Finally, I managed to significantly improve image generation quality. 🎉

After countless iterations and days/weeks I finally managed to significantly improve my models. Models were good for some but quality was not nearly as good as I would want. Did countless iterations of Dreambooth training settings and managed to increase resemblance significantly.

Still some finetuning left to do but will be able to start providing significantly better images & pivot towards photorealistic CV/curriculum pictures soon!

On top of that, the first customers that have used the new model were very happy with their results 😄

✅ My gameplan next:

1. Image optimization & compression on the backend + filtering on frontend

   • Currently still stored as PNG and big size - slow to load + consumes too much bandwidth

   • Duplicate images check

   • Corrupted images check

   • Filter and check for bad images on client-side

   • Add big modal explaining bad input = bad result. Still too many ppl send very bad images.

2. Write and add 20-40 high quality prompts on top of the current ones to generate a variety of CV images (80 to 160 images)

   • Add styles to landing page & highlight more

   • Add styles to Uploading section + add select all + add possibility to select type (profile pics vs. avatars/anime/ ...)

3. Start email marketing

   • 7k emails from current users who have never been contacted. Time to change that

   • Goal would be to start a bi-weekly newsletter with info about AI models + improvements on ImagenMIA + discounts + refer to InterioresIA too

   • Need to have at least mails for the first 4 editions before I start the campaign - to allow for a good buffer

4. Start running ads

   • Purchased a course and will start consuming it & sharing notes

5. Programmatic SEO per image & background (similar to what I did for InterioresIA.com)

   • Step up the SEO game even more

   • Improve copy

Spent the last two days trying to improve the models 😅

And man, this shit is hard. Been doing a lot of iterations on training settings and playing around with parameters. I have managed to make some slight improvements, but far from being where I want my models to be.

Also spent a lot of time prompting to generate new and good prompts for professional photoshoots/CV pictures.

That's much harder than avatar profile pictures. With avatar profile pictures you have some room for error since it's an avatar/cartoon/anime version of the user. Here you don't!

Will keep at it and let's see where we get! This is important to be able to pivot into more CV type pictures, which is what visitors of the site obviously want!

Will also be good for me as I will be able to start charging more. Currently at 5-6$ per sale does not leave me with a good margin.

Other to-do's:

• Add image filtering on client side

• Failed models error handling

• Start mailing

The changes I deployed last week made conversion worse, not better.

In a rush before my vacation last week, I redesigned the entire website and I deployed the updates with the end goal of boosting conversion rates.

But the opposite occurred. Not a single sale during the entire week. I wasn't too bothered about it as I thought it was just a random drop/seasonality effects or something similar. Until my sister came to visit and I asked her to test out my website.

That's when I realized there were 2 issues:

• It wasn't rendering properly on older browsers (OKLCH not supported)

• Google Auth was no longer working for new users This resulted in lower activity on the website, less sign-ups, and less sales.

I did the following to fix this:

• Re-implemented Google Auth integration with Supabase

• Added fallback CSS code to my TailwindCSS file to handle fallback scenarios for DaisyUI on older browsers
  
  

Result: woke up to 3 sales and currently have 5 sales in the last 24 hours! 🔥