Compositing is still the job
The professional's actual workflow
There is a widespread assumption that using these tools well means getting a finished image out in one go. Almost nobody working commercially does that. The output of a generative model is treated as stock — raw material, gathered and then assembled.
A typical commercial piece is:
- A generated background, upscaled and graded.
- A subject generated separately, cut out.
- Two or three small elements generated or sourced individually.
- Type set in a real typesetting tool.
- Colour, grain and grade applied over the whole thing.
Every one of these steps happens in an ordinary image editor. None of them is a prompting problem.
Why assembly beats generation for exactness
The reasons are the mechanisms from the previous module, all at once. Counting works when you place the objects. Symmetry works when you flip a copy. Relations work when you position things. Text works when you set it. Colour matching works when you sample it.
Every failure that comes from the absence of a global constraint disappears when a human supplies the constraint. That is not a workaround. It is the correct division of labour: the model produces surfaces and textures that would take hours to paint, and you produce the structure it cannot hold.
The free tools, named
Everything here can be done for nothing, and the free tools are genuinely capable rather than a consolation:
- GIMP — layers, masks, selections, colour. Free, offline, runs on modest hardware.
- Krita — better for painting and drawing over generations, excellent brush engine, free, and its AI Diffusion plugin runs local generation inside the canvas.
- Photopea — runs in a browser, opens PSD files, and behaves like the industry tool closely enough that skills transfer in both directions. Free with advertisements.
- Inkscape for vector work, Scribus for print layout, Blender for 3D and its surprisingly good compositor.
The paid names — Photoshop, Illustrator, InDesign, After Effects — matter because job advertisements name them, and it is worth being able to say you know what they do. The operations are the same operations. Someone who can composite in GIMP can composite in Photoshop within a day.
The five techniques that carry most of the work
If you learn only a handful, learn these:
- Layer masks. Never erase; hide. Everything stays editable, which means every decision is reversible.
- Cutting out with a decent edge. Selection, refine, a small feather, and attention to hair. The single most common tell of an amateur composite is a hard cut edge.
- Matching light and colour. Sample the background's shadow and highlight colour and grade the subject toward it. Most "it looks pasted on" problems are colour, not geometry.
- Adding a unifying pass. A shared grain, a shared slight blur, a shared grade over everything. This is what makes disparate sources read as one photograph.
- A final low-denoise generative pass at 0.15 to 0.25 over the assembled image, which smooths the remaining inconsistencies. This is the one step that did not exist before, and it is the reason composites are now much faster than they used to be.
The argument for learning this properly
The generative part of this work is the part being automated hardest. Each model release makes the prompt matter less. What does not get easier is knowing that the shadow is falling wrong, that the edge is too hard, that the type is set badly, that the colour is off by a degree.
Those judgements are ordinary craft, they are the same judgements a retoucher made in 1995, and they are what the person paying you is actually buying. The rest of this track — design foundations, image editing, vector work, print — is exactly this material, and it is worth more than another year of prompt vocabulary.
The honest limitation of compositing: it is slower per image and it does not scale to volume. If you need four hundred variations, you need generation and automation, not assembly. Know which job you are on. Assembly for the hero image; generation for the long tail.
There is a second limitation, and it is about records rather than craft. A composite has several sources, and each of them carries its own rights position — one generated, one licensed stock, one photographed by the client, one taken from a reference the client supplied without saying where it came from. When a question arises months later it arrives about the finished piece, and the only way to answer it is a layered file with named layers and a note saying where each element came from. Flattening a composite and keeping only the export destroys that record permanently. Keep the layered file.
The one thing to keep
Generated elements assembled by hand in an ordinary editor beat single-pass generation for anything that must be exact, which is why the finishing skill matters more than the prompting skill.
Before you move on
Why is a hand-assembled composite usually more reliable than a single-pass generation for a commercial hero image?
Pick the one you would defend. Nobody sees your answer.