Creations by GravidaSomnia

[gravidasomnia]

I've just uploaded a Lora! - PreggoXL on civitai.com
alt link: PreggoXL on huggingface.co


It's definitely alpha version far from perfect but it's a good start. V1.0 will need a lot more work but I think I can get better quality and more nuanced prompting out of it.

   
Prompt: your pregnant big tiddy goth GF wearing a sheer gothic lace dress <lora:preggoXL_v0.1a:0.7>
Negative prompt: crosseyed
Steps: 32, Sampler: DPM++ 3M SDE, CFG scale: 4, Size: 768x1024, Model: xl_realismarkPlus_realismarkPlus, VAE: sdxl_vae_fp16fix.safetensors,  Lora hashes: "preggoXL_v0.1a: 780246d31fe8"

   
Prompt: pregnant clubgirl friends dancing under disco lights wearing party dresses <lora:preggoXL_v0.1a:0.4>
Negative prompt: blurry, blurred, painting, illustration, cartoon, 3d, cgi, render, vagina, explicit, porn, young, teen
Steps: 20, Sampler: Restart, CFG scale: 4, Size: 1024x768, Model: Photosomnia XL v0.9, VAE: sdxl_vae_fp16fix.safetensors, Denoising strength: 0.3, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Hires sampler: DPM++ 2M, Hires upscale: 1.5, Hires steps: 12, Hires upscaler: 4x-UltraSharp, Lora hashes: "preggoXL_v0.1a: 780246d31fe8", Discard penultimate sigma: True, Extra noise: 0.02

   
Prompt: pregnant woman at the park<lora:pxl2:1>
Steps: 32, Sampler: DPM++ 2M SDE, CFG scale: 4, Size: 768x1024, Model: Photosomnia XL v0.9, VAE: sdxl_vae_fp16fix.safetensors, Denoising strength: 0.2, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Hires upscale: 2, Hires steps: 16, Hires upscaler: 4x-UltraSharp, Lora hashes: "pxl2: 780246d31fe8", Discard penultimate sigma: True, Eta: 0, Sigma noise: 0.9, Extra noise: 0.02

   
Prompt: a pregnant female ww2 fighter pilot propaganda poster <lora:pxl2:0.8>
<no negative prompt>
Steps: 30, Sampler: Restart, CFG scale: 4, Size: 1024x768, Model: xl_realismarkPlus_realismarkPlus, VAE: sdxl_vae_fp16fix.safetensors, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Lora hashes: "pxl2: 780246d31fe8", Discard penultimate sigma: True

   
Prompt: pregnant preggo catgirl wearing a traditional french maid costume patrolling the postapocalyptic wasteland with a wolf <lora:SDXL pregnancy v0.1a:0.4>
<no negative prompt>
Steps: 32, Sampler: DPM++ 3M SDE, CFG scale: 4, Size: 768x1024, Model: xlm2.fp16, VAE: sdxl_vae_fp16fix.safetensors, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Lora hashes: "SDXL pregnancy v0.1a: 780246d31fe8", Eta: 0, Sigma noise: 0.9

   
Prompt: pregnant preggo babe at the poolside <lora:SDXL pregnancy v0.1a:0.4>
Negative prompt: crosseyed
Steps: 32, Sampler: DPM++ 3M SDE, CFG scale: 4, Size: 768x1024, Model: xlm2.fp16 VAE: sdxl_vae_fp16fix.safetensors, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Lora hashes: "SDXL pregnancy v0.1a: 780246d31fe8", Eta: 0, Sigma noise: 0.9

   
Prompt: pregnant babe at the poolside <lora:preggoXL_v0.1a:0.6>
Negative prompt: crosseyed
Steps: 32, Sampler: DPM++ 3M SDE, CFG scale: 4, Size: 768x1024, Model: caroselloxlOldFlat_10, VAE: sdxl_vae_fp16fix.safetensors, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Lora hashes: "preggoXL_v0.1a: 780246d31fe8", Eta: 0, Sigma noise: 0.9

   
Prompt: pregnant blonde woman in a one-piece swimsuit at the poolside <lora:preggoXL_v0.1a:0.7>
Negative prompt: crosseyed
Steps: 32, Sampler: DPM++ 3M SDE, CFG scale: 4 Size: 768x1024, Model: xl_realismarkPlus_realismarkPlus, VAE: sdxl_vae_fp16fix.safetensors, NGMS: 0.2, Hypertile U-Net: True, Hypertile VAE: True, Lora hashes: "preggoXL_v0.1a: 780246d31fe8", Eta: 0, Sigma noise: 0.9

[gravidasomnia]

I'm trying to shift to linux because most of the local running video tools don't work on windows (mostly bc Triton isn't supported) and unlike animatediff/SVD they aren't likely to be getting a1111 extensions/comfyUI nodes. it's frustrating as heck to muddle through it all if you're not a developer and only have limited time to work on getting the software working. But looking at what it can do in the animated AI thread it's gonna be worth the extra effort. 

My lora does okay bellies but it's lousy for some other details e.g. no pubic hair, and cutout crotches on legwear, which I'm sure is appealing to some but isn't what I'm after when doing clothed pics.

                   
       
           
           


The latest Patreon/Deviantart set for supporters is loose dressing gowns:
               

[gravidasomnia]

I've taken down the PreggoXL_v0.1a Lora pending further training - with the right (wrong?) prompts it replicates some of the dataset all too well and strongly resembles one of the models rather than being random girls. A replacement v0.2a should be available shortly with the issue addressed. I've also started the manual captioning process because the auto tool created a mess of inaccuracies & irrelevant tokens to the point that I can't believe it worked so well. It will be interesting to see what a difference the improved captioning will make but for now I think it'd just be good to have something that works at all available for anyone to use, for all its flaws the lora still gives a big improvement over the base models.
Also I'm hitting a brick wall trying to get more advanced video tools to work, the standalone demos seem impossible to install properly and the comfyUI nodes/A1111 extensions seem to break other nodes/extensions.



       

           

[gravidasomnia]

Apologies for not posting more here, I'm spread too thin across too many platforms and don't have as much time to spare as I used to.
The preggoXL lora has been reuploaded to civitAI with the overtraining issue addressed. It's still a minimum effort Lora until the captioning is done manually, hopefully I can make some progress on that this week. 


   
   
   
   
   
   
   
   
   
   
   
   

[gravidasomnia]

                                       

[bpuk]

you've done some great work thanks for sharing

[gravidasomnia]

Thanks! The AI does 98% the work and anyone who can run Stable Diffusion can get similar results, it's easier than ever to install thanks to Stability Matrix, but the hardware requirements to run SDXL limit accessibility. 

   
   
   
   

[gravidasomnia]

Happy Valentine's Day!

   
   
   
   
   
   
   

[gravidasomnia]

   
   
                   
   

[gravidasomnia]

Also I've updated the preggoXL lora, I never got around to finishing captioning for the dataset but I have tried out civitAI's training with what I had on hand. If you use SDXL please give it a try, eventually it should be available to use on civitAI as well. Here's the results with some simple test results tried out across a range of SDXL checkpoints:

                   
                   
                   
                   
           

generation data is available on the civitai page.