The 80% Problem: Why the Most Valuable Part of Visual AI Isn't Image Generation

The visual AI conversation over the past three years has centered almost entirely on generation. New models are benchmarked on how well they create images from text prompts. Product launches are measured in resolution and photorealism. The headlines are always about what AI can generate from nothing.

But commercial visual content does not start from nothing.

Brands have millions of existing images: campaign assets, product photography, catalog imagery, stock libraries, user-generated content, partner-supplied visuals. The work that consumes the majority of creative production time is not creating new images from scratch. It is adapting, modifying, repurposing, and scaling what already exists.

Changing a background for a different market. Removing an object that should not be in the frame. Adjusting the lighting on an existing image to match a campaign mood. Restoring and enhancing older assets to current quality standards. Extending a product image to fit a different aspect ratio. Erasing elements from a video. Removing a background from video footage.

This is the 80% that generation does not solve. And for most enterprise teams, it is where the real production cost lives.

The problem is that editing has been treated as an afterthought in the visual AI landscape. The tools that do exist tend to fall into three categories, none of which work at production scale.

Manual creative tools that are powerful but require skilled operators for every edit, making them impossible to scale programmatically. Single-purpose AI tools that solve one problem well – background removal, for example – but force teams to stitch together multiple vendors, models, and integrations to cover the full range of editing operations. And generation platforms that bolted on basic inpainting as an add-on but never built a comprehensive editing layer.

This gap pushes enterprises that want to edit visual content at scale to assemble fragmented toolchains, managing multiple vendors, and still relying on manual work for anything beyond basic background removal. The generation side is automated. The editing side? the larger part of the work is not.

The shift is structural. What enterprise teams need is not another single-purpose editing tool. They need a comprehensive editing layer – a single integration that covers the full range of visual modifications, from simple background operations to complex text-driven scene edits, across both image and video.

Consider what a typical creative production workflow actually requires. A campaign launches with a hero image, and then that image needs to be adapted: background replaced for a different retail partner, reseasoned for a holiday campaign, restyled for a different market aesthetic, expanded to a wider aspect ratio for a billboard, relighted to match an evening setting, and elements removed or added to comply with regional requirements. In video, the same asset may need background removal, object erasure, or resolution upscaling for broadcast delivery.

Today, each of those operations is a separate tool, a separate vendor, or a separate manual task. Every handoff between systems introduces delay, cost, and quality inconsistency. The more edits required, the more the toolchain fragments.

What needs to change is the integration model. When every editing operation, from removing a background to reseasoning an entire scene, from erasing an object in video to increasing resolution for broadcast, runs through a single platform with consistent behavior, the production model transforms. Editing becomes programmable. Multi-step workflows become automatable. And the 80% of production time currently spent on adaptation becomes as scalable as generation already is.

Not every platform that offers an editing feature is built for production-ready editing. Five capabilities separate a comprehensive editing layer from a generation platform with a few add-ons:

Breadth of operations under a single integration. A platform that handles background removal but not object editing, or style transfer but not lighting adjustment, forces you back into the multi-vendor toolchain. The editing layer needs to be comprehensive: background operations, object-level editing, scene-level transformations, enhancement, expansion, restoration, and video editing. Count the operations. If you need five vendors to cover your editing needs, you don't have an editing layer – you have a procurement problem.

Editing described as structured operations, not manual interactions. Production-scale editing needs to work for people and automated systems equally. Edits should be structured instructions that a creative director can specify once, and an automated workflow can repeat across thousands of images without reinterpretation. If every edit requires a human guiding the output, you've automated generation but left editing manual.

Consistency across operations. When background removal comes from one model, style transfer from another, and object editing from a third, the results are visually inconsistent. Colors drift between operations. Quality varies. Edits interact unpredictably. Look for a platform where the editing operations share a common model architecture, so the visual language stays consistent across every modification.

Image and video in the same platform. Creative content is increasingly cross-format. A campaign asset might start as a still image, get edited, and then need the same operations applied to a video version. If image editing and video editing live in separate platforms with separate integrations, the workflow doubles. Look for video background removal, video object erasure, and video enhancement alongside the image editing suite.

Commercially safe from input to output. Every edit creates a derivative work – so the model needs to be built on licensed content with clear legal coverage. But safety goes beyond training data. Every prompt and every output should pass through content moderation and governance checks before anything reaches production, regardless of whether it was triggered by a person or an automated workflow. And for organizations that can't send assets externally, the platform should run in your environment with full data sovereignty.

Over 20 image editing operations and a dedicated video editing suite, through a single integration. Background operations, object-level edits by text, scene transformations – relight, restyle, reseason, colorize, restore – image expansion, resolution enhancement, and sketch-to-image. For video: background removal, object erasure, and resolution increase up to 8K. All under one model family, so every edit is visually consistent with the last.

That's what Bria's Visual Engine delivers.

One integration replaces the fragmented toolchain. Edits are structured and programmable – describable by people, repeatable by automated workflows, chainable across operations. Every output is commercially safe: models built on 100% licensed content, content moderation on every input and output, and infrastructure that runs in your cloud, your data center, or on-site.

Generation gets the headlines. Bria built the editing layer that makes visual AI actually work in production.

The best way to evaluate an editing platform is to test it against your own images. Bring an asset from your library and see what the editing APIs do with it. The full suite is available for experimentation in the Bria Developers Platform. Full documentation at docs.bria.ai.

Why does it matter whether AI editing is structured or prompt-based? Prompt-based editing is ambiguous – the same instruction can produce different results each time. Structured editing means every visual parameter is explicit, independent, and reproducible. Change the lighting without affecting the composition. Adjust one object without the rest of the scene drifting. This is the difference between editing you can direct and editing you have to retry until it works. It's also what makes editing automatable: structured operations can be chained, validated, and repeated across thousands of images by automated workflows and agents, not just by a person in a UI.

Can AI edit videos the same way it edits images? The most production-relevant video operations – background removal, object erasure, and resolution increase – are now available as programmable endpoints. The key is whether video and image editing live in the same platform under the same integration. If they do, your production workflow stays unified across formats. If they don't, you're managing two toolchains for content that increasingly needs to move between still and motion.

Is AI-edited content commercially safe for enterprise use? It depends entirely on the platform. Every edit creates a derivative work, so the same questions that apply to generation apply here – often with higher stakes, since edits are typically applied to existing brand assets. Look for models trained on fully licensed content, legal coverage that extends to edited outputs, built-in content moderation on both inputs and outputs, and the ability to run the platform in your own infrastructure when your data can't leave your environment.

What is the difference between a platform with editing features and a dedicated editing layer? Most visual AI platforms are generation-first, with a few editing add-ons – typically basic inpainting or background removal. A dedicated editing layer means 20+ independent endpoints covering backgrounds, objects, scenes, enhancement, expansion, restoration, and video. The difference shows up in production: when your next editing need is already covered by your existing integration rather than requiring a new vendor evaluation.