This model has been renamed to WoopWoop-General, as it is now a general-purpose model that can handle a huge variety of styles and subjects.
Latest release is 2.0 BETA, please read the version description for information on how to use this model. This is a TEST MODEL, if you want something with predictable results please use version 1.1.
A model trained to make detailed, anatomically-accurate (hands, faces, genitals, etc) illustrations of humans and non-humans alike. This is the model aiming at photorealism. Trained to require few negative prompts (see examples). For the anime/hentai model, see here.
Suggested style prompt: ((photograph, photographic photo, photogenic, rule of thirds))
I highly suggest using vae-ft-mse-840000-ema-pruned.ckpt with this version.
Easy Launch on Vast.AI:
Follow this guide on how to launch Stable Diffusion on Vast.AI, except where it says to put "/start.sh" in as the on-start script, instead copy/paste in the contents of this file instead.
Description
Significantly improved generation of sci-fi/fantasy themed content, landscapes, etc.
Introduced training for a hundred different art style varieties including various styles in painting, watercolor, pencil drawings, and photography.
Dramatically improved the amount of data available for creating character concept art for human and non-human subjects.
Dramatically reduced the likelihood of "burning" images with high CFG or step count, but still retains excellent results with low CFG and step counts.
Dramatically improved results on non-standard initial resolutions and aspect ratios -- try 960x540, or even 1280x540 to see what's possible!
Dramatically reduced the likelihood of accidentally creating NSFW content. This may result in lower quality NSFW content for those who are using it for that, if you run into this please let me know so I can address it.
FAQ
Comments (36)
Looks great. Did you by chance merge this with other models? I'm seeing a little bit of SXD there if I'm not mistaken.
Or did you train it all from scratch?
This is a complex merge of 6 different models, with a good amount of extra training on top. No SXD though.
Thanks so much for the info. Would you mind letting us know which models you used to merge with? I do the same on my URPM model in case you're interested in my recipe as well.
I'd prefer to keep the recipe private for now, as I'm still working on it. I'll release the recipe once I'm finished.
Sounds good. The only concern then until then, is if you merged with anything that included dreamlike or seek.art. If so, you'd be bound by new license modifications, and in turn people that merged with you would be bound by it as well. This is why it's extremely important as well to let people know your recipes imo.
For now I'll say I explicitly avoided using any models which would cause license restrictions for either myself, or other models downstream from me. In fact the dreamscape mode I created specifically because the Dreamlike models on here have tight restrictions.
Amazing to hear. Keep up the great work. The photo one is strong competition with my URPM model. In some ways you do it better. Glad to have some healthy competition!
Works pretty good at cunnilingus! There's no model out right now that's trained on that, so happy to see this one actually understand the concept. Others seem to always just do blowjobs when prompting for cunnilingus. My only guess as to why this happens is due to merging the blowjob model in and it seems to take over any cunnilingus. So in the future if you do merge one of those in, keep that in mind.
I explicitly avoided any specific blowjob models for this reason. Serious problems. It also tends to make /every/ facial expression blowjob related unless you work entirely too hard...
If you like the model, please leave a review!
Really glad to hear it, and look forward to any new releases!
*of human women
fixed it for you :)
Actually, it does really well with men EXCEPT male genitals. Everything else is pretty good.. I'm working on some samples to help everyone go off of. I'm also working on training a model on male genitals, the ones that exist are... not great.
Hey, is there a way i can use vast.ai startup script to have auto1111 loaded with this model and spit out a link for the webui?
totally, ill reply here with some instructions tomorrow
thank you so much!!! @zoidbb
https://docs.vast.ai/guides/stable-diffusion Here's a guide on getting automatic1111 running on vast.ai. When it comes to the part about the on-start script, copy/paste the contents of this instead: https://gist.github.com/zoidbb/3750efecbfb58662f31263fb51e33119
thank you so much...will give this a shot!
amazing... it works!
