top of page

Adobe MAX Sneaks previews explorable photos and AI characters

21 minutes ago
11 min read

Adobe previewed 10 research projects at Adobe MAX Sneaks, including experiments that turn photographs into explorable scenes and develop fictional characters through conversation. Announced on November 11, the projects investigate new ways to direct creative software. They also leave an important distinction intact: a research presentation is not a product release.

The most interesting connection is control. Adobe is exploring how people can communicate decisions through camera movement, conversation, and examples of their own edits. The question is whether software can carry those decisions forward while leaving creators enough room to inspect and revise the results.

That puts the experiments alongside an existing challenge for tools such as Runway: maintaining creative intent across changing scenes and repeated generations. Adobe’s research suggests a different interface for that challenge. Instead of describing every adjustment from scratch, creators establish a direction and ask the software to apply it elsewhere.

Adobe MAX Sneaks moves beyond the prompt box

Adobe’s creative research previews span video, sound, photography, and design. Sight Span explores new views from an image. Screen Test develops characters through dialogue. Rinse Repeat adapts edits across image regions, while Scene Sweep targets changes across video shots.

The shared ambition is to make a creative decision reusable without forcing the creator to specify every operation.

That matters because producing an attractive image is only part of a creative assignment. A campaign needs related assets. A film needs continuity. A client might approve a direction, then request changes that must carry through the entire project.

Those situations expose the difference between generating something and directing work toward a specific result. A plausible output can still miss an approved detail. A beautiful revision can create another problem elsewhere. Creators need ways to express both what should change and what should remain stable.

The experiments approach that problem through familiar activities. Moving a camera expresses composition. Talking about a fictional person expresses character direction. Copying an edit expresses a reusable decision. Each interaction provides context that a standalone text prompt might leave ambiguous.

This does not eliminate prompts. It changes their role within a broader editing process. Language becomes one input among visual examples, movement, and existing work. The resulting interface asks creators to show more of their intent through actions they already understand.

There is a practical distinction between that approach and an all-purpose conversational assistant. A conversation can describe an outcome without providing a useful way to inspect it. A specialized creative interface must connect the request to visible, adjustable changes.

Consider an illustrative advertising assignment. An art director approves a product image but asks for a different composition across several placements. Generating unrelated alternatives might produce attractive options. It does not necessarily preserve the relationships that made the original image acceptable.

The value of contextual control is therefore selective change. The creator wants the software to understand which parts of a decision should travel and which should stay local. That is a demanding requirement because visual similarity alone does not establish creative correctness.

Adobe MAX Sneaks presents research directions rather than evidence that this requirement has been solved across production work. The projects are useful signals of what Adobe is investigating. Their eventual value depends on how well those interfaces handle revisions beyond the examples selected for presentation.

Sight Span turns a photograph into a place to compose

Adobe says Project Sight Span builds a lightweight, three-dimensional preview from an image, then generates a finished shot using the chosen viewpoint. Its presentation includes a panda photograph, a portrait, and an exploration of Johannes Vermeer’s The Milkmaid.

A new viewpoint offers creative freedom, but it does not reveal verified information about everything outside the original camera’s view.

The distinction follows from the input. A photograph records a particular perspective. It does not contain a complete description of hidden surfaces and spaces. Moving beyond that perspective requires the system to infer what those unseen areas might look like.

Adobe’s project description acknowledges that inference. The result should therefore be understood as a generated interpretation grounded in an image. It is not equivalent to returning to the original location with a camera or measuring a physical scene.

For concept development, that limitation can be acceptable. An imagined room can help an art director evaluate a composition. A generated angle can help a filmmaker discuss staging. The output’s usefulness depends on the assignment, rather than a universal standard of photographic accuracy.

For documentary or evidentiary work, the same ambiguity carries different consequences. A convincing generated view should not become proof of an object’s hidden appearance. Readers and clients need to understand when an image represents a creative construction.

The interface is especially interesting because it makes composition spatial. A creator can evaluate a viewpoint through movement rather than repeatedly translating camera intentions into written descriptions. That creates a more direct relationship between the action taken and the image being considered.

Imagine a production designer reviewing a proposed interior with a director. The useful question might be whether a lower viewpoint strengthens the scene. Exploring that possibility visually can support the discussion before anyone commits to building or photographing the setting.

That example describes a possible workflow, not a documented customer deployment. It illustrates why an explorable preview can matter even when it is imperfect. Early creative discussions often need a shared visual reference more than a finished production asset.

However, the preview and finished output must remain meaningfully connected. If the final image changes important relationships, the creator has to reconsider an already approved composition. An interactive interface loses value when its apparent precision does not survive generation.

Useful evaluation would therefore examine continuity between the selected view and the final result. Creators should inspect identity, spatial relationships, and details important to the brief. A pleasing image is insufficient if the requested camera decision becomes difficult to recognize.

