Article

How ReTiNA uses recorded brainwaves to restyle films

For experimental filmmakers and curious viewers, ReTiNA shows how recorded brainwaves can steer AI movie effects—and what remains untested.

Editorial illustration for How ReTiNA uses recorded brainwaves to restyle films: a document represents the research briefing. Not documentary evidence.

ReTiNA-Cinema gives recorded brainwaves a role in movie effects: steering how AI restyles existing film frames. A preprint submitted on 19 November 2026 describes the approach for experimental filmmakers and interactive artists, with examples drawn from Frankenstein (1931) and Popeye the Sailor Meets Sindbad the Sailor (1936).

The important qualification arrives before the popcorn: the demonstrations used prerecorded EEG—electrical activity measured through electrodes on the scalp. The authors conducted no user study and did not evaluate a live viewer–system feedback loop. This is a creative research setup, not demonstrated audience control.

The central argument of this analysis is simple: a brain signal need not describe a picture to influence its appearance. The creative opportunity lies in designing the signal-to-style relationship, rather than extracting a film from someone’s imagination.

What changes in the movie?

ReTiNA changes the visual treatment of existing footage. The authors describe transformations of film frames, not explicit edits to plot, dialogue or scene order. A different appearance might affect how a scene feels, but that is a possibility to investigate, not an audience response established by this work.

That is a different creative task from turning film situations into playable actions. A branching story asks what happens next; ReTiNA asks how the next frame looks.

The idea also has a history. The author’s project page records a presentation at the NeurIPS Workshop on Creativity and Generative AI in 2024. The November preprint is an expanded research account, including two-stream control and scene-aware prompt generation—not the debut of a consumer cinema app.

How brainwaves choose a visual style

Three ingredients do different jobs: the film provides the scene, text prompts describe possible looks, and measurements from the EEG steer the visual treatment. The brain signal is a control input, not a hidden screenplay.

In the single-stream design, the software calculates Lempel–Ziv complexity, a measure of variation in the recorded signal. It compares that value with running statistics from the signal. Values above the resulting baseline select one prompt; values below it select another. At zero, the original frame appears. A greater distance from zero increases the transformation strength.

Think of it as a style selector joined to an intensity dial. That analogy describes the mapping, not a person deliberately turning the dial. A higher complexity value is not, by itself, a declaration that the viewer is excited or wants a particular effect.

Nor must the available styles remain fixed. The paper describes both manually written prompts and scene-aware prompts. In the latter mode, a visual AI model examines a sampled movie frame and proposes four styles. Sampling defaults to every 30 seconds and is adjustable. The single-stream mode uses two of those styles; the two-stream mode can use all four.

This is distinct from how GALA matches brainwaves to picture choices. GALA tries to identify a picture someone was shown. ReTiNA uses a numerical property of the signal to steer a transformation whose visual possibilities come from prompts and the source film.

There is a useful design consequence: the image can change because the signal changes, because the available prompts change, or both. For a later experiment about perceived control, keeping the prompts fixed—or recording when they change—would help separate those influences. That is proposed evaluation guidance, not a result demonstrated here.

What a second brainwave stream adds

The two-stream configuration combines the two complexity values to choose among four prompts. A separate statistic describing coupling between the signals controls effect intensity. This creates another way to connect measurements to aesthetics; it does not establish that two people share an emotion.

The authors explicitly treat these quantities as computational controls, not validated measures of engagement or shared cognition. The repository’s default setup also replays the same sample recording into both inputs. Live input can be configured, but two input channels are not evidence that two viewers have jointly directed a film.

Fast effects do not establish intentional control

The authors report roughly 15 generated frames per second on hardware including an RTX 4090 laptop GPU and an RTX 3090 desktop GPU. That is a reported technical performance result, not a guarantee for other equipment or evidence that viewers can intentionally regulate the effect.

The signal’s timescale matters too. The 30-times-per-second feature updates use heavily overlapping two-second windows. Refreshing a measurement quickly does not turn it into 30 independent readings of a person’s feelings each second.

The open questions are therefore experiential as well as technical. Can a viewer learn to influence the styling? Does the connection feel meaningful? Does it improve enjoyment or distract from the film? The paper leaves these questions for future user studies; its generated frames cannot answer them.

What creators can use now

The public project files offer a starting point for an experimental installation, with integration work attached. The documented setup requires TouchDesigner, the goofi signal-processing software and a separately obtained StreamDiffusion node. Scene-aware prompting adds a visual-language-model dependency; custom prompts are another option.

It is not an established all-in-free package. TouchDesigner has a no-charge noncommercial edition, but that alone establishes neither compatibility with this entire project nor its total cost. The separate node’s current access price and the complete setup and running costs remain unverified.

The repository licence is CC BY-NC 4.0. Its stated permissions are noncommercial and carry attribution conditions. Public files do not establish permission for a particular paid exhibition, its source footage or every external component.

For a creator, the useful next decision is artistic: choose what should change in the picture, which signal measurement should steer it, and why that relationship belongs in the work. Then assess whether an audience experiences the intended connection. ReTiNA offers a research route to signal-driven film styling; the viewer’s sense of participation is still a separate question.