am I just now finding out you made new versions of these?? Smh, I'm a fake fan.
Just know I always love using your DigustinglyReal and now your newer versions. They're leagues more balanced in how I want to imagine a "realism" model compared to, whatever results I be getting from Nova Animal. Of course, this is down to personal preferences but I can't deny it pleases a lot more boxes from me than any other realism model ever will.