For months I kept doing the same thing over and over. I would ask an AI for one small change to a picture. Move the logo a little to the left. Warm up the sky. Drop the character's arm a bit. And almost every time, what came back was a completely different image. The lighting had shifted, a face looked subtly wrong, and the one part I actually liked was gone.
So I would try again, a little more carefully, and lose something else instead. By the third attempt I was not really designing anymore. I was gambling, and the house kept repainting everything I liked.
I want to explain why this happens, because it is not your prompts and it was not mine. It is the file itself. Figuring that out is the whole reason I ended up building Rayzia the way I did.
A PNG has nothing to edit
Here is the thing that took me too long to accept. When a model hands you a PNG or a JPEG, it hands you a grid of coloured pixels. There is no logo in there, no sky, no arm. Just millions of tiny squares that happen to look like those things to your eye. So when you ask it to move the logo, it cannot pick the logo up, because as far as the file is concerned there is no logo to pick up. All it can do is paint a fresh grid and hope the result reads as the same picture with the logo moved. Every pixel is back in play every single time. That is why everything drifts.
Diffusion models, which is what most image tools run on, actually make this worse. They build the picture by cleaning up noise across the whole frame at once, so one small instruction spreads out over the entire canvas. There is no line between the bit you wanted changed and the bit you wanted left alone, because the file never had that line to begin with.
"Editable" usually means the wrong thing
For a while I thought the fix was just to export an SVG. SVG is a vector format, and vectors are supposed to be editable, so surely that solved it. Then I opened a few of them. What I found was one enormous path, traced automatically off the pixels, with hundreds of nodes drawing a curve I would have made with maybe six. A little sunburst that should be a dozen points showed up as eight hundred.
You can technically edit that, in the same way you can technically rewrite a paragraph by dragging individual letters around. But the structure I actually needed, the part that says this is the sun, that is the roof, here is the headline, was never in there. It got reconstructed after the fact from a flat image that never had it. So I stopped believing that "editable" meant anything just because a file ended in .svg. What matters is whether the file knows what its own parts are.
So I made the AI draw objects instead of painting pixels
That is the whole idea behind Rayzia. Instead of generating an image and handing you a file, the AI sits inside a real vector editor and draws actual objects onto the canvas. A rectangle, an ellipse, a path, a line of text. The same objects I would make by hand with the pen or the shape tools. Each one is a real thing the document knows about, sitting in its own place in the layer list.
Two things fall out of that, and they are the entire point.
The objects have names. When the AI draws a house, it does not dump "Path 1, Path 2, Path 3" on you. It calls them Roof, Door, Window, Sun. Open the layer panel and it reads like the picture.
An edit only touches what you named. Ask to move the sun and the sun moves. Nothing else in the scene budges, because nothing else was addressed. There is no repaint of the whole frame, because there is no frame being painted. There are just objects being moved. And every single thing the AI does is one undo step, so if it reads me wrong, one Ctrl+Z puts it straight back. I am never risking the whole canvas on one sentence.
"Move the moon left and make it warmer"
Say the AI drew me a little night scene: some hills, a few stars, a moon. I click the moon and type:
move it left a little and make it warmer
The moon slides over and its colour warms up toward amber. The hills stay put. The stars do not flicker. The composition I was happy with is still the composition I was happy with, because that instruction reached one named object and stopped there.
Then I ask for the thing a generator can never really do. I type add a second, smaller moon, and it shows up as its own new layer next to the first one, with the rest of the scene left alone. I am adding to a drawing, not pulling the lever on a slot machine again.
And then I take the pen
This is the part I care about most, and it is the part a pure generator can never give you, because there is nothing underneath to grab. The moon the AI drew is just a normal vector object. So I can open it with the node tool and drag a single anchor point to carve one edge into a crescent by hand. I can change its gradient stops. I can give it a soft glow from the effects. Anything the AI made behaves exactly like something I drew myself, because it is the same kind of object. There is no special "AI layer" that lives by different rules.
The AI is a fast first draft and a tireless pair of hands for the boring parts. I still hold the pen for the decisions that actually matter. That is the thing the pixel loop could never give me. It could only hand me a finished picture and then take the whole thing back the moment I asked for a change.
Try it in your browser
Open a canvas, type "move the moon left and make it warmer", and watch only the moon move. It is free, there is nothing to install, and there is no sign-in wall in the way before you can start drawing.
Common questions
Why does AI regenerate the whole image when I only ask for a small change?
A PNG or JPEG is a flat grid of coloured pixels with no separate objects inside it. There is no logo or sky for the model to pick up and move, so it repaints a whole new grid every time. Diffusion models work across the entire frame at once, so any instruction spreads over the whole picture and things drift.
Is exporting to SVG enough to make an AI image editable?
Usually not. An SVG traced automatically from a raster image is often one huge path with hundreds of nodes describing a shape a person would draw with a few. You can open it, but the structure you actually need, like knowing which part is the sun and which is the roof, was never captured.
How does Rayzia avoid changing the whole picture?
Its AI does not generate an image. It works inside a real vector editor and draws separate objects on a canvas, the same ones the pen and shape tools make. Each object is named and sits in the layer list, so an edit only reaches what you addressed. Move the sun and only the sun moves, and every action is a single undo step.
Is it free, and does it use my own AI account?
The editor is free and runs in the browser with no sign-in wall before you start drawing. The AI runs on an account you already own. You connect Claude or GPT, and it works on your key, so Rayzia does not add any AI charge on top of what you already pay your provider.
One honest note. I build Rayzia on my own, and it is in public beta, with new things shipping most weeks. It is not finished, and I am not trying to replace your whole toolkit. I just wanted to fix one specific, maddening thing: that asking an AI for a small change should change one small thing.