Bo27's Stable Diffusion work

[User 77373]

Absolutely stunning! Pregnancy details aside, the realism seems to exceed most of the publicly available photorealism models, which often need super-low CFG values to get photorealistic results at the expense of detail. That's the kind of quality I hope to achieve one day, plus the girls are definitely more or less my type ;-). Your hard work is paying off amazingly well.

[bo27]

These are done with CFG level 7 or 7.5, haven't tried to lower to get more realism yet.

[User 77373]

It'd be interesting to see what changes,  lower CFG also reduces how responsive the AI is to the prompts. Sometimes it's a pain trying to find a CFG value that's just right for some of the pics I've worked on, many promising compositions have had to be scrapped because they're beyond my touch-up skills to fix.

Using "twins, triplets" etc. with a regular model is virtually guaranteed to produce multi-character outputs, but your model seems to be laser focused on pregnancy content with those tokens and I assume that's the case for other much-loved pregnancy descriptors  (e.g. I suspect that the underlying interpreter doesn't distinguish between "buttons" and "bellybuttons" based on the outputs I'm getting.) As a comparison here's the metadata for one of your pics run through various photorealism models:

Bo27's original
vs. same prompt and settings, different models
                                                   

So in case you were wondering, that's what you're missing out on lol. Maybe the settings need adjusting but aside from the mismatching and other errors the pic quality from the other models is lower all around using those parameters. Your model seems to do a lot more than just great pregnancy details and imo you're getting in at the ground floor of something that's going to grow. I don't expect your model to be made public but if you could find a way to host it somewhere on your own terms that would be awesome.

[peterwein]

@bo27 Wow amazing pics again!

Did you use any LORA for the belly buttons or is it straight from your custom model?

[bo27]

@bo27 Wow amazing pics again!

Did you use any LORA for the belly buttons or is it straight from your custom model?

My custom model.

[bo27]

It'd be interesting to see what changes,  lower CFG also reduces how responsive the AI is to the prompts. Sometimes it's a pain trying to find a CFG value that's just right for some of the pics I've worked on, many promising compositions have had to be scrapped because they're beyond my touch-up skills to fix.

Using "twins, triplets" etc. with a regular model is virtually guaranteed to produce multi-character outputs, but your model seems to be laser focused on pregnancy content with those tokens and I assume that's the case for other much-loved pregnancy descriptors  (e.g. I suspect that the underlying interpreter doesn't distinguish between "buttons" and "bellybuttons" based on the outputs I'm getting.) As a comparison here's the metadata for one of your pics run through various photorealism models:

Bo27's original
vs. same prompt and settings, different models


So in case you were wondering, that's what you're missing out on lol. Maybe the settings need adjusting but aside from the mismatching and other errors the pic quality from the other models is lower all around using those parameters. Your model seems to do a lot more than just great pregnancy details and imo you're getting in at the ground floor of something that's going to grow. I don't expect your model to be made public but if you could find a way to host it somewhere on your own terms that would be awesome.

Interesting. I posted earlier that I am using everydreamtrainer2 if you want to check it out, its way better than dreambooth, every picture I reference in the training set I have extremely detailed captions and I go into intense detail. If the face is ugly but the body looks great, i call it out, if her face and body look beautiful but her hands or feet are swollen, I call it out, but then I can call out those details and get great results. The expression, the lighting, the level of detail, if its in focus, out of focus, blurry, all called out, if the photo is exceptional I make sure to say "reference hands, detailed skinned, extremely accurate details, female hands, pregnant navel', same if parts of the body look great but other parts are blurry, etc. I mention their stance, gait, positioning, do they have some stretch marks, a lot of stretch marks, perfect complexion, et, are breasts full, firm, or do they sag, etc..Maybe the photo quality looks good (studio) etc, but resolution or details are low, i will call it out. Auto captions is terrible for pregnancy, hands are on the hips, or butt exposed, will be the outcome, and its mostly wrong. So I overwrite most of it. I am updating a newer model, I literally and using model hands etc. Pexel for example has a lot of UHD copyright free material. The language used builds the model. not sure if it helps, I mean I caption bellybutton or belly button, but I think navel is the proper term, so navel, pregnant navel, etc. Stretch navel, distended or popped navel, etc.


Also I raised and lowered my config 2-12, it looks mostly good all the way through, hard to say which way I like it better, but pretty nice.

[Sweetboy2500]

Killer work on this model

[bo27]

I need to get the hands right, heres where I am at, most of it is just the flexibility of keywords

[bo27]

More examples here

[bo27]

Final bits for today