Did DPM++ SDE Karras disappear for anyone else? It was there a couple of days ago, now its not in the list.
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.
update to UI, its under schedule now or something
Regarding the installation steps at the top of the thread,...In step one of Installing:
If the "Download Zip" method is used, after downloading and unzipping locally and running the "webui-user.bat" it throws this error:
ERROR: Could not find a version that satisfies the requirement torch==2.1.2 (from versions: 2.2.0, 2.2.0+cu121, 2.2.1, 2.2.1+cu121, 2.2.2, 2.2.2+cu121)
ERROR: No matching distribution found for torch==2.1.2
If installing using the "get clone" method, no error is seen but there is no way to know where the "webui-user.bat" is physically located in order to run it from Windows Explorer.
Trying to open in a browser it is simply treated as if no such thing exists.
If the "Download Zip" method is used, after downloading and unzipping locally and running the "webui-user.bat" it throws this error:
ERROR: Could not find a version that satisfies the requirement torch==2.1.2 (from versions: 2.2.0, 2.2.0+cu121, 2.2.1, 2.2.1+cu121, 2.2.2, 2.2.2+cu121)
ERROR: No matching distribution found for torch==2.1.2
If installing using the "get clone" method, no error is seen but there is no way to know where the "webui-user.bat" is physically located in order to run it from Windows Explorer.
Trying to open in a browser it is simply treated as if no such thing exists.
Is this superior to using fooocus? I often use the image prompt part of fooocus, does this have the same?
did anyone knew what prompt or settings for better result for photo that taken by flashlight?
Prompt :
(nudity:1.2),(shirtless:1.2), full nude, a woman, nipples, small breast, asian, completely nude, small nipple, small areola, pale skin, perfect breast, shaved, perfect pussy.
Result :
Fill (this the best result and i try Lazimix, URPM, PornMaster, PicxReal) but the result there always some kind of shadow in it.
Latent Noise
Latent Nothing
Original
Prompt :
(nudity:1.2),(shirtless:1.2), full nude, a woman, nipples, small breast, asian, completely nude, small nipple, small areola, pale skin, perfect breast, shaved, perfect pussy.
Result :
Fill (this the best result and i try Lazimix, URPM, PornMaster, PicxReal) but the result there always some kind of shadow in it.
Latent Noise
Latent Nothing
Original
my pc is
i5-4590
RAM 16Gb
RX570 4Gb
my seting is
set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--precision full --no-half --upcast-sampling --opt-sub-quad-attention --disable-nan-check --medvram --skip-torch-cuda-test
but why take 30 min to inpainting 1 pic?
i5-4590
RAM 16Gb
RX570 4Gb
my seting is
set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--precision full --no-half --upcast-sampling --opt-sub-quad-attention --disable-nan-check --medvram --skip-torch-cuda-test
but why take 30 min to inpainting 1 pic?
CUDA is only for NVIDIA GPU, you are only using your CPU + RAM
Try some of these things:
- Inpaint area: only masked and "Fill" is the correct mode, start with fill to get something and when you get close to what you want, switch to "Original" to finalize.
- You won't get a good result in one generation, to save time, drag and drop the result image back into the inpaint window again to reuse the same mask each time.
- putting "clothes" in the negative prompt.
- Up the CFG Scale. Standard is 7, up to 9-10 maybe even more if it works it works. Same with denoise strength 70-80 to begin with, 15-30 when finalizing.
- Lower sampling steps: Try 20 at first. Fewer steps give a more basic result which can be good for adding new stuff.
- When masking, mask the main part you want, boobs in this case. Then add small dots above her head and below her pelvis. Sometimes all the way down at her feet. The purpose is to make sure the software detect enough of the person to place the tits in the right place. Only what you mask is what is processed. This will also help picking the correct skincolor etc.
- Instead of using dots, you can also increase the "Only masked padding, pixels", it will accomplish something similar.
- Keep the prompts very simple in the beginning, for example "naked asian girl with perfect tits" and negative: clothes. It makes it easier to detect if any of them are causing problems. When you're finalizing you can select more of the image and use more prompts to get exactly what you want.
- The new "Soft Inpainting" can help, but I have not used it a lot yet. But just the default settings seem to do wonders sometimes.
- ControlNet really is a gamechanger, allows you to run fill with way higher Denoising strength settings without destroying an image. It takes some time to download, but so worth it in the long run. Stuff that would take ages before, can be done in a few generations. Look into ControlNet and the DepthAnything model for ControlNet.
Editing:
- If all else fails, you can destroy part of the image to get what you want. Either by generating using "Fill" with a low Denoising strength to give you just a blurry blob and then turn the Denoising strength back up to try and generating something new over it.
- Or what I did, using Inpaint Sketch, and send it back to regular inpaint and using "Original" with about 50% Denoising strength to generate again. Then touching it up with even lower Denoising strength more sampling steps.
Original with Inpaint Sketch, don't generate here, just paint over and move it back to inpaint.
The mask with a dot at the bottom.
First try.
Touchup with about 80 sampling steps.
Hopefully some of this stuff was useful.
I'm not a technical person and this is all a little complicated for me. I've used mage.space in the past for NSFW inpainting, but it doesn't do a great job. Are there other websites you can use to produce high-quality inpainting?
If i dont use '--skip-torch-cuda-test' always "Torch is not able to use GPU"
damn king, youre the best, i read every word you type, awsome explanation too, thank you so much dude 🔥🔥🔥