UNREEL makes the movie while you watch

UNREEL uses Gemini 2.5 Flash and H3 Max Turbo to write and generate a sequence of shots during playback, without producing the complete movie in advance.

The movie does not yet exist when the viewer presses play. UNREEL begins by showing a short introduction prepared in advance, then progressively writes and generates what follows to keep the stream running for as long as the viewer wants.

Presented by Blendi as the “world’s first fully AI-generated video streaming service,” the project borrows the visual conventions of a Netflix-style catalog. Its home page includes posters, descriptions, previews, and several genres. Behind that interface, however, it does not store a library of finished feature-length movies.

The published version includes 14 stories spanning science fiction, westerns, cyber-noir, folk horror, romance, and boxing, among other genres. Eight additional channels use a deliberately more chaotic format, with concepts such as Capybara News Network, Goblin Tax Season, Toaster Court, and Dad Rock Volcano.

The posters, hover animations, and opening sequences are produced when the catalog is built. The content shown after that introduction is generated as the viewing session unfolds. This distinction qualifies the claim that nothing exists before the viewer presses play: the presentation and beginning of each program are already available, while the rest is improvised.

UNREEL divides the work among several models. Gemini 2.5 Flash acts as the showrunner and writes the upcoming shots. Nano Banana 2 produces the posters and their typography. H3 Max Turbo then turns the instructions into short video sequences through fal.

H3 Max Turbo is not the standard MiniMax H3 model. It is a version derived from MiniMax H3, post-trained by fal and adapted to its inference infrastructure. The official documentation describes it as a variant designed for higher throughput, stronger instruction following, and improved visual quality.

The model does not generate an entire movie in a single operation. UNREEL asks it to produce a succession of sequences lasting a few seconds, then organizes, buffers, and plays them in the correct order. The impression of continuity comes primarily from this orchestration.

Once a program begins, the showrunner prepares one shot, followed by groups of four and then ten. The first shot is generated from the final image of the introduction. While it is playing, the next one is produced in the background. The buffer is intended to build gradually so the viewer does not have to wait between sequences.

For story-based programs, each new shot uses the final image of the previous one as its starting image. This technique is intended to preserve some of the character, lighting, and setting. It does not, however, provide persistent video memory spanning the entire movie. H3 Max Turbo receives a new task at every stage and may still unintentionally alter a face, an item of clothing, an object, or the layout of a location.

The showrunner reintroduces the synopsis, visual direction, and narrative elements in its instructions to limit these shifts. Continuity therefore depends on the image being passed forward, the written description, and the model’s ability to interpret those references consistently.

The so-called “chaos” channels abandon this attempt at seamless continuity. They generate two five-second sequences at a time and deliberately move between situations using hard cuts. Inconsistencies become part of the format rather than flaws that need to be hidden.

Dialogue also requires a workaround. The project’s creator explains that H3 Max Turbo produces more intelligible speech when the exact words are placed in quotation marks within the instruction. Gemini therefore writes every line verbatim, while the program adds matching audio directions and displays the dialogue as captions.

The method was reportedly verified with Whisper, although no detailed results have been provided. It does not guarantee that a voice will retain the same tone between shots, that a conversation will maintain a natural rhythm, or that music will continue without interruption. Each sequence has its own generated audio, making sound continuity more difficult than linking the images.

The speed of H3 Max Turbo makes this system possible. The UNREEL repository states that an eight-second sequence at 480p takes approximately 2.5 seconds to process. At 768p, the delay reportedly increases to nearly five seconds. The project therefore uses 480p by default to preserve enough time for the final-image handoff.

These measurements reflect the developer’s observations rather than a guaranteed processing time. fal’s queue, network conditions, instruction preparation, and file retrieval can all slow down the stream. If the next shot is not ready in time, the player holds the final image on screen. If that image cannot be retrieved, the system can use a hard cut to avoid stopping the session entirely.

UNREEL also submits two identical requests for the showrunner’s first intervention and keeps the first valid response. This redundancy reduces the risk that a particularly slow request will delay the beginning of the generation process, but it may slightly increase billed usage.

The promise of watching an “entire movie” should therefore be understood as the ability to extend the stream, rather than the existence of a complete feature with a finalized screenplay. The story is written in blocks, with no required duration or necessarily prepared ending. Closing the browser tab stops production.

The current version also lacks the features normally associated with a platform for completed movies: a final edit, a download of the entire program, the ability to resume the exact same session, or a fixed version that remains identical for every viewer. Two people selecting the same title may receive different story developments.

The option to “create your own movie” remains primarily aimed at developers. Users must download the code, install the project, provide a fal API key, and add a title to a JavaScript file. Each entry must include a synopsis, a premise for the showrunner, visual