This model was trained on 21 images of actress Ana de Armas. It was trained in sd1.5, so works best in photorealistic models, and also works best in close shots. For longer distances, upscaling and inpainting seem to work okay. One outift in particular in the training data can overpower the images, so if you notice some metallic or rainbow vertical banding, lower the hypernetwork strength. I get results that I like around .7 in Protogen, though most of my samples were at full strength in an older model of mine
Edit: Added a few Astronaut Anas made with Protogen x5.3 to fill out the 20 image limit.
Description
FAQ
Comments (17)
I like what I see here very much and will definitely test it out ASAP :)
Question from me - which app did you use to train the hypernetwork (and how much VRAM do you have?)
I've been doing dreambooth models so far but maybe it's time to switch to hypernetworks (will have to check now photorealistic looks like on this, but damn, this looks very nice!)
Thanks! It was trained in sd1.5 I have also trained a Dreambooth in a modelmix that was simliar to Protogen in its illustration/realism mix, but not as high quality. The DB gave me more accurate and reliable results, but I was stuck in the style of the model, so the hypernetwork is more flexible. There was a lot of cherry-picking involved in choosing samples, though. 3060 12GB, by the way.
Also, just looked at your profile, you have some great checkpoints. And sorry, you asked which app, not which model. My training is done on a local install of the Automatic1111 webUI.
i can't find anything by 'modelmix' or 'modelmix hypernetwork'
could you link to the repo or some site? i would like to try out those hypernetworks too :)
i have 2080 TI 11GB so perhaps it should be fine :)
thanks! i try my best :-)
ah ok, i guess the time has come to finally download the dreambooth extention - i've read that people had issues with it and it even crashed their webui installation and they had to reinstall it from scratch
i hope it will be smooth for me! :)
Sorry, I don't have the model up anywhere yet. Need to remove certain components and replace with newer, better releases here on CivitAI, I was just saying that my DB had been trained in a custom model, but that Protogen (x5.3 in particular) captured the look I was going for better than I've managed so far, so I may retrain one that one, if I can figure out the new Dreambooth extension.
Which model ckpt did you use for your examples? I see a hash a35f230f, but don't seem to have that one.
The model hash that starts with a35 was a custom mix on my machine. I haven't uploaded it yet, because I'm planning to replace the existing photrealistic component in it with something newer and more consistent. That was the images in the middle. The first four and the last 3 were made with Protogen x5.3, available here on CivitAI.
The recipe for the older mix, if you're interested, is below:
ATitanStrawBerry + RPGv2 at .5 = RinaA
bstaber + Anythingv3 at .2 = RinaB
eldnering + Float's Mix at .5 = RinaC
RinaA + RinaC at .5 = RinaAC (also neat by itself)
RinAC + RinaB = RinaFinal.
my hash is a35f230f
float's mix is available from https://rentry.org/animusmixed, along with many other cool models
(that bit was copypasta)
Thanks! I'll give it a try. I'm having some success with protogenX58 with this set around 0.3, but I wanted to see how yours would compare. :)
my Protogen samples were done at .7-.8 strength in x5.3. Have fun!
Any chance to have it as textual inversion? 🤗
I have not yet had much luck with good textual inversions. Of people, anyway. The robo face embedding worked out fairly well, though it doesn't work as well in protogen, I think because protogen specializes in very clear pretty skin, and the embedding relies on marks being drawn on the skin. It's definitely something I plan to work on in the future, but for now I get the best results with hypernetworks.
why hypernetwork instead of embedding?
Because I got more consistent and accurate results with the hypernetwork than I did with embeddings. I also tested a dreambooth model, which worked very well, but that locks you in to not only one person, but also the model it was refined from, and it's huge. At the moment, hypernetworks are the best compromise for me.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.