Not necessary at all but if you're feeling generous and want to help support my unhealthy amount of AI generating and future art endeavors:
https://www.buymeacoffee.com/OnlyJams
There is 12 files:
ckpt 7.17GB*
safetensors 7.17GB
ckpt ema only 4.27GB
safetensors ema only 4.27GB
ckpt pruned 1.99GB
safetensors pruned 1.99GB
realistic 7.5GB
realistic safetensors 7.5GB
realistic pruned 2GB
realistic pruned safetensors 2GB
realistic ema only 4.2GB
realistic ema only safetensors 4.2GB
*The main download is the 7.17GB ckpt file.
Select what you prefer, scroll down and choose the version and download you'd like using the drop down.
They are slightly different even with the same seed and prompt.
No Viruses or Pickles! Checked using this: https://github.com/diStyApps/Stable-Diffusion-Pickle-Scanner-GUI
About this bad ass beast of a checkpoint:
I merged a few checkpoints and got something buttery and amazing. Does great with things other then people too. It can do anything really. It doesn't need crazy prompts either. Keep it simple. No need for all the artist names and trending on whatever.
Example Prompts:
female, pale purple hair, frills, detailed skin, perfect face, fashion photography, photo realistic, 20 megapixel, canon eos r3, detailed skin, detailed, detailed face, (full body intricate, vibrant, photo realistic, realistic, dramatic, sharp focus, 8k)
(extremely detailed photo 8k), full body shot photo of the most beautiful artwork, beautiful woman soldier, green hair, cleavage, wearing intricate advanced futuristic blue power armor, propped up on one elbow, cinematic lighting, very detailed face and eyes, park in background, high quality photo
Negative prompt examples:
Asian, cartoon, 3d, (disfigured), (bad art), (deformed), (poorly drawn), (extra limbs), (close up), strange colors, blurry, boring, sketch, lackluster, big breast, large breast, huge breasts, face portrait, self-portrait, signature, letters, watermark
Asian, large boobs, muscular, out of frame , worst quality , text , blurred , monstrous , hideous , ugly , duplicate , cropped , mutilated , horrifying
What I merged:
**Original post maybe? for the model.ckpt (General Purpose Model): https://huggingface.co/22h/vintedois-diffusion-v0-1/tree/main
The third one is a mystery because I cannot remember where I got it but it was called model.ckpt and I uploaded it myself because it was lost somewhere I cannot find. I did some .4 but most merges were .5 weight. I have tried so many merges and this one just clicked great.
I merged SXD with the model so I had modelsdx then merged that with Handas
You can use this model commercially I pretty sure? However I am not liable if you do messed up stuff with it.
Enjoy
Description
THIS CONTAINS:
4.2GB Realistic ema only ckpt
4.2GB Realistic ema only safetensors
FAQ
Comments (30)
Which of the model versions is Model hash: 2d13223f? I found several of the samples had this. The AutoV1 Hash version on the page seems to only show the prime download (I think can't toggle it) but I assume the pruned versions or the alt versions all have different model hashes so maybe that is the reason or is it from a home brew version :). It can help when trying to recreate to realize which of the samples came from which of the variants of a version.
Sorry for the late reply. Most of the time I use the 7.7GB ckptHASDX with all the photos I create
Is there a huge difference between the pruned and full? I have to use mobile data for my internet and only get 100GB a month, just want to make sure there is a leap in quality before I download :)
I would say use the smaller pruned ones to save your data. It gets very similar results. Being completely honest I use the ckptHASDX.ckpt 7.17GB file because it's what I'm used to. I will be making a post on reddit to show the differences of the models with the same seed and prompt within the next few days.
I use pruned to keep vram consumption down, see some of mine and I have tested other prompts and get great results
Thanks for the fantastic replies, I would love to see some comparisons! Keep up the great work :)
I think ill get the full version because this model looks dope af! Would love to use it in its entirety!
Pruning a model deletes the near zero weights of the model which retains the models more confident outputs but does reduce versatility a bit depending on how the trimming is performed.
For general use, a trimmed model results in nearly the same images most the time but it is recommended to use non-trimmed for merging and for more unique and experimental prompts.
Excellent explanation! Many thanks!