Could be significantly better. What models did you merge and with what multipliers exactly? Also, how many images where used to train this specific model in total exactly?
Thanks in advance.
You don't have to share, obviously. But am happy to help if you do so! Thanks for contributing
Can you share what issues you're having? The recipe is private at the moment, but if you can provide some info on what you're running into I can look into what's going on.
Well, since you thumbs-downed my comment I guess you're disinterested in helping. That's unfortunate :-/
Is it possible to install VAE if using Automatic1111 with Colab?
Should be able to, if you save a VAE model as "model_name.vae.pt" then automatic1111 will load it up at the same time it loads the main model. For example if you have a model called "woopwoop.safetensors", save your VAE model as "woopwoop.vae.pt" then just reload the woopwoop.safetensors model.
@zoidbb First - Great work, I was not going to write but though maybe I could get guidance so here goes.
I initially downloaded 1.2, it was in my feed, had SFW previews and it looked most appropriate for some SFW testing. I saw it might be possible to use for testing male physiques on some of my character, test getting better chests etc. and I loved the colors.
So far I find 1.2 overbaking many images, high contrast or like they are over sharpened causing halos. I am using the same step values as per samples and tried a range.
I then tried some of the test/sample prompts, but after checking the data, most were made in 1.1 not 1.2 (so matching was much harder).
I then download 1.1 (I had previously avoided as seemed more NSFW) to test the prompts and make sure I was not having some vae or other problem. I found the faces more pleasing in 1.1 and the results VERY good, the images more realistic. 1.2 tends for me, so far to often deliver more cartoon, anime or cgi-ish results, even using the photo prompts which I don't even need in 1.1. The same prompt will get a better result in 1.1.
Another big problem I keep finding in my 1.2 is incorrect nipple placement. Both models like to push for cleavage on dressed images, I get that, based on the NSFW training, it is not a problem. However the result often have nipples on the edge of cleavage LOL. It is quite comical how often nipples are peeking out of the side of the material in completely the wrong places (bad cosmetic surgery maybe :) ) nipples point at each other. Freckles go OTT, so have to be dialed back in 1.2 (could be same overbake issue) look awesome in 1.1 with same value.
QUESTIONS: Is there an ideal cfg, or prompt that differs between the 2 models to lower the over sharp or contrast. How or why are the nipples wrong so often in 1.2?
Are some samplers problematic on this model?
This is interesting, I don't have this issue with 1.1 or 1.2. Can you share some specific prompts/settings so I can investigate?
What I do is start with CFG 4, 12-18 samples, using DPM++ SDE Karras. I move about from there to tune things in where I want them in terms of detail.
I will do I am running a big batch across models to get a baseline and then I will set off a few between 1.1 and 1.2 again and add here, I want it to be helpful and to see if I have missed something subtle.
I will add the images in a "review" rather than add to imgur so the metadata is available. I will set version as 1.2 but it has both for compare.
@zoidbb - Hopefully something stands out (see review) and you can point me to making great images in 1.2.
Is this only trained, or is it a merge as well? Quality seems nice, but the skin looks smoother with 1.2 compared to 1.0.
Both.
Excited to try 2.0, fails in a way I haven’t seen before when I attempt to import it after downloading. (Draw Things app on iPad)
Interesting, I'll test against Draw Things on my iPad and see if I can figure out the cause. I only use Automatic1111, so haven't ever run into this.
Try the latest update, it’s properly trimmed.
Love this model! Getting really realistic results! Is it possible to generate a PickleTensor (ckpt) version of v1.1 too? Would like to use it in fast-DreamBooth. Unpickling the SafeTensor fails when trying to use it, tried to convert the SafeTensor with diffusers locally, but that also fails :(
I've tried over 100 models on this site with various prompts and I'm actually shocked. This one blows not just the other hundred plus models out of the water, but it also is better than all of my other favorite models. Thanks for this.
Keep an eye out for updates, next major revision will be coming soon!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.