- Jan 24, 2020
- 1,401
- 3,802
- Sep 15, 2021
- 1,042
- 3,984
Indeed! One thing I discovered when collecting the dataset for the TI training - Cameron Diaz has terrible skin! Really, everyone thinks of her as being so much more beautiful than she really is.Buh but.. I prefer the first image..
- Dec 31, 2016
- 316
- 708
Some scenes from The Mask movie might be to blame for some of that...
Given that the movie is from 30 years ago there's a slight chance she's aged as well
- Dec 31, 2016
- 316
- 708
Was the TI and lycoris both trained on the same model and that model was used for generating the images?
It says to use the lora/lycoris without the trigger word, have you seen or done any tests if there's a difference between not including a trigger in the training as well?
- Sep 15, 2021
- 1,042
- 3,984
The Lycoris and TI were trained on different data sets by different users (the TI's mine*, the Lycoris belongs to the author). The model used for the TI was SD1.5, I believe the same is true of the Lycoris, but can't confirm 100%.
Using the trigger word for the Lycoris doesn't ruin the results too much, but does 'overdo it' significantly. She looks a lot more wrinkly - 'over-detailed' if you will.
The alternative would be to use the trigger word but turn down the Lycoris weight and/or de-emphasise the trigger word.
The suggestion of not including a trigger in the Lycoris/loRA training at all would be an interesting experiment. I don't really have the time to try that though as it would mean training my own LoRA on Ms Diaz.
*and rather overbaked if I'm honest. Nowhere near the author's recommendations and certainly not very flexible.
- Dec 21, 2020
- 1,570
- 3,766
Fun with the IP Adapter - converting movie scenes / game screenshots into their weird replicas.
I have read the glossary but my head is bad so here are my dumb questions say I have about 100-1000 images of a model(woman) should I train lora or embedding with that? And If my input pictures of women has similar feature, eg big tits, small hips, can lora, or embedding understand the general idea/concept I want to get to, I want to create realistic image?
so here are my dumb questions say I have about 100-1000 images of a model(woman) should I train lora or embedding with that? And If my input pictures of women has similar feature, eg big tits, small hips, can lora, or embedding understand the general idea/concept I want to get to, I want to create realistic image?
Last edited:
IMG2IMG: Does it get any better than this?
TLDR: Is creating consistent cartoonized characters (same EXACT art style, same EXACT clothing, hair, facial shape, etc) using img2img possible in a NSFW-supporting version of stable diffusion when using a bunch of photos of the same model? And AS IMPORTANTLY: is it possible to do successful img2img photo-to-cartoon 2+ model hetero porn scenes that don't turn into a freaking living nightmare? Or is the technology just not there yet for AI builds that still allow NSFW porn?
NOTE after TLDR: I did have a much better looking img2img conversion of this without any sort of "missionary" prompt, but the guy was basically eliminated from the scene entirely by the program. His penis... became a sock on the bed smack dab in between her spread legs.
Ok. I know nothing about AI art practically. But have been casually going to a few sites to see what the capabilities are of turning 2+ model hetero porn photos into cartoons via img2img features, to see if it's even worth investing my time downloading and learning Stable Diffusion right now, based on the technology available.
I figured if one of these sites that charges is able to do img2img conversions of a 2+ person sex scene that are consistent and good looking, then it's worth my time right now to learn it better myself off-site, to get exactly the styles and looks I want. Otherwise... maybe better to wait. As you can see from above, it's not looking good. lolol. (This girl was not originally a tran-- or whatever that is. And that... arm? Was not originally a penis. And I did not ask for this. I DID NOT ASK FOR THIS!!
I've seen somewhat good results in single-person scenes converting photos to an img2img cartoon. But before I delve into this, I'm wondering if anyone's had any luck doing so with multi-person scenes. I know there will be a learning curve, of course. But I highly suspect the technology won't be there for another couple of years anyway, even if I move headlong into this, training models myself, etc.
