The ten Adobe MAX 2026 Sneaks projects, from brushwork to video

Ten research projects span video, audio, photography, and illustration. Two Sneaks are available to try, alongside Custom Filters.

A brush follows the contours of a photograph; a drag-and-drop gesture adjusts an illustration’s perspective. The Adobe MAX 2026 Sneaks, presented in Miami Beach during a session hosted by Terry Crews, explore how people direct creative tools as much as what those tools can generate. Ten research projects cover images, video, audio, and vector artwork.

These are prototypes with different levels of access. Some demonstrations already have experimental entry points through Adobe Labs, while others remain stage presentations and research descriptions. Links to Photoshop, Premiere, or Firefly on those pages sometimes point to related tools already on the market; they do not mean the corresponding Sneak has been integrated into those products. Painting on surfaces and reshaping objects

For photo editing, Project Deep Brush aims to distinguish surface shape, color, and lighting. The prototype decomposes an image into layers corresponding to these physical properties. A brushstroke can then follow a surface’s contours, a new color can preserve existing shadows, or an edit can change the lighting within the scene.

The described applications range from adding signage to a wall to placing a logo on a bottle or garment. For beauty retouching, the researchers discuss applying makeup and changing hair color while accounting for facial geometry and lighting. These operations start from a photograph whose properties are inferred, without requiring the user to supply a 3D model. The team says it worked with a professional artist to design the experience.

Direct manipulation takes a different form in Project Hands On. The system analyzes the objects in an image, then builds controls directly on the canvas. Control points can appear on a mug’s handle, allowing its shape to change as the points are dragged. The user works on the relevant detail rather than describing each successive correction.

Object segmentation and vision-language models help determine what can be manipulated. The project page connects this work to earlier research on object properties already used in Firefly’s Edit Object tool. The Sneaks prototype adds a content-aware manipulation interface intended for targeted retouching, concept art, and product shape exploration. Transferring edits while keeping the results editable

Copy and paste takes on a different meaning with Project Rinse Repeat. After an edit is made to one area of an image, the system captures the operations and adapts them to another target. Corrections applied to one person could transfer to other members of a group; a text treatment could carry over to a different composition.

The result is recreated using native Photoshop tools, layers, and masks. The changes therefore remain available for manual refinement. This goes beyond applying the same setting uniformly: the prototype selects relevant edits based on the content it encounters and can extend a treatment from several examples. Product photography and campaign variations are among the proposed applications.

For vector artwork, Project Link Up interprets an element’s movement as a composition instruction. Dragging a car into an illustrated parking space can adjust its perspective, scale, and style. The gesture’s path, direction, and speed influence how the elements combine.

The resulting illustration remains editable vector artwork. Described applications include placing objects into scenes, combining illustrations with letterforms, and creating icons. The researchers identify visual continuity as a central challenge: a difference in color, scale, or layering can be enough to make the composition feel disconnected. Developing a character and exploring its setting

Before generating a film, Project Screen Test proposes developing a character through conversation. Its appearance and voice are paired with a personality that the dialogue helps define. The resulting profile then guides video generation, reducing the need to restate every trait in each instruction.

The project originated with Gabi Duncombe, a former filmmaker and designer seeking to recover some of the exchange between directors and performers. The prototype focuses on preparation: exploring a character’s reactions, trying dialogue, or clarifying motivations. Proposed uses also include game characters and conversation simulations, without demonstrating consistent performance across an entire film at this stage.

For the setting, Project Sight Span turns an image into a navigable environment. The process has two stages: a lightweight 3D scene supports camera movement and framing, followed by a separate image-generation step that produces the final shot from the reference and selected viewpoint. The team targets approximately ten to twenty seconds to build the preview environment, rather than guaranteeing that time for the complete final output.

Set preparation, storyboarding, and camera-angle exploration are the main applications described. A reference image can support exploration of a wide shot, an overhead view, or another angle on a product. Areas missing from the photograph must be inferred: the system produces plausible views to explore, without establishing a faithful record of what was outside the original frame. Editing across shots, directing a stream, and controlling timing

Continuity between shots is central to Project Scene Sweep. The system first identifies cuts and the appearances of characters, objects, and locations. An edit is then attached to its subject so it can propagate to the shots where that subject returns, including from different camera angles.

The starting point can be text or an already edited frame. According to the team, the latter provides better guidance; a dedicated editing model was trained for sequences containing multiple shots. Changing wardrobe, removing a prop, or replacing a background element are proposed applications, with the intention of preserving the original performance and pacing. The approach also extends to collections of