stable diffusion for Krita

Like it or not, it is here: GitHub - Acly/krita-ai-diffusion: Streamlined interface for generating images with AI in Krita. Inpaint and outpaint with optional text prompt, no tweaking required.

It looks really cool, actually.

13 Likes

For me, the problem is not AI in itself. Problems start when generated images/photos are presented as taken from real life. Not really different from any other “fake news” (used to be known as “lies”): it’s possible to lie about a situation with an image without any manipulation, just change the context…

As Krita is aimed at artists, with drawing as the main tool, AI doesn’t really cause a problem as far as presenting the image is concerned. How they obtain what they use to train the AI is a different matter…

2 Likes

…funny how lies are now known as “factual inaccuracies”, huh? Oh well, that’s another soapbox I’ll stay off…

2 Likes

I’ve been testing with it and it’s great as a reference for some photographic ideas.

But only as a reference, there is nothing better than taking the photograph on site and with the conditions you have at hand. It’s the best way to learn and improve yourself.

When I tried telling my boss that I “over-reached my anticipated potential physical capacity” when running for the bus this morning, he STILL told me off for being late. And yet this guy…

https://www.msn.com/en-gb/entertainment/news/watch-spacex-starship-megarocket-explode-in-rapid-unscheduled-disassembly/ar-AA1keVNM

2 Likes

nice, obviously pre-vis is an awesome application for it. I want to test it’s potential to create storyboards and matte paintings for film making applications.

seems as though I’m going to have to acquire a CUDA capable graphics card first however :grimacing:

1 Like

RTX4090 :grin:

I’ve two nvidia cards:
ASUS POSEIDON GTX1080Ti & Gigabyte GTX1070

You also need a motherboard and power supply capable to handle it. :slight_smile:

Make sure it has LOTS of vram.

I’ve tried it as well. My adsl connection was running at half speed for some unknown reason. Download started yesterday at 14h and finished at 22h, then I switched off the pc.

This morning I gave it a try and asked to create a laughing cat on a canvas of 400x640 px. That took 25 minutes (!!) on a 10 years old 4-core pc with 8GB ram, without speedy video cards. Funny result though…

krita_laughing_cat

May I ask how long it took to generate that photo, or “photo”?

Using the two nvidia cards

ASUS POSEIDON GTX1080Ti
Gigabyte GTX1070

1000 x 1000 px

Peek 2023-11-23 11-06

Thanks. So very usable for you, not for me! :wink:

1 Like

You can connect to a remote server. Maybe it improve the speed, I don´t know because I’ve almost zero connection to InterNot here so I can´t probe that option.

A little over 2 minutes for a 1024*768 on a gtx 1650 here on an old system
nvtop clearly show that at this resolution 4Gb of VRAM is enough at this resolution.

But I can’t quite get it to generate what I want, it seems a little bit “off” or maybe being used to midjourney, I input queries that are too complex …

Still I will experiment a little bit more :smiley:

512px images on a gpu with 6gb vram will take only seconds. (But longer if you use control net). Use sd 1.5 for that. This would probably cover most of your needs for inpainting/outpainting. If you want to generate a sky for sky replacement on say a 6000px image then just upscale after.

1024px images will need MUCH more vram if you want fast results. Sdxl is designed for that.

Crikey. Seeing as I only have a 2011 iMac at my disposal (with Arch Linux running via Virtualbox), I’ll sadly have to give this a miss (or be prepared to wait a week to produce something similar). Another good excuse for me to invest in a new machine, though. :slight_smile:

A little bit off-topic but I wonder if the upscaling available in this plug-in could be use to de-noise a picture, for example :
Remove chroma noise by downsizing then upscaling a picture then using the result in chroma only blend ?
Or such a trick …
I’ll have a go at it maybe next week.

Yeah, I’ve seen that. Prices are very reasonable: around $0,25 per hour with a minimum investment of $5 or $10 (that are the credits you buy). Then you run your work on a (shared) machine with 10GB VRAM, 13GB RAM, 6 vCPUs and a RTX 3080. Hmm, I guess that will run faster than my 10-years old pc! :wink:

BUT Franklin: you need an internet connection for that. :upside_down_face:

Yes, in a way people have been editing photos for propaganda purposes for almost 100 years now, if not more. I’m sure we all saw the Stalin photo where a person is removed :slight_smile: Of course it makes everything a lot easier…