Doesnt seem to be an issues with your prompts (generated using png info > send to txt2img), have you checked the extensions tab for updates?
Doesnt seem to be an issues with your prompts (generated using png info > send to txt2img), have you checked the extensions tab for updates?
May be worth trying to increase the sampling steps and drop the denoising (40 sampling and .2 denoising used here)?
- Dec 23, 2018
- 469
- 1,939
Maybe the wrong VAE? I get something similar if I use a 1.5 vae with SDXL
- Oct 20, 2017
- 639
- 676
indeed, the wrong VAE. sdxl and sd15 have different ones
sticking with 1.5. sdxl is a headache. is there really much of a difference? maybe its my eyes
- Dec 31, 2016
- 316
- 708
Been testing some upscaling and "post processing", i guess it could be called.
I "stolen" a image from this post, so credit etc for original image should go there, to see how things worked for img2img.
Been having some horrible OOM errors which really doesn't make sense, IE failing to use 400mb more when it's only using 4 out of 6 gb...so things have gone even slower than usual :/
Anyway, just some results, really need to find a way to fix those hands, nothing really worked so far, but it's sd15 so might need to do some more scaling and maybe double masking
Need to work out the memory issues first though which is really annoying. Samplers are running really slow, not overflowing, but still deciding to run at 60-140 s/it..tiled vae OOM without even hitting the vram limit...getting OOM when samplers have been running fine for the previous steps with or without overflowing, then all of the sudden it OOM in the next step. Run/continue the prompt and it works fine. Really think they broke something somewhere
Edit: Adding json workflow file
I "stolen" a image from this post, so credit etc for original image should go there, to see how things worked for img2img.
Been having some horrible OOM errors which really doesn't make sense, IE failing to use 400mb more when it's only using 4 out of 6 gb...so things have gone even slower than usual :/
Anyway, just some results, really need to find a way to fix those hands, nothing really worked so far, but it's sd15 so might need to do some more scaling and maybe double masking
Need to work out the memory issues first though which is really annoying. Samplers are running really slow, not overflowing, but still deciding to run at 60-140 s/it..tiled vae OOM without even hitting the vram limit...getting OOM when samplers have been running fine for the previous steps with or without overflowing, then all of the sudden it OOM in the next step. Run/continue the prompt and it works fine. Really think they broke something somewhere
Edit: Adding json workflow file
Last edited:
- Dec 31, 2016
- 316
- 708
Since i'm doing alot of testing and trying to figure out which nodes to use and if there's any settings i can keep "static", it's safe to say it's a huge mess of similar nodes connected to the same things and a lot of bypassing and "wiring". I'll try to find some time to clean up much of the mess and post that.
If you want a "basic" run down of things though to have a go at something similar;
- Create or load a base image
- (optional) upscale image/latent slightly
- add noise/denoise slightly and run through sampler using the same model and prompt as you create the image with
- add nose/denoise at a lower amount than in the step above, sample with same prompt by different model
- (optional) upscale image/latent slightly
- last "stage" uses a third model and it's own prompt to clean up, highlight, style, detail or whatever you want to accomplish. So the sampling can either be a finish with just the last few steps or adding a bit of denoising and running that.
For me it takes 25-30min for a full run, with two of the models being sd15, could probably cut that down by removing some of the vae nodes, but when you want/need to keep track of changes they are kind required.
I've included 3 images from my last test just to show the stages, it's not that good an example as the changes in the last two images isn't all that large, might need to keep flipping between them to really see. #1 is the base image.
Edit:
workflow file with nodes etc added to this post
Last edited:
- Dec 31, 2016
- 316
- 708
I've added a somewhat cleaned up json (zipped as it can't upload otherwise) to the post with the two first images and linked in the other post.
The json has the same nodes, linked the same way/order etc as what was used for all the posted images. Exact settings for steps, which models etc i have no idea what was used in each case. If you check the name for the first two images you might notice that they are comfyui temp files which gets wiped on restarting comfy, so i don't even have them any more.
I think the nodes are limited to base, WAS and comfyroll.
Note that this is setup for loading a base image, so that small bit can simply be replaced with a pretty much just a empty latent and sampler etc, if anyone want to use it "directly". Also top chechpoint loaded (and prompts) are set up for sdxl, so that needs to be tweaked if anyone want to use just sd1.5
I have come across a few image like this on Civitai, while it has detected certain lora used but I don't see the usual lora format <lora_name: strength> in the prompt, how is that possible?
