Desert in an hourglass (Daily challenge, 2nd place!)

16 of them 😎

QWEEEEN

Thanks for the tip

Thank you!

Thank you so much! 🙏🏼

Thank you!!!

Sure, what do you want to know?

Well: 1) How complex/lengthy are your text prompts? 2) Do you use default modifiers or have you designed your own? 3) Do you use starting images? If so, how do you decide where to set the noise level on the image? 4) How do you decide how heavily to weight the text prompt? I really admire what you've been able to create but don't feel like I understand the generation process well enough to emulate it. If you could answer some or all of these questions, that would be mint!

Most of what I do varies by creation, and especially by algorithm but I'll try to give an overview.

1. This one in particular used 16 filled prompt boxes. Four of which were negative prompts. That is more than what's needed typically, and only possible with the Artistic and Coherent algorithms (for now).

For Stable and DallE2 the limit is 60 words and only a single box so I need to be a bit more selective there, but I usually use all or almost all of the available word count. You can make stunningly artistic or realistic things with a few words, or no words if you use emoji, but they feel less like *I* made them when I do that.

A tip on word count is that some characters count "," and some don't ";" but almost every single character of punctuation has a unique effect on the outcome of a prompt.

2. I helped design some of the default modifiers. They're all generally designed by a mix of the volunteer moderators, site staff, and site creator.

Epic and the neo-impressionist paint ones are two that I did.

Beyond that pretty much everything I make is something I've made myself. I've been fortunate in having the ability to test thousands of modifiers, and the whole fun part for me is putting the word together to create something new. Once I've looked through the outcome of testing multiple things I try to get them organised into collections and display them on my blog site tdraw.art ↗️ - here's a direct link to my latest post showing examples of a few things: https://tdrawaiart.wordpress.com/2022/11/11/art-movements-in-stable-diffusion/ ↗️

When writing my prompts I'll get an idea, test it, observe what the AI understands, then test more or refine it - often looking to reduce words down to the most impactful ones I can (Epic preset) for use with stable and DallE2.

3a. I do! Sometimes. There's no set way I do things. For me a good prompt is one that can work without a start image, but can also be enhanced by one.

I've likely used some of every media possible as a start image by now. A couple of my favourites are things I've taken photos of - that look like what I want, or are more abstract and are unlikely to resemble the outcome in any meaningful way - but also hand drawn pencil sketches, digital sketches on my phone, photoshopping together a collage, random noise, or miscellaneous screenshots just to see what happens.

As for noise weights it comes down to what the start image is, what the prompt is for, and how much these two things look alike. Most of the time I'll create a sketch to use as a start image as a way of blocking out the composition for a picture, generally guiding it to where I hope it'll place the right elements. On those I'll use a low weight, often 0 with stable, so that it can doesn't add too much.

3b. It's important to note than when the sliders say from 0% to 100%, that doesn't mean that 0% won't add any noise, it will instead add the *minimum amount* that it can.

Even when doing a more abstract piece it's rare that I'll ever max out the noise as you start to lose some of the colour from the starter and well as the geometry.

A while ago I put together a big example of the different in noise, start image, and overall prompt weights have in a coherent creation. You can view that here: https://tdrawaiart.wordpress.com/2022/07/05/weight/ ↗️ It won't be exactly the same but it should help give an idea of how much different values can effect the image.

4. At this point I kind of choose weights by feel, but I've made too many things.

I'm going to assume that the overall prompt weight works similarly in Coherent and in Stable, but I'm happy to be corrected if this only applies to one of them:

When you pick a low overall weight, you give the AI more space to fill in the blanks. I will often use lower weights (30-40%) for shorter prompts or realistic faces where I'm relying less on my description and more on the AI dataset.

Conversely, a higher weight will be like telling the AI to be more pedantic and make only what it is told. I'll use a higher weight (60-70%) for things where I've been very specific or have a lot of experience with, or when creating more vibrant contrasting pieces.

Mate, I'm going to methodically try out each of the different strategies you've described, and I'm sure my art will get a lot better. Thank you so much!

No problem! If you want to know more any time feel free to ask :)

What a great read. Thank you for sharing!

thx for tip! ^-^