Important information about each version is in, wait for it... the version notes! (Suggest reading those so you know how bad each one is. Thank you.)
I strongly recommend...
ADetailer model: face_yolov8n.pt
Using a variation seed (You'll get a lot of near misses, can't imagine life without a variation seed)
And suggest...
VAE: vae-ft-mse-840000-ema-pruned.ckpt (Since it's the only VAE I've ever used...)
DPM++ 3M SDE & Automatic at 30-40 steps, 768 x 512 upscaled to whatever size at between 0.20 & 0.50 denoising strength, CFG 3.5-5 at these settings...
(There are probably better settings, but I don't experiment as much as I should)
Description
A few notes for Aeaea Cove I…
Renders fatness well, I find it produces more natural curves and rolls than some of my other checkpoints, for a heavier and less inflated appearance.
Generally good at anatomy in my experience, for SD 1.5.
Some bias towards an exaggerated pear shaped figure, but is capable of producing more natural shapes with prompting.
Responds strongly to face references.
The most stylised of my current checkpoints, depending on the prompt and settings the images tend to appear more like digital paintings than photographs.
Very error prone. Don’t expect great images with every click, or even every other click. Recommend using a variation seed so that, if you get a good composition, you can randomise it and keep clicking until you get a good variation of the image.
Faces can appear unnatural sometimes.
I give it… 3/5
FAQ
Comments (4)
Noice! You should really upload a SafeTensor file though - PT is risky, and very outdated now...
Some more demos would be great too!
Any info on how it's supposed to be prompted?
I’m a bit of an amateur when it comes to prompting… You’ll probably get a better idea from looking at the sample images, but I can give you a few tips that work for me…
I use a lot of short phrases and natural language, stuff like “an overweight married woman getting dressed”, that I know 1.5 can’t really interpret very well, but I never seemed to have as much success with just tags. Maybe you will though! I haven’t really experimented much, I’ve just settled with what I find works for me, but there are probably other ways to get good results.
Having what’s pretty much the same thing but worded in different ways seems to work well. Like “hefty woman” and “dumpy fat MILF”, between more specific phrases like “big heavy boobs” or “overflowing fat flanks”.
If you want a particular setting or background, I recommend putting it towards the end of the prompt. Also, preceding it with something like “in a…” rather than just having, say, “living room” greatly improves focus on the subject.
I feel like I get better images without stuff like “sharp focus, 4K, ultra high resolution” or whatever else at the end of the prompt. I use them sometimes, but I suspect because those words are, deep inside Stable Diffusion, generally associated with unrealistic beauty standards they can counteract what you’ll probably be trying to get with this model.
The Vista model in particular has really surprised me over the past couple of days with how it can occasionally produce interesting faces with little guidance, but it does respond well to references. I usually prompt them with “face like X” somewhere around the middle to end of the prompt.
Lastly, you can also mix references with a prompt such as “face like A and B” or “face like A, B” and it will, with varying degrees of success depending on the rest of the prompt, the seeds, and how strongly it responds to either reference, come up with one person with features of both. You may also just get two subjects, but that doesn’t happen anywhere near as much as I expected.
@Vitiator So there's nothing like Bigger Girls / anything booru based in the mix (which would have distinct tags for certain concepts)?