I did find confusing that the ema only is larger than pruned, usually ema only is the model to get pruned right? Here Realistic HASDX is smaller (I like smaller models) that EMA only, both are sub 4gb so both can work but is there a difference? I so far have these 3:
hasdx_prunedEmaOnly (original once it was pruned as it was too large for me to use before, then - realistic was released, initially again large and but HasedSDX was pruned so I downloaded it (hasdx_hasedsdx) and a few days later realistic got 2 smaller models and I took the pruned version (hasdx_realisticHASDX), I have not checked Realistic ema only, usually ema only is not a training right so far what I have works but not clear what is the best to get original HASDX and Realistic HASDX
@creeduk It's because with "pruned", people tend to mean they converted from floating point 32 to 16. That also halves the size of the model from 4 to 2 GB. Personally I noticed stronger detail differences with fp16 model versions (e.g. when I trimmed Protogen to fp16). So I would suggest to pick the ema 4 GB version when resources are no issue for you.
@robot1me That is good to know, I have been able run the 4gb ones OK with --lowvram and with the smaller ones can get bigger images before scaling. When a few of the 2.1 custom models dropped at 4gb because they load additional models I had to add xformers, but I may as well grab 4gb when it is available. Eventually a new card is needed.
Thank you for the info!
@bestjammer - Quick confirmation, Model link HasedSDX (then pruned in drop list) and Pruned/Ema Only (again pruned from drop list, so double prune?), they have different model hashes when loaded but after some testing on grids they produce (so far) the exact same result.
I was a bit confused on these 2, I downloaded both at first thinking HasedSDX was different to HASDX :)
Reading up a little on EMA vs none EMA and then some call it Full EMA, just leaves more confusion nobody seems to have really described it well. I know this is linked in some way to training and a difference of weights and biases for inference vs weights and biases for inference + momentum. I don't have the capacity to train or I might understand what that means more LOL.
For now would we say that those are the same release just different variations of Ema vs pruning? I think that means 2 basic models HASDX and Realistic version. (I am now deciding which HASDX to delete or archive maybe pruned EMA and keep pruned HasedSDX)
You have it right. Just two versions and 6 types each. All 6 produce almost the exact same images with both models. It's just what's included for extra training and other things I'm not advanced enough to understand. I am glad you enjoy the models!
Detected: Trojan: Win32/Sirefef!cfg
Almost All that I downloaded
Use an offline version like “invoke Ai.” Or in a Virtual machine
Misunderstanding or not I am going on what I have seen
from windows Defender after scan
Detected: Trojan: Win32/Sirefef!cfg
Status: Remove failed
This treat or app might not be completely remediated.
Date: 2023/01/21 12:38
Details: This program is dangerous executes commands from an attacker. “Severe”
Go whine to mommy.
THIS IS NOT TRUE! STOP SPREADING FALSE INFORMATION!
Your works looks promising. I am tempted to waste again my time with this model but before i will start, you this model is worth trying for training in dreambooth locally with my faces on this?
Whats the best version should i go for interms of training faces, is it EMa ony? do you know better what set should we have?
I am not sure on what is best to train, I still have to do that as well as a bunch of other testing.
This model has a problem that causes it to ignore the first word or token in your prompt
easiest workaround is starting your prompt with a leading comma, see their comment for more information
I had no idea
Are you not on reddit anymore?
I got banned off reddit for some reason. Every time I create a new account it gets banned. Some mod has it out for me I guess I can never use reddit ever again.
@bestjammer wow dude. What happened?
@randallaware I'm not quite sure but it's very frustrating. I asked for a reason for the ban every time and the only response is I violated a rule in the terms and service. With no explanation. So it is confusing and I assume personal. I must of made a comment somewhere that really got a mod angry and now they have a personal vendetta out for me for every account I create now.
@bestjammer have you managed to create one in the end? if not, try registering with tor. sad to see that there are reddit admins this shitty
@JakoDel I stopped using reddit entirely lol but I will try that if I want to again.
Great work ! Just a bit difficult to make a choice. Can you also add for the HasedSDX the safetensor pruned version ?
Can you specify which version of UNVAIL did you merge?
If it's not the newest, could you update your merge to include
Unvail AI 3DKX v2 ?
Thanks!
Seems like the dropdown menu for alternate files isn't here.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.