- Nov 17, 2023
- 955
- 1,518
Thanks for the heads up! I am also using a site called Tensor these days. Which features upscaling etc.
- Jan 24, 2020
- 1,396
- 3,777
Controlnet with openpose for the pose. You can try using ip adapter to try retain the likeness of the character if necessary.
Guys, help me out on this.
Is there any good tutorial on fixing the hands on SD1.5 for both the manual way using photoshop and using the inpainting tools for from the sd webui.[For Anime images only]
Is there any good tutorial on fixing the hands on SD1.5 for both the manual way using photoshop and using the inpainting tools for from the sd webui.[For Anime images only]
- Jan 24, 2020
- 1,396
- 3,777
The genre doesn't matter, hands and toes are a constant issue with SD among other things.
There have been attempts to fix it with varying results. There isn't any perfect solution that works 100 percent of the time unfortunately. However, there are of course ways to improve the result and get something that doesn't look like a mutant.
Best practice is to try avoid bad hands to begin with.
Use a good resolution ratio.
Since you are using SD1.5 this means you need to keep the resolution under 1024 for the sample image otherwise you will likely get 2 merged images, resulting in mutants and aberrations. I would suggest using 512x768 (2:3 ratio) alternatively 576x768 (3:4 ratio), 640x960 (2:3 ratio) alternatively 720x960 (3:4 ratio) if you want a bigger image.
In my experience you get better results with a portrait ratio (2:3 or 3:4) rather than letterbox ratio (square 1:1), this might depend on how the checkpoint has been trained though.
Don't use too many loras or embedding at once.
When a lora or embedding is being trained, the entire subject is always being trained. This means that there is no such thing as only an outfit lora for example even though people claim this. So when you use many loras and/or embeddings you are more likely to have clashes that will result in mutants.
if you are using openpose, take the time to refine the pose in the editor especially the hands and fingers. The preprocessors are not even close to being perfect in it's detection.
Don't keep banging your head against the wall and trying to make a checkpoint model do something it has not been trained for or simply can't. If you keep getting bad results try several different checkpoints and see which gives the best result. For this purpose xyz plot script is great.
Go through the prompt and simplify it, remove any clashing words.
Most of the time I avoid using inpainting and simply generate more images until I get something good, but sometimes it can't be helped. After detailer has a hands model the will do inpainting for you on the fly, this doesn't work flawlessly but it can improve the result significantly when it works. I highly recommend getting adetailer and learning how to use it.
I have made tutorials exploring hands issues with inpainting manuallly. It's not perfect but it will at least point you in the right direction.
Other option is using the inpainting type plugin for the free image editor krita. I will link to this as well.
XYZ Plot Script Tutorial: https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-12816116
Fixing Hands with inpainting: https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-12545892
Krita Tutorial: https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-12643242
Last edited:
- Jan 24, 2020
- 1,396
- 3,777
Ps*
If you want better and more specific help you need to post an example image with the generation data included so we can take a look at the prompt and settings etc. We can then help you better with trouble shooting. Also give more descriptive info about what your goals are so we know what to aim for.
Last edited:
You must be registered to see the links
New 112 Image dataset of 1024x1024
I plan on releasing the SD version at some point.
Maybe a Pony training depending on the popularity of the model.
Maybe a Pony training depending on the popularity of the model.
I'm at 300+ trainings and today I had one of my worst trainings followed by one of my best.
The worst I knew was going to be bad, the image quality sucked, angles where not there.
But now for the best, I'm still amazed it is producing the results it is.
The 512x512 proof of text to image
1024x1024 with a Lora value of 1.0 and Clip strength doubled at a full image de-noise of 0.7
1536x1536 full image (Not Highrez fix) de-noise @0.5 on SD 1.5 not XL
The worst I knew was going to be bad, the image quality sucked, angles where not there.
But now for the best, I'm still amazed it is producing the results it is.
The 512x512 proof of text to image
1024x1024 with a Lora value of 1.0 and Clip strength doubled at a full image de-noise of 0.7
1536x1536 full image (Not Highrez fix) de-noise @0.5 on SD 1.5 not XL
Considering it was trained on 112 Images and fairly low rank....I am still in disbelief that it consistently able to go up to 1536x1536
I recently tried SDXL after lots of messing around with SD1.5
However I find that the controlnet openpose does not seem to work very well.
I did a "git pull" update in the "stable-diffusion-webui\extensions\sd-webui-controlnet" folder to get sdxl support.
And downloaded a model "controlnetxlCNXL_bdsqlszOpenpose.safetensors" which seems to be the right one?
Do i need to get the "
However I find that the controlnet openpose does not seem to work very well.
I did a "git pull" update in the "stable-diffusion-webui\extensions\sd-webui-controlnet" folder to get sdxl support.
And downloaded a model "controlnetxlCNXL_bdsqlszOpenpose.safetensors" which seems to be the right one?
Do i need to get the "
You must be registered to see the links
" model as well (or instead?) (it is very large)
Just get back to my pc after 3 weeks, do some update and the text-generation UI stops working, I am using oobabooga. Can someone take a look and tell how to fix this?
Edit: it seems possible to install a version in the past(thanks to github) now I just have to find the one that worked for me.
Are there any other text-generation UI, hopefully lighter than it?
Edit: it seems possible to install a version in the past(thanks to github) now I just have to find the one that worked for me.
Last edited:
This thread is about Stable Diffusion image generation, not LLM oobabooga etc.
The short answer to your actual question:
Oobabooga and the underlying libraries are in a constant state of churn and breakage. seems like something updated that needs another package updated.
1st thing to try would be the update_webui.bat (or whatever it is called now)
2nd would be move your downloaded models out to a temp folder (and anything else you want to keep like character defs, system prompts). Then delete the whole thing and reinstall fresh.
Last edited:
been making depth map poses from images found on net. is there a way to get rid of the background areas and just keep the figure? cause when u use the pose it renders whatever was in the map too. i just want the model depth pose created
example:
example:
modine2021 The depth map is just an image. You could download it, use an image editor to remove the background (replace it with pure black) and then re-upload as the controlnet depth mask.
The details of downloading/saving and re-uploading depend on the tool you are using. No guarantee that online AI generators allow that functionality. Local install of comfyui etc will definitely allow the setup though.
note that the AI will fill the black background parts with whatever is in your prompt, but is unlikely to do anything "close" there because of the black -> far away meaning.
The details of downloading/saving and re-uploading depend on the tool you are using. No guarantee that online AI generators allow that functionality. Local install of comfyui etc will definitely allow the setup though.
note that the AI will fill the black background parts with whatever is in your prompt, but is unlikely to do anything "close" there because of the black -> far away meaning.
im using this.. it has a n option to remove the background. but it looks a mess. or maybe im just doin it wrong:
You must be registered to see the links
I've not used that plugin before, I assumed you were using ControlNet Depth, which has the functionality I mentioned to upload your own source image for depth maps.
i use it cause u can batch files or whole folders. and it has a standalone version
