GeneratingPicturesForFunMar 27, 2024

Small QoL suggestion!

https://lemmy.world/post/13588595

Small QoL suggestion! - Lemmy.World

Hi, I’m using perchance AI image generator for a few days, and it’s excellent, something’s been bugging me. While AI is probably training itself in the process, extensively so, it doesn’t keep track of certain flags, say, color of the hat, or some shape is different. Is there a way to let AI know “Hey, I want certain features to be persistent from one generation to another?” Like an absolute statement, it can’t be anything else, it’s defined.

Show threadVioneTMar 27, 2024

The AI doesn’t train itself every time you generate an image. It is ‘pre-trained’ and when it generates, it essentially matches the ‘text’ that you input to its ‘vocabulary’ and match those ‘vocabulary’ to an image that it learned.

It would always give variations of the text that you have given. See this prompting guide for other techniques that might be useful for your generations.

Plus, the model used in Perchance isn’t as updated as it may seem. I would suggest you to look into creating your own setup so you can have other features.

Text to Image Prompt Guide ― Perchance Generator

Show threadGeneratingPicturesForFun

Okay, very interesting. You said it “learned” something, meaning it’s “possibly” retaining some of the information from previous generation.

Either I didn’t read it, or you didn’t answer the part about generation traits being set in stone, or is that related to “make your own setup”?

How difficult would it be to make my own, on a scale from 1 to 10 for someone who isn’t technically inclined, and definitely not a coder?

Thanks for such swift reply.

Mar 27, 2024 at 5:04amWeb
Show threadVioneTMar 27, 2024

By ‘learned’ I mean, the images that it was trained on i.e. the Dataset of images that the model is trained on.

It doesn’t actually ‘retain’ information from succeeding generations but it ‘tries’ to match the ‘word’ to an image that the model is trained to.

I’m not sure what ‘generation traits’ do you mean but there is an option to add an ‘Image Seed’ which essentially locks the ‘noise map’ that the image is generated from (although it doesn’t really lock it due to the sampler used in the generation). Here is a simple example of it.

Settings:
Prompt: beautiful clever draconic magical
Image Seed: 1
Resolution : 512x512
Guidance Scale: 7
Negative Prompt: bad

All of them have the same settings which means that it would have the same ‘noise map’ where the image is generated from. Due to the Sampler (Euler A) used (see this website for a visualization) it would still have slight variations even if the image seed is the same.

As for creating a local setup, and by what I mean ‘local setup’ is downloading your own AI model and using it in your own computer. It might need quite a powerful computer to be fast in generating but it has more features than the Perchance one can provide. See this video tutorial. 

Stable Diffusion Samplers: A Comprehensive Guide - Stable Diffusion Art

Many sampling methods are available in AUTOMATIC1111. Euler a, Heun, DDIM... What are samplers? How do they work? What is the difference between them? Which

Stable Diffusion Art
Show threadGeneratingPicturesForFunMar 27, 2024

Thanks for another detailed reply.

Don’t get me wrong, I’m happy with perchance, just pointed out something. Generation trait… basically just prompts, but persistent. As for PC, yeah, you’re right, must be powerful, that comes at the cost.