WHAT I'M LOOKING TO DO: The reason why I'm looking for what capabilities img2img has is that I want the characters to be consistent from scene to scene, just converted to 2d art. Same exact hair and clothing as represented by the host image (no minor variations). Same body and face shape. Same art style. Which text to img prompts, as nice as they look sometimes, just can't do. I might be better off just tracing art and working on things in photoshop instead at least for the next 5 years until there's a leap in tech though. I don't know.
Here was one site that seemed to have its stable diffusion down well with a very nice art style. Giving me SOME hope that maybe the technology is there now. Unfortunately, it's a non-porn site that just just works on faces, cropping off everything else. And I assume it's using a new version of stable diffusion that came after that damn porn lock, where-as we're stuck using earlier versions that allow it. Could obviously use a bit of touch-up (her lips, nose, ect), plus the penis isn't exactly cartoonized. But it makes me think there's some hope possibly?
You don't have permission to view the spoiler content.
Log in or register now.
TLDR: Is creating consistent cartoonized characters (same EXACT art style, same EXACT clothing, hair, facial shape, etc) using img2img possible in a NSFW-supporting version of stable diffusion when using a bunch of photos of the same model? And AS IMPORTANTLY: is it possible to do successful img2img photo-to-cartoon 2+ model hetero porn scenes that don't turn into a freaking living nightmare? Or is the technology just not there yet for AI builds that still allow NSFW porn?
NOTE after TLDR: I did have a much better looking img2img conversion of this without any sort of "missionary" prompt, but the guy was basically eliminated from the scene entirely by the program. His penis... became a sock on the bed smack dab in between her spread legs.
Ok. I know nothing about AI art practically. But have been casually going to a few sites to see what the capabilities are of turning 2+ model hetero porn photos into cartoons via img2img features, to see if it's even worth investing my time downloading and learning Stable Diffusion right now, based on the technology available.
I figured if one of these sites that charges is able to do img2img conversions of a 2+ person sex scene that are consistent and good looking, then it's worth my time right now to learn it better myself off-site, to get exactly the styles and looks I want. Otherwise... maybe better to wait. As you can see from above, it's not looking good. lolol. (This girl was not originally a tran-- or whatever that is. And that... arm? Was not originally a penis. And I did not ask for this. I DID NOT ASK FOR THIS!!
I've seen somewhat good results in single-person scenes converting photos to an img2img cartoon. But before I delve into this, I'm wondering if anyone's had any luck doing so with multi-person scenes. I know there will be a learning curve, of course. But I highly suspect the technology won't be there for another couple of years anyway, even if I move headlong into this, training models myself, etc.
WHAT I'M LOOKING TO DO: The reason why I'm looking for what capabilities img2img has is that I want the characters to be consistent from scene to scene, just converted to 2d art. Same exact hair and clothing as represented by the host image (no minor variations). Same body and face shape. Same art style. Which text to img prompts, as nice as they look sometimes, just can't do. I might be better off just tracing art and working on things in photoshop instead at least for the next 5 years until there's a leap in tech though. I don't know.
Here was one site that seemed to have its stable diffusion down well with a very nice art style. Giving me SOME hope that maybe the technology is there now. Unfortunately, it's a non-porn site that just just works on faces, cropping off everything else. And I assume it's using a new version of stable diffusion that came after that damn porn lock, where-as we're stuck using earlier versions that allow it. Could obviously use a bit of touch-up (her lips, nose, ect), plus the penis isn't exactly cartoonized. But it makes me think there's some hope possibly?
Last edited:
- Sep 15, 2021
- 1,042
- 3,984
OK.I have read the glossary but my head is bad
As with so many things in life, the answer is "It depends". Sorry!
My advice would be: If the model has a relatively ordinary, normal look, reduce the number of images down to the best quality 50 or so (bearing in mind you're going to have to trim them to 512x512) and train an Embedding (aka TI).
If her look is out of the ordinary or you want more than the face to be specific to her e.g. massive tits, bodily features, tattoos, weird hair etc, create a LoRA instead. Use about 30 images for that.
In either case, for a good likeness do NOT select to mirror the images otherwise the face will look too symmetrical. It's often touted in more basic tutorials as a "one neat trick to double the effective number of training images for free!!1!!". It's actually generally a shit idea, unless the subject is inherently perfectly symmetrical.
