What you can make
Images
Making a picture from nothing, changing one you already have, or seeing eight of something at once, and which tools do each.
Three questions decide which image tool you want. Are you making a picture from nothing, changing one you already have, or trying to see several directions at once? The sidebar is grouped that way. ## From nothing **text-to-image** puts the current generation of models behind one prompt box: Nano Banana Pro, FLUX 2 Pro and Max, Recraft 4, GPT Image 2 and 1.5, Qwen 2 and 3, Ideogram v4. Type the prompt once, run it, then swap the model pill and run the same words again. Models read the same sentence differently, and two readings side by side on the canvas settle the question faster than rewriting the prompt for one of them. Ideogram brings its own layout canvas. Drag regions to say where the text and the objects go, then let it render into that arrangement. Lock the card with the padlock and repeated runs scatter into a grid instead of replacing each other. ## From something you already have **edit-images** is the same idea for editing: Nano Banana in four sizes, GPT Image 2 and 1.5, FLUX 2 Pro and Max, Seedream 5 Pro, Reve Remix, P Image, Qwen Image 3, all on one card. Drop up to four images onto it and address them in the prompt with `@`, as in "put @image-2's logo on @image-1". Anything already on the canvas reaches this card from its own action bar, so you rarely go hunting in the sidebar for it. ## Several at once In this group the model barely matters, because the work is in the tool. Each one art-directs a whole set for you and returns it from a single run. | Tool | One run returns | |---|---| | variations | Up to 8 takes of your image along one axis, colour or proportion or grade | | antagonist | Up to 6 cells, each breaking a single assumption while everything else holds | | diverge | Up to 6 complete key visuals, new scale, world, material and light in every cell | | same-family | Up to 8 relatives of your subject, a different individual each time, same world | | time-machine | Your object as any year you type could actually have made it, up to 6 ways | | close-ups | Four detail crops of what you dropped, or two wide 16:9 stripes | | layerize | The picture pulled apart into transparent layers, up to 17 of them | Most of them cost one generation rather than eight. The model returns a single grid image and Plnty slices it into separate canvas images afterwards, so eight takes bill as one run. The tools in that table are opinionated on purpose. **antagonist** picks an assumption your reference is making, breaks that one thing and freezes the rest, so you can name exactly what changed. Expect at least one cell to come back worse than what you started with. A grid where every cell is an improvement means the tool flattered you six times and told you nothing. ## Finishing **image-upscale** holds four engines with different jobs. P Image Upscale enlarges faithfully and fast, SeedVR does 2x and 4x with a little cleanup on the way up, and Topaz invents detail rather than enlarging it: pores, fabric weave, print grain. :::caution[Topaz is a gamble on purpose] Invented detail is sometimes exactly what a small source needed and sometimes very obviously machine-made. Reach for P Image or SeedVR when the pixels are right and just too few. ::: Around those sit the small repair tools: **remove-bg**, **extract-object**, **image-isolate**, **inpaint**, **typography-remove**, **splitter**. None of them are interesting to read about and all of them are the reason a generated image becomes usable.Three questions decide which image tool you want. Are you making a picture from nothing, changing one you already have, or trying to see several directions at once? The sidebar is grouped that way.
From nothing
Section titled “From nothing”text-to-image puts the current generation of models behind one prompt box: Nano Banana Pro, FLUX 2 Pro and Max, Recraft 4, GPT Image 2 and 1.5, Qwen 2 and 3, Ideogram v4. Type the prompt once, run it, then swap the model pill and run the same words again. Models read the same sentence differently, and two readings side by side on the canvas settle the question faster than rewriting the prompt for one of them.
Ideogram brings its own layout canvas. Drag regions to say where the text and the objects go, then let it render into that arrangement.
Lock the card with the padlock and repeated runs scatter into a grid instead of replacing each other.
From something you already have
Section titled “From something you already have”edit-images is the same idea for editing: Nano Banana in four sizes, GPT Image 2 and 1.5, FLUX 2 Pro and Max, Seedream 5 Pro, Reve Remix, P Image, Qwen Image 3, all on one card.
Drop up to four images onto it and address them in the prompt with @, as in
“put @image-2’s logo on @image-1”. Anything already on the canvas reaches this card
from its own action bar, so you rarely go hunting in the sidebar for it.
Several at once
Section titled “Several at once”In this group the model barely matters, because the work is in the tool. Each one art-directs a whole set for you and returns it from a single run.
| Tool | One run returns |
|---|---|
| variations | Up to 8 takes of your image along one axis, colour or proportion or grade |
| antagonist | Up to 6 cells, each breaking a single assumption while everything else holds |
| diverge | Up to 6 complete key visuals, new scale, world, material and light in every cell |
| same-family | Up to 8 relatives of your subject, a different individual each time, same world |
| time-machine | Your object as any year you type could actually have made it, up to 6 ways |
| close-ups | Four detail crops of what you dropped, or two wide 16:9 stripes |
| layerize | The picture pulled apart into transparent layers, up to 17 of them |
Most of them cost one generation rather than eight. The model returns a single grid image and Plnty slices it into separate canvas images afterwards, so eight takes bill as one run.
The tools in that table are opinionated on purpose. antagonist picks an assumption your reference is making, breaks that one thing and freezes the rest, so you can name exactly what changed. Expect at least one cell to come back worse than what you started with. A grid where every cell is an improvement means the tool flattered you six times and told you nothing.
Finishing
Section titled “Finishing”image-upscale holds four engines with different jobs. P Image Upscale enlarges faithfully and fast, SeedVR does 2x and 4x with a little cleanup on the way up, and Topaz invents detail rather than enlarging it: pores, fabric weave, print grain.
Around those sit the small repair tools: remove-bg, extract-object, image-isolate, inpaint, typography-remove, splitter. None of them are interesting to read about and all of them are the reason a generated image becomes usable.