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
- Dec 31, 2016
- 316
- 708
Comfyui doesn't include lora in the prompt it's loaded in separate nodes. I believe there's still subsection in a1111 that lets you select loras and set their weights directly in a set of dropdowns and input fields as well.
- Jan 24, 2020
- 1,401
- 3,791
Fixing hands etc.
Something I have experimented with is to use after detailer in inpaint and only one hand at a time. I first mask the hand as best I can and only have hand_yolov8n.pt activated in adetailer then I generate and with each attempt I play with the settings until I get a good result. You have many parameters to adjust. What seems to do the most difference is the "normal" denoising strength and the inpaint denoising strength in adetailer don't go past 0.4 seems to be the rule of thumb in either. I switch between using "only masked" and "whole picture", I also switch between using the og seed and random. What works best seems to be scenario or case dependent. You can also use a specific prompt for only hands while inpainting. Steps and resolution is something I have not experimented with very much but could be relevant.
For demonstration purposes I generated a simple image with the LCM extension.
As you can see the left hand needs some fixing, not the worst case obviously but a good example nonetheless.
When inpainting there are a few helpful quick commands you can see if you hover the mouse over "i" in the image box.
I always zoom in to make it easier to mask the hand and typically I mask the entire hand from the wrist. If you have a mouse that has dpi settings change it for a slower input, this also makes it easier ofc.
I start out with 0.1 denoising and for adetailer inpaint denosing I typically use 0.38 . Only masked and the og seed was used.
We could be happy with this or experiment a bit like I always do. I leave the normal denoising at 0.1 and lower the adetailer inpaint denoising to 0.28 .
The hand is slightly sharper but the fingernail on the middle finger is now a little deformed.
It's always worth experimenting but in this case the previous result was better.
Ok so lets finish off this image. I do the same process in inpainting for the face with adetailer mediapipe_face_mesh_eyes_only, adetailer inpainting denoising 0.22 and lips with adetailer inpaint denoising 0.3 and 0.01 for "normal" denoising and I'm also using GFPGAN postprocessing. I didn't add any prompt for the lips in adetailer.
I'm doing this to improve the face,lips and eyes and get better details before upscaling. Why I did it in inpaint is so I will not lose the progress of the hands.
So lets upscale. I'm using SD Upscale in the script menu with 4x_NMKD-Siax_200k upscaler, only 2x for this demo and with 0.01 denoising. This is because I'm using LCM models and it seems that all the numbers is decreased greatly even the denoising strength for a good result. It's important to not use postprocessing when upscaling because it will mess with the face too much and undo our progress. You can use adetailer but don't use it for hands, only eyes and maybe lips because again it will undo the progress with the hands otherwise.
The final result:
I hope this was helpful and if you have any questions or comments I will try to answer or if someone else wants to chime in, it's ofc welcome.
PS.
If you are wondering what the fudgestick is LCM? Latent Consistency Model and it's the latest "rage" or trend. The point of it is to use much less steps and lower cfg scale to decrease generation times but still get a good result.
There is "Turbo" for SDXL models and LCM as well.
Here's an article about it on civit if you want to learn more.
Something I have experimented with is to use after detailer in inpaint and only one hand at a time. I first mask the hand as best I can and only have hand_yolov8n.pt activated in adetailer then I generate and with each attempt I play with the settings until I get a good result. You have many parameters to adjust. What seems to do the most difference is the "normal" denoising strength and the inpaint denoising strength in adetailer don't go past 0.4 seems to be the rule of thumb in either. I switch between using "only masked" and "whole picture", I also switch between using the og seed and random. What works best seems to be scenario or case dependent. You can also use a specific prompt for only hands while inpainting. Steps and resolution is something I have not experimented with very much but could be relevant.
For demonstration purposes I generated a simple image with the LCM extension.
You don't have permission to view the spoiler content.
Log in or register now.
As you can see the left hand needs some fixing, not the worst case obviously but a good example nonetheless.
When inpainting there are a few helpful quick commands you can see if you hover the mouse over "i" in the image box.
I always zoom in to make it easier to mask the hand and typically I mask the entire hand from the wrist. If you have a mouse that has dpi settings change it for a slower input, this also makes it easier ofc.
You don't have permission to view the spoiler content.
Log in or register now.
We could be happy with this or experiment a bit like I always do. I leave the normal denoising at 0.1 and lower the adetailer inpaint denoising to 0.28 .
The hand is slightly sharper but the fingernail on the middle finger is now a little deformed.
It's always worth experimenting but in this case the previous result was better.
