Skip to main content

Loose painting to define areas of an image.

Feature RequestHigh Priority

Oftentimes when creating an illustration you are working with various requirements. For instance you have to leave room for a headline or the client’s product, or you have two figures fighting and want their swords to cross in just a certain way. Blocking-in the illustration lets you paint a broad stroke or lump of color for each of the elements in your prompt. Each element gets a different color and then you can prompt with <red> or <blue> tags to let the AI know where the elements go in the prompt. This is a really useful tool and takes AI from “hoping you get something useful” to creating an image that fits a requirement. All jobs have requirements.

Posted by oskyguy ·
2 comments

Log in to comment and vote

Comments2

  • oskyguy

    •

    Jun 4

    I wanted to share a technique that might make this feature and the in-painting feature happen faster. The trick is to use a control image as a way to communicate to the AI where to place various scene elements. I don’t know if this is how in-painting is done IRL but I got the idea from a Reddit discussion on how to make a fantasy map follow a sketch.

    The goal was to create a section of a bigger map that was a canyon with a waterfall falling into it. It had a very specific shape so on a whim I created a picture out of the map’s sketch and filled in each area with colors. Then used the colors in the prompt and told the AI what to put in each area.

    Here’s the control image:

    I told the AI to put the canyon down in a pit where the control image was white. The black were cliffs surrounding the canyon. Put a river where it was green, put a waterfall where it was pink, put a lake at the bottom of the canyon where it was blue. The yellow was a forest.

    I told the AI to “make this image strictly conform to the shape of the reference image.”
    It took several tries but here was the result:

    For some reason Nano Banana 2 created a second river and didn’t follow the shape of the control image at the bottom but I was surprised at how closely the rest of the control image was followed.

    It seems like if you added some really basic paint tools (just color and brush size, no antialiasing) just as a way to indicate areas on an image, you could take the user’s painted input and create a control image that was fed to the AI.

    In-painting could be done the same way. Paint some color where you wanted the AI to apply the prompt and then feed the control image to the AI along with the prompt.

    Hope that helps.

  • oskyguy

    •

    May 31

    Another example of when this feature becomes essential is when you have an element that does not fit one of your preset aspect ratios. For instance if you want a city-scape off in the distance that fits the image side to side but is not very tall. That’s very difficult to prompt into existence. You’ll always get a city that fills the entire picture top to bottom, or you have to fake out the AI by asking for a lake in the foreground just so it restricts the cityscape to the top half of the image. It would be great to just draw a rectangle of color that is the size that you want it to be and then the AI can fill it in.