peteralaoui

We use it to get consistent results, otherwise only 30% of the generated images have the right composition.

For example, Astria has its own trick to increase the number of headshot in frame, if you download the checkpoint and run it locally you will see that it doesn't handle well negative prompts (which, IMHO, is usually a side effect observed on fine-tuned style models with dreambooth). With this trick you can get up to 70% of headshots in frame but won't reach 100% without risking a loss of prior knowledge. This is my experience so far, maybe you have another opinion based on your experience.

@xupybd, cool idea, it could maybe capture the essence of you while still injecting some perturbations into the final rendering to break facial recognition.

If you want to experiment thatwith your photos I'm available on DM: twitter.com/_dulacp

You don't relate to the message? I'm curious to know what is the first thing you would change

It's a cool idea to simplify the input as you said, though as @dorkwood pointed out this would lead to bad results with a lot of overfitting, this model needs a lot of variation to only register your facial features as "you", and disentangle that from the surroundings or even from your clothes.

(this is due to how this model learns, maybe a future model will alleviate that)

Hmm to be honest I just reused the most restrictive license of the tools that compose the app.

I guess I should rephrase to something like "The rights of the images are transferred to the user who purchased the service" or hire a lawyer to have it rephrased properly.

Seeing examples from happy users might help?

Or a refund policy could seal the deal? :)

Thanks, feedback well received! I do have a new pipeline that I'm experimenting with and it's much more realistic (hopefully realistic enough to end this emoji feeling).

If you are interested in trying, we can talk in DM: twitter.com/_dulacp

Thanks! Great idea. I will probably experiment something.

Allowing some flexibility in the types of costumes might be tricky though (today it's easy because I only allow for a few outfits for "business" looks).

What if the cracks were imperceptible in the next iteration or two? Would you still say it's a noob trap? _Genuinely interested in your feedback_

It's not the first time that I hear that, thanks for your radical candor :) and to be frank too, I was also sharing that feeling of a fantasy made to mimic reality when I started this project.

It all started when we started to argue with my family members on which avatars looked more like me. It made me realize that we were much more sensitive than I thought about our self-image. Me and my partner would pick different pictures in a set of 10 samples ^^ as if we had two slightly different perceptions of reality.

Now, I changed my mind slightly and tend to see these models as an another type of compression of information. Almost like a new censor of data.

Would you mind to share which mobile device you are using? Asking because it seems to work well on Android and iOS devices that I have at home.

Thanks for the feedback, I do have a new pipeline in the making that aims at improving the photorealism just tested it today with a few users (if you want to try, I can run it manually you, DM me on Twitter @_dulacp)

Out of curiosity, which part of the image makes you say that?

Could probably add GFPGan to my pipeline to remove these eyes artifacts. Thanks for the candid feedback :) I will double down on that

Fair point :) I feel at this stage only early adopters might try it, but with the next iteration or two I'm feeling this gap will be closed

Astria is a great service, but doesn't have ControlNet integrated to the best of my knowledge.

So, if you really want to reproduce this on your own without using a service like ours, I would advise to: i) train a model with Astria, ii) download the model checkpoint from Astria, iii) install Automatic1111 UI with the ControlNet extension and now you will have a local setup that is very close to what we have, the only missing piece is that we have a custom model optimize for faces.