Visual worlds for websites are increasingly being created with generative AI. Two current developments show where the technology currently stands: on the one hand, there are research systems that no longer understand image editing as a single generation step, but as a sequence of targeted tool interventions. On the other hand, there is everyday practice, in which the work is spread across many individual tools and consistency becomes a labour-intensive chore.

Both strands describe the same fundamental problem: a single prompt in a single tool is rarely enough for professional results. What matters is how the steps in between are organised.

EditAgent: the agent as tool selector

Adobe Research, together with UC San Diego, has presented a system called EditAgent. The approach is modelled on how experienced image editors work: instead of solving everything with a general-purpose editor, the agent selects a suitable specialised tool for each subtask.

In this context, an agent is an AI system that not only generates text or images, but independently plans steps and calls external tools. EditAgent adds a second mechanism: a language model evaluates the generated variants based on concrete metrics and selects the best result. Quality control thus moves into the process itself, instead of only taking place at the end through a human.

According to the research paper, the gain in reliability comes at a clear price: the method requires around ten times more computing time than the simpler approach. This is less a technical obstacle than a question of calculation — in which cases does a better result justify the higher effort?

Consistency as the real hurdle

The second perspective comes from practice. High-quality AI images and videos are rarely created with a single tool. Instead, the creative process is spread across several browser tabs: generating in one service, post-processing in the next, upscaling in a third.

In between lie downloads, uploads, copy-and-paste of prompts and the management of file versions. What sounds like a detail problem causes real effort in projects — and it is the main cause of motifs drifting apart. If a character looks different on the home page than on the landing page, this is rarely down to the model, but to a workflow that does not pass intermediate states on cleanly.

Both sources therefore start at different points, but aim in the same direction. EditAgent automates tool selection within an editing step. The workflow perspective addresses the gaps between the steps. Consistent visual worlds require both.

What this means for image production in web projects

For agencies and internal teams, this results in some practical consequences:

  • Tool diversity is normal. The idea that one service covers the entire image production process does not hold up. Plan from the outset with a chain of several tools.
  • Handovers are the critical point. Where files are manually downloaded and uploaded again, version errors and inconsistencies arise. Every automated handover step saves time and reduces deviations.
  • Quality checks can be partially automated. The EditAgent approach shows that machine evaluation based on defined criteria can make usable selection decisions — not as a replacement for the human eye, but as a pre-filter.
  • Computing time is a budget item. A factor of ten in computing time is justifiable for a single key visual, but rather not for the mass production of teaser images. The decision should be made per use case.

Safeguarding consistency organisationally

Regardless of the technology used, it helps to treat image production like any other production process. This includes traceable storage of prompts and intermediate states, clear naming of versions and the definition of which characteristics of a motif are binding — such as colour world, perspective or recurring image elements.

The more precisely these specifications are documented, the more readily they can later be transferred into automated steps. An agent can only reliably implement what has previously been described as a criterion. Teams that already record their visual language in a style guide have an advantage here.

At the same time, a sober look at the level of maturity is worthwhile. EditAgent is a research project, not a product that will be in the production environment tomorrow. However, the direction described — agents that specifically operate specialised tools and evaluate results themselves — is clear enough to take into account in planning.

Assessment

For web and software projects, the bottleneck is shifting from image generation to process design: the challenge is not the individual motif, but the repeatable, consistent production of many motifs across a project. Anyone who now defines responsibilities, storage structures and quality criteria for AI images can later plug in agent-based tools without rebuilding.Conversely, even the best model is of little use if the handovers between the tools are still done by hand.

Sources