- Sep 15, 2021
- 1,042
- 3,984
- Dec 21, 2020
- 1,570
- 3,766
We kinda have two LORA training threads. One's listed in the very original topic that started this thread. The author takes you through making a LORA in 1-2-3 steps. Trust me, you do want to follow those just to wrap your head around some concepts.I have read the glossary but my head is bad
Then we have this community thread of training a LORA: https://f95zone.to/threads/loras-for-wildeers-lara-croft-development-thread.173873/
- Dec 21, 2020
- 1,570
- 3,766
Would anyone have a good SDXL ComfyUI setup?
The fucking thing is laughing at me.
The fucking thing is laughing at me.
- Jan 31, 2022
- 225
- 663
Here's a basic workflow I put together it has SDXL and a Refiner, I'm not sure I added the detailer correctly and I still have no idea how to add half of the other things like Controlnet etc.
To clarify things for those that don't use comfy much, as far as I can tell you need a slightly different workflow for SD1.5 compared to SDXL, this is the SDXL workflow.
Last edited:
- Dec 21, 2020
- 1,570
- 3,766
So, now you can do SDXL in real time:
Watch for just 10 seconds.
Called Turbo SDXL, yet being rolled out.
You must be registered to see the links
Watch for just 10 seconds.
Called Turbo SDXL, yet being rolled out.
- Dec 21, 2020
- 1,570
- 3,766
So, in ComfyUI I suddenly started get the nightmarish renders. Like brrr FML WTF bro.
Turns out one of the negative prompts connected as a positives into the sampler. So, yea, look at your negative prompts - that's what I was rendering for a short second. Brrr.
Here some therapy for me.
Turns out one of the negative prompts connected as a positives into the sampler. So, yea, look at your negative prompts - that's what I was rendering for a short second. Brrr.
Here some therapy for me.
Last edited:
- Dec 31, 2016
- 316
- 708
They've been doing "live" renderings of webcam input and LCM, don't have a clip at hand, just came across one on a LCM info page the other day.
There's a delay when you see it side by side but still a pretty good start.
Edit: Found the page, it's almost at the bottom of
You must be registered to see the links
Last edited:
- Jan 10, 2019
- 291
- 1,552
Just found out that there is a new thing called Deep Shrink (
There's a
Just generated in SDXL and it looks good:
Looks like this in SD, just make sure you enable the extension:
I did not change the settings that are enabled by default otherwise.
You must be registered to see the links
) also known as Kohya Hires Fix (A1111 - search for sd-webui-kohya-hiresfix in the extensions), which allows you to produce hires pictures without needing to use the HiRes Fix. Much faster. Loads of articles in reddit on it, here is
You must be registered to see the links
. It also prevents the double head or monstrous bodies generated when using large width/height (I still got weird ones, but far less and no double heads).There's a
You must be registered to see the links
by Nerdy Rodent which also covers this option. Just generated in SDXL and it looks good:
Looks like this in SD, just make sure you enable the extension:
I did not change the settings that are enabled by default otherwise.
- Dec 31, 2016
- 316
- 708
I don't know what this does so sorry if it's already mentioned in the linked stuff.
When using SDXL in comfyui you can use the target width/height (i think that's that they are generally called) in the sdxl clip text encode to deal with "copies". You shouldn't just blindly put this as a insane number as it can have odd effects, but if you are getting duplicates of ppl or things like 1,5 bodies stacked in very tall images, you can increase the relevant value to fix it.
Not sure if there's something similar in a1111 as i can't use SDXL there
- Mar 19, 2020
- 2,193
- 2,824
Hi guys! Total noob here when it comes to AI generated images, but I'm trying to get somehow started. I've read the OP and a lot of guides, also searched this thread for various topics and questions I had and already found some very useful guides as well as some tips and tricks. But I just can't scroll through 2.500+ posts to find all the information I need, so please bare with me if I'm going to ask a couple of stupid questions, which might have already been answered already and I just missed to find the right posts....
