I'm just testing some embeddings. I tried using the urpm model and it spat out images with fluoro/neon colours and deformed features. When I tried with the stable diffusion base the colours and shapes were more natural looking.
Guide Installing Stable Diffusion Webui & Nudifying (Inpainting) Guide
- Thread starter softarchitect
- Start date
-
Welcome to the Fakes / AI / Deepfakes category
When posting AI generated content please include prompts used where possible.
It depends on the image.
Most upscalers in A1111 are not great, Real-ESRGAN works fine for upscaling an already good quality image (> 500px) to higher resolution (x2 or x4). Codeformers works best if you only want to increase the quality of the face.
There are other models (based on ESRGAN) here: , add them in models/ESRGAN
Have not tested any of them, but heard that "remacri" is good for landscapes and "lollypop" for portraits, YMMV.
IMHO for best quality, ppl buy Topaz products (i.e. )
Something that I have been doing recently that hasn't really been said yet (or I missed it):
Go bit by bit! When inpainting an image, try inpainting a single section with a specific prompt, then, pick the best output and upload it to inpaint another section and repeat. I find this to be effective in images with multiple subjects or "difficult" lighting, clothing, filters on top, etc.
Doing this not only allows you to be autistically specific with the prompt you can also change the denoising, cfg scale, etc. per section resulting in a better image than if you just used a single seed over the whole image, in my experience.
Go bit by bit! When inpainting an image, try inpainting a single section with a specific prompt, then, pick the best output and upload it to inpaint another section and repeat. I find this to be effective in images with multiple subjects or "difficult" lighting, clothing, filters on top, etc.
Doing this not only allows you to be autistically specific with the prompt you can also change the denoising, cfg scale, etc. per section resulting in a better image than if you just used a single seed over the whole image, in my experience.
Hi I'm running this with an AMD Radeon RX580 8GB, had some issues at first but got it running with arguments: --opt-sub-quad-attention --lowvram --disable-nan-check
Problem now is that it starts generating the nude, finishes without error, but returns the original with black mask. Any ideas?
Problem now is that it starts generating the nude, finishes without error, but returns the original with black mask. Any ideas?
Last edited:
From my understanding, Stable Diffusion by default will try to run at half-FPU meaning fp16. This requires an nVidia CUDA core to run that way, so you will need to add the argument --no-half to your startup so you will use full floating point math in the generation. I have a 6950 XT which is a few generations over your card, but the arguments I'm using are --no-half --precision full --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 ; so you could try to add a few more to your arguments and see if that helps. It did for me.
You can but it's super difficult. You'd actually be better off just learning to "manually" edit dicks in mouths (using photoshop or a free alternative).
This was helpful for my all AMD system. Faster than the old fashion way these. Yep, smexy results too! Thank you to those others, including the original poster, for helpful advice. 👍
depends, if youre inpainting i havent found something good that works more then 10% of times, if your creating using txt2img and not inpainting then use the tgirls Lora in civitai it works pretty well after some generations(especially if you merge the models he suggested to merge and use it)
my prompt is mainly "tgirl", "RAW image of a nude woman shwoing her erect penis" or w/e you want those just work good with the txt2img tgirls Lora.
I am inpainting. I will continue testing and see if I can figure it out
Thanks, this helped on my RX6900XT , no longer an blurry result even if you can see it working during proccessing
hey i been getting this error when i tried
RuntimeError: "log_vml_cpu" not implemented for 'Half'
ajusted with
--skip-torch-cuda-test --precision full --no-half --no-progressbar-hiding --opt-channelslast
thanks to the people on the tread
RuntimeError: "log_vml_cpu" not implemented for 'Half'
ajusted with
--skip-torch-cuda-test --precision full --no-half --no-progressbar-hiding --opt-channelslast
thanks to the people on the tread
Last edited: