Yeah of course, any specific questions you have?
For now I can say that the basic workflow was:
1. Collect dataset and train an image AI (ru-DallE). Generate Images.
2. Use google lens to find names of interesting things in the image to possibly use in the description.
3. Use GPT-J on a cloud TPU to generate the descriptions with results from step 2.
4. Final filtering of images/text, then make this app. The images and descriptions are hosted for free on cloudflare.
I'm guessing #2 is the implementation of Looking Glass? Or is that used elsewhere?
And I suppose, for the most part, the workflow the main thing I wanted to know about. How much did you train the AI prior to building the app? And was there a reason to use ruDallE over just DallE?
As well, did you have much experience prior to this with AI? With how quickly AI had developed in the last few years, it's been a bit overwhelming to try and find resources to understand everything.
Yep, I used looking glass for that.
I chose ruDallE because at the time at least it was the only available version of DallE that was open source and I could train myself. Not sure if there are other versions of DallE right now that allow custom fine-tuning. If I remember correctly I did 20,000 iterations, not sure how many epochs that was. I think in total the training time was around 8-9 hours.
Basically no experience with AI before this, it's indeed pretty difficult to start out, and I think right now I could already do a lot better with DallE2 and ChatGPT existing. Unfortunately I can't really point you to good resources to learn, I just looked for different info all over the place.
Hope this answers some questions though!
I know you mentioned about can't point to good resources, 3 months later. Is there anything you can share to get started doing something like. SDXL is out and I believe on civit ai there are a lot of LoRAs thatn ca be used to train stuff like yours. I wanted to know what did you use as far as hardware GPU to train it.
I don't really have any experience with the more modern tools like stable diffusion, so I still can't really give good advice on that. My personal GPU is pretty old so I used sagemaker studio lab which gave me a pretty good free GPU to use (I think it was a tesla v100). Google Colab (pro) should be a very good option right now as well.