Ok so lets finish off this image. I do the same process in inpainting for the face with adetailer mediapipe_face_mesh_eyes_only, adetailer inpainting denoising 0.22 and lips with adetailer inpaint denoising 0.3 and 0.01 for "normal" denoising and I'm also using GFPGAN postprocessing. I didn't add any prompt for the lips in adetailer.
I'm doing this to improve the face,lips and eyes and get better details before upscaling. Why I did it in inpaint is so I will not lose the progress of the hands.
So lets upscale. I'm using SD Upscale in the script menu with 4x_NMKD-Siax_200k upscaler, only 2x for this demo and with 0.01 denoising. This is because I'm using LCM models and it seems that all the numbers is decreased greatly even the denoising strength for a good result. It's important to not use postprocessing when upscaling because it will mess with the face too much and undo our progress. You can use adetailer but don't use it for hands, only eyes and maybe lips because again it will undo the progress with the hands otherwise.
The final result:
I hope this was helpful and if you have any questions or comments I will try to answer or if someone else wants to chime in, it's ofc welcome.
PS.
If you are wondering what the fudgestick is LCM? Latent Consistency Model and it's the latest "rage" or trend. The point of it is to use much less steps and lower cfg scale to decrease generation times but still get a good result.
There is "Turbo" for SDXL models and LCM as well.
Here's an article about it on civit if you want to learn more.
You must be registered to see the links
Last edited:
- Dec 31, 2016
- 316
- 708
There's lora versions for turbo too and some are made in versions aimed at specific samplers and they work fairly well with most models i've tested it with. Seems more "trainers" are including one or both of these methods in their models as well so if ppl are updating those it's worth checking as it can screw with out results if you're running at old cfg/steps.
There's something called DPO as well, from what i gathered from glancing at some stuff, it's meant to follow prompts more "accurately". If it actually works or is any improvement i can't say, not had a chance to test it yet
- Jan 24, 2020
- 1,401
- 3,791
There is a LCM sampler, it is a bit muddy atm how you get this though. According to the article you get it when you install animatediff. I have the sampler but I'm not sure when or how I got it. I had installed animatediff while experimenting with gif and vid making and then I also got the LCM extension. It was after this I discovered I had a new sampler. It works much better for any additional LCM model you download other than what is included with the extension. It's worth getting the extension though for very fast simple image generation. It comes with a special integrated version of dreamshaper 7 specific for the extension. The settings and prompt is very limited but it's for small and simple images only and you can make batches of 100 images. It also has img2img and vid2vid.
Last edited:
- Nov 29, 2022
- 64
- 2,735
This is a broad question and maybe this has been answered before, but what are tips for writing prompts that involve sex positions?
I can do single woman poses with no problems but when ever I do a prompt with a Lora pose, it turns into a nightmare image.
Sometimes I would even copy prompts word for word of another picture and get the same nightmare results.
Any idea how I should approach this?
I can do single woman poses with no problems but when ever I do a prompt with a Lora pose, it turns into a nightmare image.
Sometimes I would even copy prompts word for word of another picture and get the same nightmare results.
Any idea how I should approach this?
- Jan 24, 2020
- 1,401
- 3,791
The issue is not the prompting. Any interaction between subjects is very problematic with stable diffusion, it seems to be the interference between objects and between persons that is the problem. What is the difference between two different surfaces touching and one single object?.. Loras and controlnet etc is likely the solution.
Last edited:
- Nov 29, 2022
- 64
- 2,735
I've been using loras but not controlnet. So, I am guessing all those AI sex images (blowjob, reverse cowgirl, etc...) were done using controlnet?
I guess I was just hoping it was going to be a lot more straightforward lol
- Mar 19, 2020
- 2,075
- 2,609
I'd try ControlNet as well instead of LoRAs -- because the few times I've tried to use LoRAs for a pose, it always kinda "bleed" too much into the original checkpoint/model I was using and making the picture worse if that makes sense.
The way I'd do it is to find a decent reference picture and use ControlNet and/or "MultiArea Conditioning" if you want to have multiple characters in your picture that are not directly "connected" to or touching each other. Probably won't work with most sex poses though...
The way I'd do it is to find a decent reference picture and use ControlNet and/or "MultiArea Conditioning" if you want to have multiple characters in your picture that are not directly "connected" to or touching each other. Probably won't work with most sex poses though...
- Nov 29, 2022
- 64
- 2,735
I might just try to make poses in Daz and make it work with Controlnet. Thanks for the info.