Unstable single image dataset test, use at your own risks!
This is a test I was looking to do on training on a SINGLE image. In my quest to perfectly imperfect beautiful women, I did one image I liked a lot and I decided to try and train a single image lora with it. This is the result.
Description
Epoch 13, 650 steps on a single image.
FAQ
Comments (9)
Was the training data set truly comprised of only a single image, or did you "bootstrap" (synthetically augment) from a single image into a still-small dataset? I ask because I've been interested in the techniques around doing this, for those cases where a checkpoint gives you just what you want...one time...but alas never comes close again. Thanks for any info you may wish to provide as insight and thanks for this LoRA (she is wonderfully imperfect)!
Thanks! It is a trained on a single image with a long detailed caption. The training was directly on civitai generator, as you said, I was able to get that large teeth, narrow face and large mouth on a single image while playing with prompts, with the cost of flux this was a coslty experiment… I read some articles on this single image theory and I gave it a try. I was thinking I could get various angles of the same woman to do a better lora later on, but so far I struggle to do this, it seems hard to recreate the face features in another angle because the forward facing view is integrated in the lora, like the looking away, it thoses conditions are not respected, I start to lose face features! I did 1000 totals steps splited in 20 epochs, I think it is the #13 here. I think flexibility might be better at a lower epoch, right now I’m taking a pause on this because my buzz are running low😅.
@ElectronLibre So, one theory I had was that I could use an "Inflexible" 1-Image LoRA to produce some number of reasonably diverse images, even if they do lose resemblance a lot, select some that come closest, and then manually do an IMG2IMG Inpaint of the face (and hair I suppose) with IPAdapter Face v2 and the 1 "ideal" reference image, with it setup for "ControlNet is more important" and w/o a Prompt at all, and see whether that can lend a little more of the desired likeness back in. It seems like a fancier (and hopefully more capable) way of doing what Roop did with the InsightFace 128x128 model.
There's also something called PuLID for Flux, but I've only used it a little bit on HuggingFace and it seems like produces pretty much no variety in the face or expression, or very little anyway, so it doesn't seem useful to me for synthetically augmenting a 1-image training dataset.
Just some brainstorming I did...I haven't set aside any time to test this notion out myself yet though.
@XenoTech Seems like a good idea, unfortunately I don’t have the necessary computer for this, all I can do is use civitai generator, so I’ve been working on multiple lora trying to recreate what I’m looking for by mix and matching them. Using « oblong fave » and « large mouth », I believe I got what I was looking for.
OMG I love women with that kind of “ugliness” - those big teeth on a long, thin face 😍 I find them very sexy. Please keep going with this style!
Thanks! I’m currently training a lora for « large mouth and teeth » to create that exact kind of look, I should publish it tonight if it gives good results, and I’m quite committed at it, same for the whole « perfect imperfections » stuff I try to achieve with some of my lora. I don’t know why but I love that type of face and I’ve been surprised at how many people felt the same way.
I think I got the style dialed in https://civitai.com/posts/7367881
For some reason, she makes me think of Katherine Ryan...
I didn’t know her, I can see the similarities.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.