AI Technical Discussion and Info

[SonicIsBack]

I've been working on recreating a classic "candid" style circa mid 2000's smartphones, I used my own candid pictures for the LORA training and with a few tweaks it's looking pretty good so far. I've figured out a few tricks to recreate the slightly crappy photo quality, this definitely gives it more of a real look imo. I find one of the main problems with AI is it makes everything look slightly "off" and plastic/soulless, but adding a bit of low quality vibes makes it way more believable....even if the sizes are not!

Here are the LORAs that I used other than my own unreleased LORAs (these are the ones available online) .

<lora PAseer-SD15-LCM Quick:1>
<lora:skin_slider_v1:0.4>
<lora:add_detail:0.4>
<lora:anti_longtorso_v10:1>
<lora:depth_of_field_slider_v1:-3>
<lora:lyco_humans_v10:0.4>
<lora old_siemens:0.2>
<lora:breastsizeslideroffset:> anywhere from 0.4 to 1.4 depending on the next LORA, higher values if using lower weight values
<lora:weight_slider_v2:> -2 to 2

 

[hindrancestrickish]

Candid street photos of busty women have always been the pinnacle for me. For the last year I’ve been experimenting with AI in the Photoleap app trying to create the most convincing “veritas“ images with some pretty good raw results. Then I remembered I purchased the Hipstamatic app years ago which mimics analog film lenses and flashes and processed my raw AI images for an added dimension. This prompt is of Christina Hendricks walking in New York generated in the IOS PhotoLeap app and face swapped with another model in FaceApp where I also added glasses. Finally I uploaded through Hipstamatic for a Lomo style candid snap of a stacked bra buster making her way home on a warm evening.

 

[SuperBoobs]

So, I got this error and wrestled with it for a couple days.
The solution looks like you need a degree in Python programming language.
Although, I did see comments about decreasing batch size and the number of batches, and so on.
What finally worked for me was to keep my width and height measurements reasonable. I was pushing out 800x1000 and have decreased to 500x700. It seems to have fixed it up for me.

What works well for me is to keep the first image size low and do a resizing afterwards.
You can even repeat reszing to create really large images.

 

[SonicIsBack]

I'm finally taking the plunge into SDXL......which means remaking all of my LORAS in XL format, hoping the results will be worth it. Does anyone have any recommended checkpoints/LORAs? I've heard RealVisXL is really good and it has a lightning version which is needed as I'm pushing the upper limits on VRAM.

 

[Upscaler]

I'm finally taking the plunge into SDXL......which means remaking all of my LORAS in XL format, hoping the results will be worth it. Does anyone have any recommended checkpoints/LORAs? I've heard RealVisXL is really good and it has a lightning version which is needed as I'm pushing the upper limits on VRAM.

RealVisXL works very well. If you want X-rated scenes and can tolerate a dip in quality check out GroverNSFWmix.

 

[SonicIsBack]

RealVisXL works very well. If you want X-rated scenes and can tolerate a dip in quality check out GroverNSFWmix.

Yeah, RealVis is by far the best I've used so far....the lighting and backgrounds are such a step up from 1.5, it's amazing!

 

[cy5]

So FYI (and also a plug for me) - if you're just starting out or don't have the hardware to do stable diffusion, I highly recommend Diffus. It's easy to use and has a fairly extensive Library of LoRAs and checkpoints.

If 3 more people sign up we all get 1000 credits, which is more than enough to get started generating or inpainting.
Link is: https://webui.graviti.com/

Here are some examples of what i've made:

 

[SonicIsBack]

So FYI (and also a plug for me) - if you're just starting out or don't have the hardware to do stable diffusion, I highly recommend Diffus. It's easy to use and has a fairly extensive Library of LoRAs and checkpoints.

If 3 more people sign up we all get 1000 credits, which is more than enough to get started generating or inpainting.
Link is: https://webui.graviti.com/

Here are some examples of what i've made:

Civitai is also really good if you don't have the GPU power.

 

[SonicIsBack]

SDXL is a game changer, it's amazing!

 

[CAgunner]

a third school had students make AI images of other student in California, and i cant figure any of this out

 

[Upscaler]

Upscaling shitty photos using SUPIR.
ComfyUI workflow: https://flowt.ai/community/supir-v2-plugandplay-edition-n5acf-v
Tutorial: https://www.youtube.com/watch?v=EMAz8KktB5U

 

[41boner]

Upscaling shitty photos using SUPIR.
ComfyUI workflow: https://flowt.ai/community/supir-v2-plugandplay-edition-n5acf-v
Tutorial: https://www.youtube.com/watch?v=EMAz8KktB5U

Have you used automatic1111? If so, what do you like about comfyUI better?

 

[Upscaler]

Have you used automatic1111? If so, what do you like about comfyUI better?

I actually prefer using automatic1111. I just don't know of an easier way to use SUPIR.

 

[SonicIsBack]

I've been experimenting with img2img, using PonyDiffusion to create images then running them through img2img (and controlnet canny) using a realistic checkpoint.

I created a ComfyUI workflow, it's a bit of a mess but does the job. Or you can just use Automatic1111, it just takes a bit longer.

Pony allows you to create poses/breast shape and sizes that you normally couldn't with realistic checkpoints.

 

[Tezcat]

I've been experimenting with img2img, using PonyDiffusion to create images then running them through img2img (and controlnet canny) using a realistic checkpoint.

I created a ComfyUI workflow, it's a bit of a mess but does the job. Or you can just use Automatic1111, it just takes a bit longer.

Pony allows you to create poses/breast shape and sizes that you normally couldn't with realistic checkpoints.

Love it! do you mind sharing the workflow?