my PNGInfo send to txt2img img2im etc. quit working. it clicks but nothing shows up in prompts etc etc. no errors in console show either. so idea what causing it. im tired of copying and pasting. so i need this working
I skimmed over the training post linked in the lede at https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10306324..
One thing that's not clear to me is if it's viable to feed results of a batch of generated images back into a training program - and have those be the sole source of information. My main thinking on this is while random pictures if done well are nice, in order to have any application for games there needs to be a way to generate consistent characters - e.g. all generated images for a given set resemble this same girl you started with.
Does that make sense or has anyone done something like this? Are there tips like once you have a base face, you train it with 1000 images of that very same face (which have in turn be generated from a prior session?) And if this works, does that mean one needs to repeat the training process every time they have a new character, or is there a way to associate faces you've created with prompts?
One thing that's not clear to me is if it's viable to feed results of a batch of generated images back into a training program - and have those be the sole source of information. My main thinking on this is while random pictures if done well are nice, in order to have any application for games there needs to be a way to generate consistent characters - e.g. all generated images for a given set resemble this same girl you started with.
Does that make sense or has anyone done something like this? Are there tips like once you have a base face, you train it with 1000 images of that very same face (which have in turn be generated from a prior session?) And if this works, does that mean one needs to repeat the training process every time they have a new character, or is there a way to associate faces you've created with prompts?
- Dec 23, 2018
- 481
- 1,998
There is a cooperative drive to make a Lora from Wildeer's Lara Croft here which does something similar - https://f95zone.to/threads/loras-for-wildeers-lara-croft-development-thread.173873/
I also did one which worked enough for my own needs. What I did was use a pic of a render that I quite liked as a base and ran it through Reactor which is an extension for Automatic1111. It overlays the base face into any subsequent pics. So I made 20 pics using that base, and then used those 20 as a training set to make a Lora. Worked quite well
- Jan 24, 2020
- 1,401
- 3,793
I have had this also. It usually works again if you reload the image, meaning exit the image and load it again. This can happen when you try to load an image in png info while generating at the same time. If it still doesn't work then reload the ui, next is restarting the bat file. Update etc. And last is reboot your pc. I doubt you'll need to go that far though.
- Jan 24, 2020
- 1,401
- 3,793
The best approach imo is what Sharinel described. Lora's are nice when you need them but they are becoming more and more obsolete. With Reactor you can make a face model also, it is simply compiling the batch of images you feed it and save it as a safetensor file. Next time you only need to select this model in your list rather than navigate yourself to an image. There is also an alternative to reactor, faceswaplabs. With this extension you can also make a facemodel and in addition you get a preview of the face model you have created, and can exclude or add an image to get a better result or improved likeness.
Last edited:
- Mar 3, 2019
- 228
- 737
I would have to disagree with that statement. As someone who basically focus on creating images of real people all the available face replacement technique, like Reactor, are very limited to achieve good likeness because they don't affect head shape and not everyone has the similarly shaped heads, all they do is change eyes, noses, and mouths; furthermore, they won't do anything about body shape. You can get images that look a bit like someone with these techniques, but if you want accurate representations of specific people, there are no replacements for LoRAs or Dreambooth models yet.
- Dec 21, 2020
- 1,523
- 3,589
Without getting into the discussion, but rather only focusing on this snipper: "they won't do anything about body shape" I'd say if I had to nail the body shape, I'd go with ControlNet for Depth/Canny or even Tile. Then snap Reactor/whatnot and there's your [famous person] doing [consipracy theory].
I developed a distinct dislike of LORAs for they are fucking witch brews that completely fuck with one's mind because of how uncertain the training outcome is. I mean this in a supportive way, this is not a criticism of anyone, thank God we have these. Rather a statement that due to me being fucking inept at things, LORAs are a miss for me.
Now, for real I meant to ask if you or if anyone else tried InstaID
You must be registered to see the links
?I am procrastinating as usual cause FML. So, wanted to ask if folks go "yey!" or "fuck no" over it.
- Jan 24, 2020
- 1,401
- 3,793
I messed with it a bit but since it was at the time only for SDXL it didn't do much for me. My old card can't really handle SDXL. Insta id is just a different flavor of ip adapter. The basemodel has support for sd1.5 but at the moment I tried it the extension didn't. This might already be corrected as I have been living under a rock for a bit I got no clue.
- Jan 24, 2020
- 1,401
- 3,793
Yes it's true that faceswapping leaves out the body of course but with different techniques you can get that pretty close. The issue with trained models such as loras is that they are notoriously inconsistent and not at all straight forward to create good ones. It takes a lot of time and dedication to achieve a decent one and most people doesn't have the fortitude that it takes, so places such as civitai are filled with utter garbage loras. Unless you are completely anal about it you can get close enough with faceswap and then work on the body separately with controlnet or other methods. This is imo much more consistent and easier to get right than loras. Every person is of course free to choose their own path. This is only my own experience and opinion, though many would agree with me clearly not everyone.
- Mar 3, 2019
- 228
- 737
The problem with that technique is only being able to generate basically a copy of another picture in a different style. I want to make new images of people in different poses and situations, and not just recreate existing images of that person. That's why loras work better for me.
