Captioning

[irakli-ff]

Hello there!
Thanks a ton for this work - it is lot of hard work that you are sharing so openly.

I think I saw it somewhere but can't find it again:
what kind of captions did you use?

And any thoughts on experimenting with short captions vs long captions?
related would be caption shuffling and keeping n tokens (relevant for long captions only)

[InfluxOW]

Hello! That's where you could see it.
I don't recommend using captions for a training on a single person's photos. I made over 50 different LoRAs for over 50 different people with the described settings and the result was good without captions. So, I think that captioning is a waste of time in this case.
If you want to merge a several characters in a single LoRA or you want a style LoRA, then captions is a necessary thing as far as I know. But I haven't been delving into this so I can't tell you about captions in these cases.

[irakli-ff]

Thanks!

So what I'm trying to do it figure out how to train a model of a person but with only 1 image instead of 20 or 50.
And my thinking was that if I use detailed captions I could get a bit more flexibility from the model I get.

My understanding is that if you use short caption, model tries to fit everything that's in that picture in that small capture, including unrelated things like chair that the person is sitting on.
But if your caption includes "sitting on the chair", model already know what that looks like and it will not learn the chair part in the special token you are using. Essentiatially, in theory, this should be the way to exclude certain things being learnt. In practice - well, I'm just starting the experimentation.

[InfluxOW]

I saw some videos and articles on how to train a model with a single photo and it requires a different technique. Otherwise model won't be versatile enough and most like will just reproduce the image it has been trained on. I guess, adding captions won't be enough to achieve a good result.
What about captions, as I know, you should include everything unrelated to the person itself in the captions, i.e. scene, pose, clothing and etc.