I'm not switching but putting LORAs alongside LyCORIS so I guess best of both worlds? :)

Some people prefer the high details of faces and other people prefer the flexibility of LORAs. (and I guess you could mix and match)

May I ask you about your workflow? The LyCORIS, in general, are great at faces, and definitely shine in the close-up department, but when I do further back - when using high.res.fix I'm increasing the denoise factor the further away the face is. So for close-ups, I use denoise between 0.1 - 0.3 but if the face covers half the picture, I increase it to 0.4 - 0.5 and if the face is in like 1/3 - 1/4 of the picture - I go even higher, 0.6 - 0.7 (I try to go lower if possible since higher denoise also means higher changes in the whole composition).

With those denoise factors, the face improves from "barely recognizable/not recognizable" to "good enough/quite good".

If I want to make the face even better, I upscale it again in extras, send it to inpaint, and do the face correction (inpaint only masked, denoise around 0.5)

@civitaijugg also, this is a good place as any -> I'm looking at the LyCORIS popularity and I'm going from top to bottom with LORAs. But if you would like to see some specific person sooner (that might not be as popular as LyCORIS) - just let me know (this goes for anyone reading this :P)

@zzzzzzerza both of them are already done and waiting, will upload them either tomorrow or on the weekend :)

thanks for the support and good word! :)

@civitaijugg I see, well I'm planning on releasing a bit more of my models also in the LORA format as I see that there is still demand for it even though the LyCORIS for those subjects is present.

I'd appreciate if you could do that. Thanks a ton.

the samples have been uploaded, have fun and thanks! :)

That is an interesting question. I have seen great LyCORIS done via training so it is possible to achieve great results directly I think it's much easier to get a better quality Dreambooth (and therefore extract to LyCORIS) because there were several techniques (and/or repos lets say) and also because people were able to experiment with it for much longer.

Would you say that training a dreambooth then extracting the LoRA/LyCORIS is a better technique for training a model of a real person

I think game characters turn out exceptionally well with this technique so I wouldn't limit it to only real people.

Does the model's subject matter or would you say that training dreambooth + extraction is always the better option regardless of what kind of thing you're trying to teach the model?

I would say this: if you have a great Dreambooth then an extraction to LyCORIS/LORA should yield very good results. We've seen varying concepts made into Dreambooth. But then again, there are great LORAs and LyCORIS made via direct training too :)

I know this response might be a bit "diplomatic" but the point I'm trying to make is this: you can achieve great results when you focus on specific things and try various options, parameters, datasets, and so on. My first Dreambooths were really quite bad, but only through perseverance, trials, and errors (and a lot of reading and checking how others do it) I've managed to get to where I am today :)

There are two main reasons I stick with Dreambooths and extractions:

* I'm happy with the dreambooth results that I make

* you can get other stuff from dreambooth, and since those represent the best quality, if something else comes up (that might be better or have another purpose) - it is likely (but of course, not guaranteed), you would be able to extract it to that new format eventually.

I'm not sure if I answered your question :) If you can, try both and see which one you are better at :) (assuming you can do both, some people can only train LORA because it requires less VRAM)

@excalibu good luck, would be nice if you could share your findings later on :)