Home Artists Posts Import Register

Content

Something new is brewing! Our new brew feature is the easiest way to create really great AI images on your first try, without any technical knowledge or long prompts, but with all the benefits of someone that knows how to do those things manually.

Example:

/brew A tiger is eating a chicken salad

Behind the scenes, brew is reducing anatomical imperfections, selecting the most appropriate AI <concept>, setting the /guidance, the /sampler, and all that other technical stuff -- it does it all for you. And over time, brew will get even smarter.  Let's take a closer look.

In the figure below, we can observe Morgan's simple input up top, and the two outputs that follow from PirateDiffusion reveal the magic:

A lovely picture, no fussing with concepts and styles and commands. But the long technical prompt is there if I want to get into the nitty gritty with /render.

Or I can brew something new, /remix, /more, and so on. Everything you already learned is still there. Brew is an optional new feature in your toolkit. It is the easy mode of PirateDiffusion, and what we recommend for beginners.

From here on out, we'll refer to /render as the pro mode, or manual mode, like a setting on your DSLR camera.  If you know what you're doing and want crazy control with all the knobs, use /render with all of its 30 parameters. If you're learning or just want something good fast, use /brew. I'm overexplaining this, right? You got this.

Now, let's look at when things go wrong

If you're an old timer, skip this part.  For new people, this is frustrating and I want to help you. Look at this prompt below. We have requested a turtle drinking a beer, but I didn't get my beer.

How do we solve fix this?  Raise your hands if you know!

The AI doesn't know what part of your request is the most important, so you have to tell it what is crazy important to you.  Wrap the most important words with (parenthesis) to make them more important.  You can nest (((((multiple))))) parenthesis to make things even more important. This is called a positive prompt. Imagine each pair adds roughly 15% more power to words. That's the gist of it.

You can also do [[[[negative]]]] prompts and add multiple words, separated with commas.  For example, we don't want a pet turtle, so we can add:

[aquarium, indoors, tank, glass] and those elements will be discouraged.

This is useful when you have unusual requests that go against the forces of nature, such as creating a ((((hairy)))) turtle. We can make all kinds of wild stuff with it.

And what a difference that makes.  Same prompt, with a little uuumphh!

Don't want the extra words in your prompts?

Use /render instead of /brew. That will remove the extra words, like going fully manual. Don't forget to add /size and /steps or the quality will be iffy.

Brew helps train the robot pirate

Right now, our brew system is not very smart because it is working on very little data. Maybe we don't want a cute tiger, even though its adorable.

Please help us improve by hitting favorite on images produced by both /brew and /render, as well as slamming bad on images you don't like.  The system will become smarter over time based on your opinion.  Your voting will help make PirateDiffusion a much more powerful AI, and other folks like it when their stuff gets liked too, so double win for everyone.

More models this week!

Stop on by to our spreadsheet of doom to see what's hot and new. We added URPM, Definite, and more. We'll eventually move this to our website and make it more dynamic, I promise.


For pros -- New sampler, too!  DPM++ 2M Karras

Type /samplers to see the full list

If you don't mind an extra 15 seconds and want a little more detail, we now support DPM++ 2M Karras, which we are calling dpm2m for short.

But our new favorite sampler is upms because it can often do in 20 steps what others do in 50, like a real superhero. It's short for UniPC Sampler. Saves time!

A sampler, (also called a noise scheduler) is like a modulator that impacts image quality in a much more subtle way than selecting a concept. Thus, comparing images from different samplers may not look that different at first, but there is a difference.

The difference is very subtle, but for perfectionists, we think you'll nice. For most of us, a random default is fine. 

You can safely ignore this if you think this is moon speak. But if you'd like to learn more, we have a slightly longer technical blog for you. 


Share your art!  Pirate Diffusion has a subreddit

We're always posting something fun on our subreddit, and it is open for submissions. You can post whatever you like, we're hoping for more examples of what PirateDiffusion can do as our product can be so hard to explain. You can upload 20 images at a time using the Images tab.  We look forward to seeing you there and will be upvoting you.


Render-Ahoy!

Files

Comments

No comments found for this post.