hunkins

Deployment jobs are queueing and not being picked up. Vercel status page indicates no issues.

Our core models (v1.5 and v2.1 Stable Diffusion) will remain free and unlimited, and have for six months and counting!

Ah, yes we use StabilityAI's NSFW detector. It uses CLIP, so it's pretty good, but far from perfect - false positives are pretty normal. A re-run tends to fix!

We use sensible limits to ensure people don't automate things, which is the primary concern. CloudFlare is great as well. But it's not cheap.

Things are starting to work, but my actions are stuck in the queue.

If this is true, they really need to do a code freeze and figure things out.

Trying to get a production build out and... it's down.

"It stores compressed copies of the training images and uses them to recombine the images into new "collage" artwork based on the existing artist-made image."

This is painful to read. So much misinformation on the tech.

If you're looking for an API, can personally recommend PhotoRoom: https://www.photoroom.com/api/. Used it in one of my projects, and had zero complaints. They have a free tier as well I believe.

Having built something that had to use the netRequest API due to new extensions being forced into MV3, I can attest to it being viciously limited and complex.

Ads will win if MV3 is mass adopted and enforced. Time to set-up Pi-hole.

Using an opinionated library can help with some of these issues. Personally, I love https://mantine.dev.

Sorry for the late reply: we're running via Banana, they quote half a cent per generation via Stable Diffusion. Rough estimate. Lots of optimization to do though.

They're very early in the space. Would recommend looping back with them on Monday in their Discord. Haven't seen that specific issues personally yet though.

These definitely help.. but still painful.

Early days, but as a user: this is so much easier than setting up a K8s cluster w/ GPU node pools and cheaper than dedicated GPUs. Check it out.

Note: launched a Stable Diffusion app at https://mage.space via this. Check out the speed.

Most deep learning these days requires CUDA acceleration to enable GPU / TPU for the libraries (i.e. PyTorch, Tensorflow), which is an absolute nightmare to set-up.

We're using GPU serverless (via banana.dev), so it's actually not bad. Will have limits at some point, for now go wild.

Specifically images, so admittedly yes. Generating the right prompt is a tricky science, but if what you're trying to generate contains elements found on the internet, you're probably in luck.

https://lexica.art is a great place for prompt inspiration.

You should be able to get a regular prompt to generate in 5 seconds via https://mage.space. A100s under the hood. Unlimited too. (Note: I am the creator)

Love it. If you don't have an M1 Mac, or don't want to wait, https://mage.space does unlimited generations currently. (Note: I am the creator!)

Just fixed a speed issue with the API. Pesky async snuck into the code. Should be much snappier now!

My best attempt: https://twitter.com/greghunkins/status/1571025487362953218?s...

Unfortunately, these models don't do great with counting.. (yet!) Might want to try a higher guidance scale if you need the image closer to your prompt.

Great feedback. Noticed this as well. Just pushed out a change where the notification shows a loading bar and remains until the search ends.

Twitter is actually @greghunkins. Tired.

Definitely will be adding img2img capabilities. Easy to turn a celebrity into any video game style, but will need img2img to do that for non-celebrities.

Should be able to generate icons via this tool now, albeit it is not focused on that.

Hi all, creator here! Should have put the "Show Hackernews:" entry on the post. Would love feedback and thoughts on next features! Will be posting progress on Twitter (@ghunkins) if interested in updates.

Creator here. Yes, hacked this together over the past two days. Community is going wild, and planning to integrate everything sensible.

Next will be: (1) img2img (image-to-image): combine prompt with a given image (i.e. turn your cat into a renaissance painting by Leonardo Da Vinci) (2) inpainting: use text to fill-in an area of a photo (3) outpainting: expand a photo beyond its bounds

Would love thoughts & feedback though.

This release changes society forever. Free and open access to generate a hyper-realistic image via just a text prompt is more powerful than I think we can imagine currently.

Art, media, politics, conspiracy theories; all of it changes with this.

Stable Diffusion, a direct competitor to DALLE-V2 and Midjourney in terms of image generation quality, is now released and available to the public. It is the first large-scale text-to-image model to be released to the public amid concerns of misuse and abuse.