Gloria Guida, actress.
If you like my embeddings, please consider to support my work in
Description
FAQ
Comments (16)
Assuming you mean the Italian actress? 90% of her credits were in the 70s, if you want to change that tag... 😋
Oh yes that true, thanks.
I think I've only watched one of her movies to be honest, but I really like her face. It has something special.
@Punkimaster She was indeed a stunner back then! 😍 I was shocked to learn she's currently 70, but I guess it happens to us all eventually! 😂
The problem I see is the face dont resemble to Gloria Guida at all. I know well this actress, so I would say: face is pretty but the resemblance is just 20~30%. Hoping in a v2.0 version, soon or later...
I agree with you. I really appreciate the creator's work especially in doing these cult film actresses but it's really only a slight resemblance. I think it's the TI format. I've never had much luck with them in general getting a good resemblance. Wish someone would tackle these actresses as Lora's.
@Snipes Get his "edwige fenech", it's' a bomb and works well with a lot of checkpoints....
Yeah! I'm guessing why the other are not so good nor so stable. Probably is a matter of what images you can have as dataset...
Oh, it could be the dataset? I haven't found too many pictures with enough quality. I don't know why is so hard to find. The actress is well known. I think I made this embedding with 13 pictures (I normally use a lot more). I'll try to train it some more and we'll see.
@settima_ai Yea I have all of them. I tried Edwige, results are okay, I did try it with the model you used and if I do it at 1:2 I get a closer result. Still I never have a lot of success with TI's, Lora's just seem to work better with the models I use. I created my own checkpoints for Edwige and Bouchet and they work pretty good but I'm limited with what I can do with them. If I had the time, and the computer power I'd try to do my own Lora's but I don't. ;(
@Snipes How many VRAM do you need to make a LORA? I've read it's pretty consuming,
@Punkimaster I don't know. I've heard you can do it all online and not even use your own V-ram, but I haven't had time to look into it, here's the guide I bookmarked: https://civitai.com/models/22530/guide-make-your-own-loras-easy-and-free. All I know is I'm on a Windows 7 machine and that pretty much limits a lot that I can do because I can't run newer versions of Python. Someday I'll switch to Linux (I don't think I can stomach Windows 10). I did try inpainting a face with one of your TI's and got a little better result.
@Punkimaster Depends on settings. 1 batch with x amount of epochs and x amount of steps will place you at 5 gigabytes (and a bit), 2 batch close to 8 gigabytes. And then you got bunch of other options then can reduce or increase the VRAM.
Do not bother with Dreambooth if you decide to try.
Use this https://github.com/bmaltais/kohya_ss.
Watch his video tutorials from his youtube channel.
Latest version of bmaltais gui has "presets" folder with few examples of training configurations to test.
LoRA = ok overall for everything, LoHA much better for characters.
@frfromg Must say, i'm pretty confused about "TI vs LORA" fact. Everibody say Lora are better, but in my esperience I found a lot of unusable LORA and, au contrair, valid TI. Probably Lora is better but only if done in the right way. For example the 'herbalist' lora works ONLY with the checkpoint used during the training, while the 'edwige fenech' TI works well everywhere!
In other words, is still better having a good TI, than a bad Lora!
Anyway, for sure, two or more LORA dont work well togheter... so, I would like to understand if having a character as 'TI' may be usefull to reduce problems when you need to add a Lora in your prompt.
@frfromg How long it takes to make a LORA (3000 steps) vs embedding? A very simple embedding it takes me 90 minutes at the very least (3000 steps) and with all the optimizations you can use in SD applied. And that counting you don't have to repeat it from scratch. It's pretty unsustainable in the long run (electricity is not exactly cheap here). If LORA is really a lot quicker, I'll give it a try. That would fix a lot of problems and concerns.
@settima_ai Unfortunately I have no examples of my own (yet) to post, but you are partially correct.
Example: TI is mostly trained on a base model and it works well with all models, many LoRAs are trained on whatever model LoRA author picked and can give (sometimes) unpredictable results. Another thing I have found is base torch models used can have impact on end result as well.
Torch 1 vs new Torch 2 : torch version 2 trains better than torch 1, but torch version 2 embedding (textual inversions that is) use on torch version 1 software does give weird and ugly output.
So what you say that well trained TI is better is true. LoRAs can turn out bad (really bad).
His Edwige is example of very well trained because it works for all distances, close, medium and far.
@Punkimaster
How long does it take for 3000? I can't be precise but it takes less time than TI - at least via that github link program I left. And it learns faster.
Dreambooth broke my Automatic so I deleted it and started using those Kohya scripts.
That developer has his own youtube channel - it is better to take a look at his two videos he linked on his github page and then take a look at few other ways he uses that kohya gui.
Anyways it is just an option, I did try that Edwige Fenech embedding you made, it is top quality work.
To see one example of well trained (not prefect but close) LORA check out https://civitai.com/models/84137/enji-night-cosplay
Works well when I tried it.
I'll try to make a new embedding of Guida with the things I've learned these last months. Be at ease.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.