Learn 1 thing there is no best

most of these AI faces don't look "natural", they look plastic surgery performed on them.

same! though even this inversion can't save majicmixRealistic from my testing. that otherwise great realistic model appears to only know asian people exists :)

@argh I tried out most checkpoints that claim "realistic" images, and photomerge seems to be better than most. use with negative prompt "asian"

'superior' is highly subjective here. By that you mean "the usual method everyone already knows."

First, this isn't meant to replace that method, it's just meant to be another one. It works, and it achieves different results as well, which makes it equally valuable. I've done testing using your "method" and this negative embedding, and even both together. All three are completely different.
There is no 'problem' with this method whatsoever.
There is no 'right' way of doing any of this.
It's all about taking the tools provided and bending them in different ways to make something new.

Second, this embedding helps a lot on models that are heavily trained on asian datasets. I implore you to continue your testing on models such as this:
https://civitai.com/models/25494/brabeautiful-realistic-asians-v5
You'll find that neither method works well on it's own, but combined they do a pretty great job.

Just pointing out that this resource's value doesn't diminish based on your personal tastes and opinions whatsoever, and claiming things are "superior" or "ideal" is putting what you prefer up on a pedestal and can deter others from testing things for themselves. Considering you have zero uploads and zero contributions to this community, I would think anyone would be wise to take your opinion with a grain of salt. Especially when commenting on an upload from someone like @Zovya who has far more established credibility and contributions.

@Zovya I used to use the negative asian tag, among others, but have come realize that while it works it is not optimal. Like you say, they don't tag asian, also asian can mean many things, not just a persons face, and so on.

Putting (Asian) in the negative just made them even more Asian for me.

@DannyCool I tend to use "Asian Ethnicity", since the internet is very western biased anything with "ethnicity" is generally coupled with non-caucasian people

My testing indicated that the combined use of both methods in specific amounts based on the weights of the models, the number of tokens in surrounding textual inversions, and the strength of certain LoRA types is, in fact, optimal.

I suggest you continue with a more rigorous experimental regime for even more superior results. In this there can be no preference - only what is optimal. I applaud that you tried. You should be proud of yourself. Continue to learn more a more scientific method, however.

For anyone wanting the optimal solution I suggest genetic testing be performed on your images.

/s

@Fenn You failed to read what I actually said and just want to argue so I'm going to keep it simple and if you continue after this post I will not bother wasting further time. When I stated superior I gave context as to why but also pointed out where this specific textual inversion can still have merit. I clarified quite clearly the textual inversion produces biased results which defaults it to being an inferior method which is an actual fact. It skews to specific types of people of given races and lacks the richer diversity, more accurate, and less biased results of the method I stated. Nothing you stated refutes or is even relevant to anything in my initial comment. Please don't waste time and space on pointless arguments, especially when someone's post was intended to be insightful, helpful, and not trolling or inflamatory.

@thelustriai This depends on the model. Most models you will not want to use both because it can produce inferior results and has a strong bias. For models that are very specifically tailored to Asian results then, as Fenn and you stated, you would want to use both because those models are inherently not capable of properly producing non-Asian results so the textual inversion provides more data to work with and is, essentially, the only option there. However, you are better off not using Asian specific models for such purpose to begin with.

This comment was helpful—I'm not sure why it's getting downvoted.

@shaozi88 Not too sure honestly, but thanks. If you are looking for a good model for this suggestion the best all-rounder I've found is probably Level4 so it may potentially be worth checking out if you haven't already found a better one to use with Positive / negative modifiers for intended race/nationality. https://civitai.com/models/17449/level4

Harvard needs a " Smaller Noses" TI

Yeah, setting anything to more than 1.5 seems to stretch the canvas of reality

most people don't understand what embeddings means. It literally mean to remove something from stable diffusion. Some people remove ugly face, some remove ugly hands. His embedding remove asian face. U train it what to remove pretty much, learn how to make embeddings and problem solved

You can convert it yourself

asian girls in most of AI models all look like they have crap ton of plastic surgery to maximum that any doctor would say "no! its impossibruuu!!"

Just putting the literal word "Asian" in your negative prompt can often help