Hacker News Reader

Comment by vslira | Hacker News Reader

Is there any service offering user-friendly access to open source DL models? Paid is fine.

If you have over 4MB VRAM you can run it locally. I've been experimenting recently and find that even with 10MB VRAM I can only get 256x256 resolution images. I have a Dockerfile I can share that packages up the install process and removes censorship if anyone is interested. I find the censoring is extremely conservative.

Please do share. I did the same and have it but having trouble deploying it to a prod GPU server, still figuring it out.

Here you go: https://gist.github.com/gvbl/9231406c54e7c9fd37abdfa6c697fe5...

You can run it with a command like this (I'm on windows): docker run -it -v <model file path>:/stable-diffusion/models/ldm/stable-diffusion-v1/model.ckpt -v <outputs folder>:/stable-diffusion/outputs -v <inputs folder>:/stable-diffusion/inputs -v <cache folder>:/root/.cache --gpus all knightley python /stable-diffusion/scripts/txt2img.py --W 256 --H 256 --prompt "a horse wearing a top hat"

assuming you build the image and tag it "knightley"

Huh, I've been using the docker container by cmd2 and been doing 512x512 just fine with 8GB. Are you on Windows by any chance?

Yes, windows with a RTX 2070 Super. The logs say the app is trying allocate just a few hundred MB more than what I have. I'm reasonably happy with 256x256 for now, just messing around.

http://beta.dreamstudio.ai is the official paid Stable Diffusion webUI.

I liked it until I noticed that it costs me $10 a day. I'm also not sure if they support img2img. I have used huggingface and then replicate for it(NSFW filter is very annoying).

Unless you’re looking at something else, it’s not per day, it’s a credit system that amounts to about 1 cent per image. How many images do you plan to generate a day?

yes, it's per image but for some reason I easily rushed through the 1000 generations in one day.

Well you’ve gotta either run it locally or pay for it. The GPUs this stuff runs on are too expensive to offer unlimited use for free.

I'm not against paying but it should be somewhat raisonnable. to put things into perspective: Colab pro only costs about $10/month and you will probably be able to generate at the same speed.

For uses like that Colab costs Google more than they earn. It would make no sense to base a business around that model.

I did too, but I bought another 1000 generations and haven't run out yet. Many days I don't use it at all. I like it better than a monthly subscription like MidJourney.

Just use Google Colab, it's free indefinitely because Google is paying for it.

Midjourney uses Stable Diffusion now too in their beta version.

The stable diffusion repo itself is free, and you can get the model from hugging face, also for free. Runs on my 1070 easily.

Reply on news.ycombinator.com