Open AI Images on HackerNoon: The Conundrum

HackerNoon and the time the team compared different Open AI Image Generators for in-house use. This discussion paved the way for some awesome features like our headline generator and featured image generator, both already live for your enjoyment! Read and follow along our behind-the-scenes journey.

This Slogging thread by richard-kubina and David occurred in Hackernoon's official #random channel, and has been edited for readability.

DavidAug 30, 2022, 9:30 PM

richard-kubina, maybe time to revisit the headline generator? Ideally, the machine reads a story by contributor and suggests 3-5 headlines. gpt3 seemed really good for this (except for the cost)

richard-kubinaAug 30, 2022, 11:50 PM

It'd be neat to sort out a platform for open-source AI models... for features images, stable diffusion is a bit like Dall-E but open source, we could run it on cloud servers we run directly, which would be the cheapest way to access that functionality. I really dig its results too

DavidAug 31, 2022, 2:41 AM

Oh, interesting! Given the cost maybe a featured image is only after the first story published? Is this the image generation alternative you mean?

https://github.com/CompVis/stable-diffusion

richard-kubinaAug 22, 2022, 7:21 PM

OpenAI is reducing the cost of GPT-3 models

We are excited to share that we are reducing the price per token for our standard GPT-3 models and Embedding models.

Please find our new pricing below, effective beginning September 1, 2022 00:00:00 UTC. You can also find more details on our Pricing page. Note that this pricing change does not affect fine-tuned models.

MODEL BEFORE ON SEPT 1
Davinci $0.06 / 1k tokens $0.02 / 1k tokens
Curie $0.006 / 1k tokens $0.002 / 1k tokens
Babbage $0.0012 / 1k tokens $0.0005 / 1k tokens
Ada $0.0008 / 1k tokens $0.0004 / 1k tokens
Davinci embeddings $0.6 / 1k tokens $0.2 / 1k tokens
Curie embeddings $0.06 / 1k tokens $0.02 / 1k tokens
Babbage embeddings $0.012 / 1k tokens $0.005 / 1k tokens
Ada embeddings $0.008 / 1k tokens $0.004 / 1k tokens

We have been looking forward to this change for a long time. Our teams have made incredible progress in making our models more efficient to run, which has reduced the cost it takes to serve them.

Our hope is that even more developers will now be able to experiment and build on top of our API so that more people can access the benefits of powerful AI.

richard-kubinaAug 31, 2022, 12:17 PM

Yep!

DavidAug 31, 2022, 6:02 PM

richard-kubina is there a way to test it in the browser? just found the paper https://ommer-lab.com/research/latent-diffusion-models/

richard-kubinaAug 31, 2022, 6:31 PM

I believe there is a hugging face model to test with, just takes a while. Let me find that again.

I used to pay for google collab pro to try out the resource-intensive models out, that's pretty easy to test in the browser.

richard-kubinaAug 31, 2022, 6:32 PM

https://huggingface.co/spaces/stabilityai/stable-diffusion

richard-kubinaAug 31, 2022, 6:32 PM

https://beta.dreamstudio.ai/home

richard-kubinaAug 31, 2022, 6:33 PM

This application is too busy! Try again soon.

ah shucks, there's that issue too. It was working for me the other day
edit: working now.. anywho we could see about other providers like google collab, apparently, there are more that are lesser known, cheaper

richard-kubinaAug 31, 2022, 6:48 PM

already got into dream studio. A lil over a penny an image

richard-kubinaSep 1, 2022, 2:25 PM

I signed up for a https://www.paperspace.com/gradient trying out stable diffusion, it would be $8 a month to run on their GPUs, but I think other than that it would be 'free'. We could run it as much as we want.
Looking into their SDK/CLI may have to spin up a python API in google cloud to access this from the editor. If we like it and get a lot of users we could scale up by paying for more GPUs💰

they had a stable diffusion project template ready to go.. with the prompt "Kermit the Frog on the Iron Throne"

richard-kubinaSep 1, 2022, 2:28 PM

Dreamstudio could work too where we pay for the month and don't have to worry about using up credits, but paperspace is more flexible, we can run more models than just stable diffusion, like if we wanna try other models like the pixel art one, or open source GPT-3 alternatives for text generation/summarizing/categorizing/etc models

richard-kubinaSep 1, 2022, 2:44 PM

oo node.js access exists :thumbsup: https://www.npmjs.com/package/paperspace-node#programmatic-access-via-paperspace-node

DavidSep 1, 2022, 3:20 PM

AI images made John Oliver this week https://www.youtube.com/watch?v=3YNku5FKWjw&t=8s focuses on https://www.midjourney.com/ and https://www.craiyon.com/ in addition to dalle

richard-kubinaSep 1, 2022, 3:27 PM

Midjourney is 10 bucks for ~200 images. not sure what the commercial terms are

but I know stable diffusion doesn't mind if you use it commercially "for any purpose whatsoever, including without limitation commercial, advertising or promotional purposes" https://stability.ai/stablediffusion-terms-of-service

DavidSep 1, 2022, 3:32 PM

Craiyon's terms are way worse - they want 20% of NFT sales made with it, accreditation everywhere, and contact sales for use for companies like ours https://www.craiyon.com/terms

DavidSep 6, 2022, 10:15 PM

richard-kubina for ref, here's ux example of someone turning a wordpress text editor block into a dalle powered image gallery https://twitter.com/UglyRobotDev/status/1565030948428828675

DavidSep 6, 2022, 10:17 PM

lol sidenote discovery https://twitter.com/bakztfuture/status/1528173885517987840