The broader implication is that image generation is becoming an environment for making decisions, rather than merely receiving outputs. Sight Span’s research direction is compelling on those terms. Its accuracy, practical limits, and eventual availability remain separate questions.

Screen Test makes character direction conversational

Adobe’s Project Screen Test uses conversation to build an editable personality profile that guides video generation. Presenter Gabi Duncombe developed Blork, an anxious alien, through discussion of motivations and movement. The presentation explicitly identified later video examples as pregenerated.

The experiment treats personality as creative direction that should persist beyond a character’s face and voice.

That is a meaningful distinction for storytelling. Recognizing a character visually does not guarantee that the character behaves consistently. A familiar figure can still feel wrong when gestures, posture, or reactions conflict with the intended personality.

Conversation offers a way to develop those expectations gradually. A creator can respond to questions, clarify motivations, and revise a description after hearing how it sounds. The interaction resembles a workshop more than a single instruction submitted for generation.

Its value depends on what survives that workshop. A lively exchange is not enough if later scenes ignore the resulting decisions. The useful artifact is the direction that can be examined, changed, and applied to subsequent creative work.

An editable profile helps make that direction explicit. It gives the creator something to inspect rather than relying entirely on an opaque conversational history. If a trait is misunderstood, a visible description provides a place to correct it.

There is still a substantial gap between describing a personality and producing a convincing performance. People communicate through timing, restraint, hesitation, and context. A label such as “anxious” does not specify the appropriate behavior for every scene.

Overapplication is a foreseeable creative problem. A character should not express the same emotion at the same intensity throughout a story. Useful direction must allow variation while preserving an underlying identity. Consistency should support dramatic range rather than flatten it.

For an illustrative training scenario, a designer might want a fictional customer who begins a conversation frustrated but becomes more receptive. The character needs continuity across that change. Repeating a fixed mood would undermine the purpose of the exercise.

That example is a way to assess the concept, not evidence that Screen Test currently supports the assignment. The research preview does not establish performance across arbitrary scripts, languages, or production conditions. Those questions require broader access and independent testing.

Runway offers a relevant comparison through its character reference workflows. Its documentation describes preserving characters and environments using reference images, with iteration recommended for complex changes. Those references address visual continuity; Adobe’s experiment adds a conversational route to behavioral direction.

The comparison does not establish a winner. These are related problems approached through different controls, and no shared evaluation establishes their relative effectiveness. The useful question is whether a creator can preserve an approved decision as the surrounding scene changes.

For writers, the appeal is an externalized character brief that can be challenged and refined. For directors, it is a possible connection between discussing a performance and generating one. Neither requires treating a fictional character’s conversational responses as independent understanding or creative judgment.

Copying an edit raises the standard for creative automation

Adobe’s Project Rinse Repeat adapts demonstrated edits to other image regions using native Photoshop tools, layers, and masks. Its presentation includes transferring product shadows and group-photo retouching. Adobe says the resulting operations remain editable.

The important distinction is between repeating an operation and adapting the reason for that operation.

A conventional repeated action can be useful when inputs are sufficiently similar. More varied images demand judgment about placement, scale, and relevance. Applying an identical adjustment everywhere can create consistency in process while producing inconsistency in appearance.

An adaptive system attempts to close that gap. The creator establishes an example, and the software interprets how it should apply elsewhere. That shifts the burden from specifying every action to checking whether the interpretation was appropriate.

Editability makes the proposal more practical. A creator who receives adjustable operations has somewhere to intervene when part of the result is wrong. The alternative is accepting a finished image or attempting another generation without a clear route to local correction.

This matters in client work because approval is rarely final at the first presentation. A requested adjustment might affect only a small part of an otherwise acceptable result. Rebuilding the entire output introduces opportunities for unrelated details to change.

The useful promise is therefore continuity through revision. Automation should reduce the work needed to reach the next acceptable state. It should also preserve enough structure for a person to understand what happened and take over when necessary.

There is a corresponding risk in showing software an example. A demonstrated edit can contain both general intent and local exceptions. The system has to decide which elements belong to the reusable pattern. The creator may not recognize a mistaken interpretation immediately.

Suppose an editor adjusts a product photograph to compensate for unusual lighting. Applying that adjustment across a campaign might be appropriate for some images and harmful to others. The correct response depends on the target image, not merely the existence of a previous edit.

That is an illustrative evaluation problem rather than a reported failure. It explains why editable results are valuable even when automation performs well. The creator needs a way to correct exceptions without abandoning the entire workflow.

Scene Sweep extends the same concern into video. Adobe says the research explores carrying edits across shots while accounting for characters and locations. The practical stakes are different because a change must remain convincing through movement and cuts.

For a post-production team, local quality and sequence quality are distinct standards. An edited frame might look acceptable in isolation while drawing attention during playback. A useful system must support evaluation at the level where viewers experience the work.