I absolutely agree, training good loras is hard and it's very rare to find good ones on Civitai. But that's why I train my own, or have them trained for me by a friend with a more powerful GPU who happens to find the process fun. I currently have about 80 loras of different cosplayers and instagram models of whom I make images of as my past time. Many (if not most even) of these girls have very distinct faces and head proportions that would be near impossible to recreate a good likeness of simply by replacing their face in a different picture, and I often make full body pictures of them, and I don't want their face in a generic woman's body, I want to see their body recognizable as well. Loras are still the best way to achieve that, without question; if there was a better way, I would be using it
- Jan 24, 2020
- 1,401
- 3,793
Don't get me wrong I love using loras too. They are just very inconsistent. I think it's all about the use case scenario though. If you are happy with the result you get, that's all that matters. One thing doesn't exclude the other though. You can use both, or neither.The problem with that technique is only being able to generate basically a copy of another picture in a different style. I want to make new images of people in different poses and situations, and not just recreate existing images of that person. That's why loras work better for me.
I absolutely agree, training good loras is hard and it's very rare to find good ones on Civitai. But that's why I train my own, or have them trained for me by a friend with a more powerful GPU who happens to find the process fun. I currently have about 80 loras of different cosplayers and instagram models of whom I make images of as my past time. Many (if not most even) of these girls have very distinct faces and head proportions that would be near impossible to recreate a good likeness of simply by replacing their face in a different picture, and I often make full body pictures of them, and I don't want their face in a generic woman's body, I want to see their body recognizable as well. Loras are still the best way to achieve that, without question; if there was a better way, I would be using it
It's very difficult to get perfect likeness with a lora in my experience no matter how good it is. With a faceswap you get closer though. There is a different option that has not been mentioned and that is outpainting. Essentially only re generating the face from an image and outpainting the rest. You can just mask the face and select inpaint not masked. Have you tried to use a good lora and then use a faceswap over it? This would make use of the bone structure from the lora and get the better likeness from the faceswap. Essentially using the best from both. You can use real images with openpose etc for the different scenarios.
There are always new things to try. I think the fact that SD is not perfect is one of the things that keeps it interesting. If it were easy there wouldn't be any "sport"..
- Mar 3, 2019
- 228
- 737
Yeah, I tried a few different methods, my current one actually involves making an initial image without the lora, but using IP Adapter for a basic likeness with SDXL models to get a first pass with better colors and composition (but lower facial fidelity), and then using img2img with ControlNet to recreate that visual with SD1.5 + Lora, to add the likeness back to the image. I don't usually like using faceswap methods after because I feel they tend to make facial expressions look too bland for my taste, though I can't say I have really tried ReActor yet. But at least in my personal experience, and for my needs I feel that loras give me higher fidelity than any other method I have tried while giving me full control over character pose and composition on the prompt, and I usually don't feel the need to add another step to increase that fidelity.
While I won't say that I never have issues getting a good resemblance when using loras, I feel that it's usually the result of a lora that didn't get trained with the best source images or just the random nature of Stable Diffusion, but I have some loras that to me are basically as good as it gets, so the inconsistency problems you describe, again in my experience are less due to an issue with limitations of the technology, and more a case of bad implementation (i.e. poorly trained loras).
For example, here are some generations using my custom models compared to photos of the real women trained in the loras:
So yeah, that's why I don't think loras are going anywhere. If you want to be able to create images from the prompt without being restricted to get a likeness from pre-existing photo, loras are a lot more flexible and lead to better quality than the alternatives... to me at least. But my quality bar for likeness is perhaps higher than the average user, so don't take my words as gospel.
- Jan 24, 2020
- 1,401
- 3,793
Very nice images.
Yes you might very well be right that it's poorly trained loras and/or the general nature of SD that is the reason for inconsistencies, probably both. . An alternative that I already mentioned in a different post is faceswaplabs. It has some interesting features. Like reactor you can create a facemodel from a batch of images to use instead of a single input image, but unlike reactor it allows you to use an input image also with the facemodel, you can blend the facemodel with the input image and the generated image with a slider. So you can fine tune and balance the likeness more.
Ip adapter does get you there partially but it's not as accurate as a faceswap at least in my attempts.
It was along time ago I trained a lora and your conviction of it being superior inspires me to perhaps give it a go again. Good talk.
What method do you all use for aging people? I'm having a hard time setting ages. Everyone comes out young looking? Sometimes I want to have middle-aged or older people. But no matter if I enter 30yo or 90yo they come out the exact same.
I've even tried added "slight wrinkles" and some other wording.
I've even tried added "slight wrinkles" and some other wording.
- Dec 21, 2020
- 1,523
- 3,589
What does your prompt look like? Would be best if you posted an image that contains your prompt. I am pretty sure you'd have reworked piece back in no time.
- Sep 24, 2017
- 13
- 50
AI models tend to generate younger-looking faces by default, making it tricky to create older characters.
To depict older ages effectively, incorporate detailed descriptions that focus on features associated with aging, rather than just stating an age number. Use terms that hint at life experiences and the passage of time.
try a prompt like 1 woman in her mid-60s, with long silver hair gracefully framing her wise, smiling face, eyes that carry the twinkle of youth yet reflect the depth of experience, with laugh lines that tell a story of a life fully lived.
Yes sorry I didn't attach
You don't have permission to view the spoiler content.
Log in or register now.
What a great shot! Yah I think I just have to be much more descriptive.
What if the LORA you are using is a younger model? Would it work in the prompts or to age them in img2img?
- Dec 21, 2020
- 1,523
- 3,589
Kinda put the age token first and "max" it out. If this is on the younger end of what you want, keep adding age and add things like "grandmother".
We kinda had a few posts about aging a few month back.
That's working much better thank you! Yah I have been trying to go through every message, sorry if things have been repeated. I'll try using the Search more extensively next time. Thank you all!