Layers are the interface
A generated image becomes truly useful when we can still reach the decisions beneath the pixels.
Most image generators are very good at endings. They produce the flattened result: one rectangle of pixels, ready to admire and difficult to change. But visual design rarely happens that way. A design is a stack of decisions—layout, hierarchy, type, shape, color, subject, background—and each decision remains useful only while it stays addressable.
This is why I think layered generation is more than a file-format problem. The layers are the interface between the model's synthesis and the designer's intent.
The gap between layout and image
Layout models understand boxes, relationships, and spatial constraints. Image models understand appearance. The interesting problem sits between them: how can a system carry structure into synthesis without treating the layout as a weak hint, and without making the image feel assembled from disconnected pieces?
Grounded attention and region-level alignment are useful steps, but control should survive the generation process. If the user asks to move one object, replace one texture, or reveal the background, the representation should make that operation natural rather than forcing a full restart.
Generate a working artifact
My working hypothesis is simple: the target should not only be a convincing image. It should be a convincing working artifact. That means coherent RGBA layers, meaningful grouping, stable occlusion, and enough semantic structure for another person—or another model—to continue editing.
The best generation system may not be the one that gives the most finished answer. It may be the one that leaves the most useful possibilities open.