- Jan 24, 2020
- 1,401
- 3,793
Thanks guys. Though.. A blonde bombshell with huge boobas .. And everyone is like.. Hey look at the water.. And the lovely tile work, must have been expensive.. Those guys building it, how talented.. ![Laugh :LOL: :LOL:](data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7)
Faacinating, and thank you for taking the time to do this. I always train on SD1.5, but only because people say it makes for a more flexible and accurate model - nice to see someone test it.Been doing some testing to how training on different models affect the end result. Thanks to an update that added 3 hours to my already long training time it took a lot longer than hoped...
(each image is marked with the model used to generate the image)
Train on deliberate v2:
View attachment 2674951
Trained on dreamshaper v6:
View attachment 2674952
Trained on sd 1.5:
View attachment 2674953
All training is done with the same settings, same dataset, captions etc
The one trained on and generated with sd1.5 is definitely the one closest to the dataset and overall it seems to give the "best" result.
Doesn't come as a completely surprise but i'm rather surprised at the amount of aging/wrinkling (and size/weightloss) in the others. I guess it can be used if people are trying to create that effect considering how bad the AI is at generating "middle aged" at times, it's either 20 or 80.
Not sure how helpful (or informative) this is to anyone else, but as i'd done it i thought i'd share just in case.
I've found the most versatile photorealistic checkpoint to be CyberRealisticv3.0Hi, just getting my feet wet with stable diffusion after trying it out once months ago, does anyone know what are the most photorealistic checkpoints now? gonna bother Mr-Fox as well
PS nice blondie, but not as nice as the water
PPS send help for my poor 8gb 1070 Ti, she's screaming in pain
Also, my $0.02 is that the sampler matters a lot. I.e. Heun does result in great images, pretty much regardless of a model you use:Hi, just getting my feet wet with stable diffusion after trying it out once months ago, does anyone know what are the most photorealistic checkpoints now? gonna bother Mr-Fox as well
PS nice blondie, but not as nice as the water
PPS send help for my poor 8gb 1070 Ti, she's screaming in pain
Absolutely.Adding to the two posters above, going with the more commonly downloaded "realistic" models will keep you pretty well covered, minor differences, one might suit your needs better in one case, then another the next time. If you use lora/embeddings you might need to use a specific model for "best results". Depending on what you consider "best results" at that time.
I strongly advice staying away from low tier "merges" that's infesting things though. Generally those are poorly done and are leaching off of other peoples work, often without even bothering to give credit to those they are stealing from. And you end up having to fight issues with your renderings that you shouldn't have to deal with simply because others have taken shortcuts etc.
Depending on situation/imagery, i've had pretty good results with these, just to start off with:
You must be registered to see the links
You must be registered to see the links
You must be registered to see the links
You must be registered to see the links
You must be registered to see the links
If i'd have to pick just one i'd probably agree with Jimwalrus and go with cyber. It's being updated/worked on pretty often and has inpaint model too. So if bandwidth or hdd space is concern, i'd start there, and as Sepheyer mentioned, sampler matters too, so try different ones and see which you like.
Holy ! something.. Lost for words and this doesn't happen often believe me..I've been on thisYou must be registered to see the linksmodel kick, it even replaced Clarity for a quick minute:
View attachment 2686138
Holy ! something.. Lost for words and this doesn't happen often believe me..This is definitely your best work yet.
Amazing, I gotta try this one out.
I as well jumped into the A-Zovya photorealistic wagon.You must be registered to see the links
Great work.Does anyone old enough to remember the DOS game Dune from 1992?
Chani was one of my early crushes
View attachment 2687439
Well this is the remake from 2023
You don't have permission to view the spoiler content. Log in or register now.
If you want to get a decent prompt add the images to img2img and click Interrogate CLIP and it will give you details of the images. Also, click Interrogate DeepBooru and it gives you porn prompts.Where did you find them? Do you have the original png files? Those ones don't have the metadata you need to get an easy route to copying / cloning them
Interrogating the image though gives us this for the prompt that maybe you can use in a number of different Civitai models to see which fits best:
a sexy woman in a blue costume posing for a picture, hyper-detailed and realistic, fantasy dungeon, plus size woman, doll, hyper-realistic lighting, dark golden hair, featured on amiami, trending on artstartion, real-life big mom, blue skintight closed longcoat, ellora
--
Refining that with a bit of trial and error (we can do this better if we keep going), we can have this:
Positive: (masterpiece:1.2), (best quality), (intricate details:1.2), (photorealism), wallpaper 8k, a beautiful woman, wearing sexy revealing detailed dress, hyper-detailed, fantasy ornate castle window view, cleavage, wide hips, large breasts, sexy cloak, necklace ornament, (sheen to her skin), vibrant colors, (dynamic lighting:1.3), (vibrant cinematic lighting:1.3), (glowing lights and effects and lamps:1.2), lustrous golden hair, realistic skin, smiling, looking at viewer, facing front <lora:hipoly3DModelLora_v10:0.6>
Negative: poor quality, mutant hands, ugly, blurry, low quality, bad-artist, EasyNegative, facing front, boring, dull
--
That gives us something like the attachment - which we can see needs more lighting, and more colors... so you just keep adding to that in the prompt until you get something you need. If you have stable diffusion / automatic1111 - just throw the png into the png info tab and you'll get all the generation details.
View attachment 2531032 View attachment 2531040
It's a very good way to get started on the process of making a good prompt but it's very basic and needs a lot more. I haven't tried the DeepBooru option yet, thought it was only aimed at anime..If you want to get a decent prompt add the images to img2img and click Interrogate CLIP and it will give you details of the images. Also, click Interrogate DeepBooru and it gives you porn prompts.