Together, these experiments place pressure on creative software to preserve decisions through repetition. That is an editorial interpretation of the research direction, rather than a measured productivity finding. The announcements do not establish how much work teams save after review and correction.

Availability and independent testing remain separate questions

Adobe’s experimental tools playground lists Deep Brush, Hands On, and Custom Filters for exploration. It also lists conversational character direction as coming soon. Adobe identifies Deep Brush and Hands On as research previews from this year’s Sneaks.

Access to selected experiments does not establish general availability for the full research lineup.

As of November 17, readers should distinguish the project descriptions from tools they can evaluate directly. An announcement can explain an intended capability without establishing its release schedule, supported applications, or suitability for a commercial assignment.

That distinction affects purchasing and production planning. A team can follow a research project without depending on it for a deadline. Until access and operating conditions are clear, a demonstration belongs in an evaluation backlog rather than a committed delivery process.

The same caution applies to the word “interactive.” Different stages of a workflow can operate at different speeds. A responsive preview does not establish the time required for a finished output. Prepared presentation footage does not establish performance under unscripted conditions.

Screen Test’s disclosure of pregenerated examples is useful in that respect. It separates the conversation shown onstage from the time needed to produce the later clips. Readers should preserve that distinction when describing what the system does.

Sight Span raises a related distinction between exploring a scene and rendering a chosen view. Rinse Repeat raises another between creating editable operations and ensuring those operations are correct. Each project needs evaluation against its own task rather than a single general claim about AI capability.

A meaningful assessment would include difficult inputs, ambiguous instructions, and requests to revise only part of a result. Those conditions reveal whether the interface helps creators recover from mistakes. Recovery is a practical requirement, not an edge case to consider after adoption.

Teams should also distinguish technical feasibility from permission to use material. A workflow involving people, voices, or client assets still needs appropriate authorization. The existence of a creative control does not answer every question about the input or intended output.

The available sources do not provide a shared independent benchmark for these projects. They also do not establish broad customer adoption or production savings. Inventing those findings would turn a research story into an unsupported product endorsement.

That leaves room for a measured positive assessment. Adobe is investigating problems that arise after the first generated result, including continuity, direction, and revision. Those are concrete creative concerns, even while the proposed solutions remain at different stages of access and validation.

The useful skeptical position is to ask whether control survives ordinary work. Can a creator explain an unexpected result? Can someone else revise it? Can an approved decision persist through a changed brief? Those questions connect the research promise to the work it intends to support.

What to watch after Adobe MAX Sneaks

The next signals should concern access, portability, and repeatability, in that order. These are practical tests for the research direction rather than predictions of scheduled releases. They distinguish a promising interface from a dependable addition to creative work.

Progress will be clearest when more people can test the tools, retain their decisions, and reproduce acceptable results.

First, watch for changes in public access and documentation. A project moving into a usable preview would allow creators to test their own material. Clear limitations would make that access more informative by showing which assignments the team intends to support.

Expanded access would strengthen the case that Adobe is turning research into something users can evaluate. Continued restricted access would leave the central claims more dependent on selected examples. Neither outcome alone would establish whether the underlying concept is useful.

Second, watch what happens to the decisions created inside each interface. A chosen camera view, character profile, or adapted edit should remain understandable during revision. If that direction disappears into a finished output, part of the proposed advantage disappears with it.

Portable, adjustable decisions would strengthen the case for these interfaces as workflow tools. Results that require starting over after small changes would weaken it. The relevant standard is the creator’s ability to continue working, including after handing the project to someone else.

Third, watch independent evaluations that report corrections alongside successful outputs. Useful accounts should identify what the creator requested, what changed unexpectedly, and what work remained. A selected gallery provides less information than a documented attempt to finish a real assignment.

Repeatable performance across varied material would support the case for adoption. Heavy correction or unstable revisions would suggest a narrower role, perhaps during exploration rather than final production. That would still be useful, provided the limits are understood.

For creative teams, the sensible next step is to define an evaluation brief before an experiment becomes available. Choose a task with clear requirements and an existing way to complete it. Then compare the entire process, including review, revisions, and handoff.

Keep the original brief, approved references, and evaluation findings together in a personal knowledge base. Record which decisions survived and which required correction. That creates a reusable account of suitability rather than a collection of impressive outputs.

Adobe MAX Sneaks offers several compelling ways to communicate intent. Follow the projects through actual access, then test them against work that matters to you. Ask whether each tool preserves the decisions you approved, exposes mistakes clearly, and leaves a usable result for the next person. Compare that experience with your current process, including the time spent correcting exceptions. An explorable photograph or conversational character can make an effective presentation. The more consequential result is a creative workflow that remains understandable when the brief changes. Which decision would you trust the software to carry forward, and what evidence would you need before relying on it?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page