Adobe opens two MAX Sneaks previews as it tests finer AI editing control

Creating a 3D environment from an image is already available elsewhere. World Labs, an AI developer building models of three-dimensional environments, made its Marble world model generally available on November 12, 2025. Its launch announcement describes generating worlds from a single image, editing and expanding them, and exporting 3D representations or videos. Its experimental Chisel mode also lets users guide a world's structure with a rough 3D layout. Direct creative control is therefore part of the competitive baseline, too.

World Labs introduced its next-generation Atlas model on September 1, 2026. In that announcement, the company says Atlas generates new views from one or more reference images using specified camera positions and angles, and was available to try. These capabilities establish an alternative to waiting for Sight Span; they do not establish equivalent output quality, speed or workflow compatibility.

Adobe’s Scene Sweep illustration shows wardrobe variations for video editing. Company-reported research demonstration. Source: Adobe. Source: Adobe.

Ten experiments target different editing bottlenecks

Adobe's announcement describes the following capabilities. These are company descriptions and demonstrations, rather than independently measured performance results:

Some demonstrations addressed familiar manual work. In the on-site account, Deep Brush followed folds in fabric as lettering was painted onto it. Hands On reshaped a square perfume bottle into a diamond and stretched a sofa. Rinse Repeat transferred blemish-removal work from one face to three others, adapting the changes to different blemish locations.

The significance of Rinse Repeat and Link Up is also in what remains afterward: native Photoshop edits and vector artwork can be refined further. That preserves a route for correction rather than making the generated result the endpoint.

Adobe is testing interfaces as well as models

The brush and drag controls have a research rationale. Adobe's research, described in the session report, draws on eight studies involving 82 working and aspiring creative professionals from 2024 to 2026, plus a review of research papers. The researchers argue that creators often discover their intent through making and revising work rather than fully describing it beforehand. They acknowledge that the sample leans toward visual disciplines and is not representative of all creative workers. Those findings do not establish a productivity gain from the Sneaks tools.

Hands On also builds on an existing Adobe effort. According to the report, Adobe says it grew out of the Edit object tool in Firefly, its generative AI offering, and that models were initially too slow for the idea. Adobe's announcement describes the current approach as combining object segmentation with vision-language models to create controls on the canvas. Adobe also says, as relayed in the session report, that its team trained its own editing model for Scene Sweep's multi-shot video edits. The sources do not supply hardware requirements or latency measurements to show how these systems would perform in production.

There is a path from Sneaks to commercial software. The demonstration report traces Photoshop's Harmonize feature, introduced in 2025, to Project Perfect Blend at the October 2024 Sneaks. The on-site account also relays Adobe's statement that Relight, shown at the previous year's Sneaks, is now a standard Photoshop feature. Those precedents explain why the experiments matter to existing users, without guaranteeing a release for this year's projects.

The next test is work outside the demonstration

The Playground extends that feedback process beyond the stage. Alongside the two Sneaks previews, Adobe offers Project Custom Filters, a separate research preview of a possible Photoshop capability. The session report describes it as turning a text description into a live filter with adjustable sliders. It is not an eleventh project in the Sneaks lineup.

The outstanding questions are practical: how reliably edits hold across difficult shots, how quickly object controls respond, and how much correction a generated viewpoint requires. The Sneaks announcement and reports provide no comparable quality benchmarks, measured time savings, pricing or production-cost figures for the lineup. They cannot yet support a cost or performance comparison with alternatives.

Adobe's next decision is which experiments merit product integration, and on what terms. Feedback from Deep Brush and Hands On can test a narrower question now: whether direct manipulation helps creators reach and refine an intended edit. Broader claims about faster or cheaper production will require evidence beyond the stage demonstrations.