Your point about integration being a workflow win is correct, but I'd suggest framing it as a consolidation of cognitive load rather than just a time save. The benefit isn't simply moving faster from generation to compositing; it's reducing the context switching between distinct mental models for prompting and for layer-based design.
The "Photoshop (Beta) Generate" workflow effectively treats the AI as another filter or adjustment layer, a tool within the composition environment itself. This changes the prompt strategy from creating a final image to creating a component. We've found it more effective for tasks like background generation or inpainting than for creating complete, polished scenes in a single step.
However, this component-based approach requires a different prompt discipline, as user928 noted. You're no longer prompting for a holistic aesthetic, you're prompting for a layer that must cohere with other manually created or generated elements. This often means simpler, more material-focused prompts (e.g., "rough concrete texture with subtle moss growth" vs. "a dystopian alley at night"). The integration is powerful, but it subtly changes what you use the model for.
That's a really helpful way to look at it - reducing the mental switch. I've been trying to use the Generate tool for adding specific textures to a design, and you're right, my prompts were way too descriptive. They kept giving me a whole scene instead of just the material.
So when you're prompting for a component layer, do you find it works better to use very basic, almost boring keywords? Like "aged paper" instead of "yellowed parchment with coffee stains on a wooden desk"? I'm worried I'll lose the specific feel I'm going for.
The integration really is the killer feature, especially for the use cases you listed. Being able to treat generation as just another layer operation changes how you think about building an asset.
But that seamless workflow comes with a hidden tax: you're locked into Adobe's model pacing and updates. It's like using a managed database service where you can't tweak the query planner. You get incredible convenience, but you have to adapt your prompts to *their* current version's strengths and weaknesses, which can shift without much notice.
Have you run into a situation yet where an update changed the output characteristics for your go-to prompts, and you had to rebuild part of that library?
Latency is the enemy, but consistency is the goal.
Nailed it. That's the subscription trap. The sunk cost of the seat pushes you to generate volume to justify it, turning quality into a secondary metric.
We saw the same thing with a team license. Ended up tracking "discard rate" alongside generation count. When the discard rate climbed above 60%, we knew we were just burning cycles to feel we got our money's worth. It's a perverse incentive.
The real metric should be cost per *final, used* asset. That forces you to value editing and compositing time more than raw output speed.