Looks cool, although i personally stopped at the login prompt.
I would probably also tone down the continuous generic pop-up that someone generated something. It seems a bit too frequent, repeating and generic to be trustworthy.
That being said, I like the use-case and it seems like nice work in a short amount of time. I have a couple of questions that I am curious about
1. Can you elaborate a bit more on your fine-tuning process? Did you "just" feed the model a bunch of regular emojis? Have you considered using any RLHF/DPO approaches?
2. You mention you generate a vector emoji. As far as I know the flux model just generates bitmaps, how do you handle that conversion?
Thank so much for your feedbacks. I think I may need to open up 1 generation for people without logging in. I toned down the pop-up completely now.
1. yes, right now it is bunch of images with data. I will be looking into RLHF/DPO when I have more time and finance, since it is costly.
2. I tried to train it based on a lot of vector images, and also do some optimization on my back-end too.
Please reach out once/if you start looking into further refinement (my email is in my profile). I am with a startup specializing in collecting preference data quickly and efficiently - so efficient that we even offer a free demo (https://app.rapidata.ai/compare). Feel free to hit me up and I will make sure you get a good deal.
That being said, I like the use-case and it seems like nice work in a short amount of time. I have a couple of questions that I am curious about
1. Can you elaborate a bit more on your fine-tuning process? Did you "just" feed the model a bunch of regular emojis? Have you considered using any RLHF/DPO approaches?
2. You mention you generate a vector emoji. As far as I know the flux model just generates bitmaps, how do you handle that conversion?