This is a blend of models that I've been working on for a while, with the primary goal of creating realistic character portraits in a variety of genres, but it's a well-rounded model that's been good at everything I've thrown at it. It's probably not very good at anime, but if that's what you're after there are plenty of other models to choose from.
This model inherits the CreativeML Open RAIL-M and Dreamlike License. By downloading, you agree to both.
No need to use a VAE. The model has vae-ft-mse-840000-ema-pruned.ckpt baked in. I also strongly recommend using the following negative embeddings: neg-sketch-2 by JPPhoto and neg-anime by dwringer.
The example images were done with InvokeAI. If you want to reproduce them with a1111, you'll need to convert the syntax.
Description
FAQ
Comments (23)
Not sure If I am in the minority or not but when I see nice samples that have metadata stripped out of them I just don't want to bother.
With so many models out there there is not much time to really experience everything so having a model that provides some sample generations with metadata is highly recommended.
I do LOVE the dog generation, I wish I could recreate it, but alas, without the prompt/settings it is impossible
I didn't strip metadata, civit just didn't pick it up, probably because I use InvokeAI. I'll see about manually copying it over in the morning.
@whosawhatsis thank you very much!
@whosawhatsis also, would be good to tag someone from Civitai and mention this problem so the data from InvokeAI is also properly imported :)
Oh hey yup! Those images totally do have metadata, and civitai doesn't recognize it. I posted a feature suggestion on the civitai Discord to suggest they add recognition for InvokeAI's format - https://discord.com/channels/1037799583784370196/1076836878935924806/1076836878935924806 (Discord inv if you need it: https://discord.gg/UwX5wKwm6c )
oh wow! thanks @mcmonkey when you download the images and check their content the metadata is all there, it is in different format so that is why it is not picked up by civitai
thnks for the tip! i will definitely give it a try!
@whosawhatsis it's true the final upscale from InvokeAI DOES strip all the data
@DarkAgent It's not stripped, it's still there if you download the image and toss it into PNG Info
I can confirm: opened the first image with notepad and the metadata is there.
<x:xmpmeta xmlns:x="adobe:ns:meta/" x:xmptk="Adobe XMP Core 5.6-c138 79.159824, 2016/09/14-01:09:01"><rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#">
<rdf:Description rdf:about="XMP template with common namespaces" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:xmp="http://ns.adobe.com/xap/1.0/"><xmp:gm rdf:parseType="Resource"><xmp:prompt>abstract colors, texture, film grain, skin pores samdoesart++, intricate RAW full-body dramatic portrait photo of a nerdy goth chick, petite susan coffey natalia dyer, neon hair, pale skin freckles, analog style eye contact selfie, standing on a (cyberpunk)1.2 city street at night in the (rain)1.1, backlit (neon lights)1.2, punk undercut pixie cut mohawk, straps belts leather (jacket).5 tight black jeans, film grain, posing modelshoot style</xmp:prompt><xmp:sampler>k_dpmpp_2</xmp:sampler><xmp:steps>30</xmp:steps><xmp:cfg_scale>6</xmp:cfg_scale><xmp:seed>1775344174</xmp:seed><xmp:negative_prompt>anime, cartoon, bad, ugly, deformed, bad anatomy, bad hands, hard light, text, logo, jpeg artifacts, comicbook, stippling, topless, oversized loose</xmp:negative_prompt><xmp:model_name>proto-SamDoesArt_v2.4
I manually added prompt info to all of the example images.
@whosawhatsis the prompts are all the same
@epinikion They're not... There are a lot of prompt fragments that I like to reuse, and a couple of images might have been generated from nearly identical prompts (the superhero ones, for example), but they're not all the same.
This model definately leans towards porn / lewd
Don't you get tired of complaining?
I uploaded a new version that is more imaginative and less horny.
Not complaining at all just stating a fact. I'm looking forward to checking out the new version
But I would like to know how you achieved some of these results
if you download the files and inspect them, the metadata is all there!
@danGorst would you be so kind to tell us how to inspect the checkpoint file to see it's metadata? Do we use an text editor ?
Really great model, but as others have already stated, LORAs doesn't match very well with vae baked models. Can you upload a non vae version ? It would be very much flexyble.
Hmm, I don't think there's any good way of reverting the VAE. I haven't worked with LoRAs, but I find it hard to believe that vae-ft-mse-840000-ema-pruned.ckpt would conflict with them, since it's just a better-trained version of the default VAE. I can certainly imagine some of the anime ones tuned to do color correction conflicting, though.
ok, I downloaded the baked vae, but if you manage to upload the non vae version, I'll be ready to download it too. Anyway, great model.
This model is a sleeper hit.
It lets me do things it would otherwise require multiple lora's.
Nice job.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.