So, I downloaded and installed ComfyUI and kinda get it working on my main Windows PC: I've also downloaded some checkpoints, and was able to use those different checkpoints to get different results for the same prompts with the "default" ComfyUI workflow. Then I followed some guide about SDXL and I think I got it working as well. But I still have some questions, and if anybody could help me with that (or just link me to the some posts which already have covered those questions), I'll be more than grateful!
1) Using multiple PCs over network: when it comes to GPUs, my main PC has a 2080Ti, and I have a 2nd PC with a 1080. Is there a way to include the 2nd PC's GPU when using ComfyUI, is it worth the hassle to set it up and how would I do that?
2) Using the "default" SDXL setup/workflow -- is there a way to include other models/checkpoints/Loras I've downloaded from Civiatai, and how would I do that?
3) Are there any "must have" ComfyUI custom nodes I need to add to be able to setup a decent workflow?
4) What are your personal "must have" negative prompts to avoid those horrible deformed bodies with missing/additional limbs? Are these prompts checkpoint specific or do you have some "default" prompts you always use no matter what checkpoints you use?
5) I've seen a lot of "grid" pictures from you guys, and I was wondering how to create those with ComfyUI to then to select a couple of pictures from that grid to upscale/improve only selected ones? What's your workflow on how to accomplish this, or is this something that's not possible with ComfyUI and I just misunderstood how "grids" works?
6) I've read about being able to include all your ComfyUI setting into a saved PNG, but so far I couldn't figure out how to do it. Is there any guide on how to write that information into a PNG (or into a separate file that corresponds to a specific PNG) so I can go back to a specific PNG I've save and just mess with it?
Sorry again if those are some very basic questions, so please feel free to ignore this post (or me in general), but any reply is with useful links or tips and tricks is highly appreciated! Thanks a lot in advance!
So, I downloaded and installed ComfyUI and kinda get it working on my main Windows PC: I've also downloaded some checkpoints, and was able to use those different checkpoints to get different results for the same prompts with the "default" ComfyUI workflow. Then I followed some guide about SDXL and I think I got it working as well. But I still have some questions, and if anybody could help me with that (or just link me to the some posts which already have covered those questions), I'll be more than grateful!
1) Using multiple PCs over network: when it comes to GPUs, my main PC has a 2080Ti, and I have a 2nd PC with a 1080. Is there a way to include the 2nd PC's GPU when using ComfyUI, is it worth the hassle to set it up and how would I do that?
2) Using the "default" SDXL setup/workflow -- is there a way to include other models/checkpoints/Loras I've downloaded from Civiatai, and how would I do that?
3) Are there any "must have" ComfyUI custom nodes I need to add to be able to setup a decent workflow?
4) What are your personal "must have" negative prompts to avoid those horrible deformed bodies with missing/additional limbs? Are these prompts checkpoint specific or do you have some "default" prompts you always use no matter what checkpoints you use?
5) I've seen a lot of "grid" pictures from you guys, and I was wondering how to create those with ComfyUI to then to select a couple of pictures from that grid to upscale/improve only selected ones? What's your workflow on how to accomplish this, or is this something that's not possible with ComfyUI and I just misunderstood how "grids" works?
6) I've read about being able to include all your ComfyUI setting into a saved PNG, but so far I couldn't figure out how to do it. Is there any guide on how to write that information into a PNG (or into a separate file that corresponds to a specific PNG) so I can go back to a specific PNG I've save and just mess with it?
Sorry again if those are some very basic questions, so please feel free to ignore this post (or me in general), but any reply is with useful links or tips and tricks is highly appreciated! Thanks a lot in advance!
- Sep 15, 2021
- 1,042
- 3,984
Not being a ComfyUI user (or SDXL - can't get the bastard to work!) I can really only answer #4
I really like to use negative embeddings:
For photorealistic I pretty much always use
You must be registered to see the links
&
You must be registered to see the links
(the latter from our very own devilkkw).The former also works pretty well in cartoon/anime checkpoints. I believe devilkkw has also done an extreme negative for these, but haven't tried it yet.
For hands, I use
You must be registered to see the links
for cartoons and
You must be registered to see the links
for photorealistic.