Web developer John Polacek released “Deviant” — an 84-minute sci-fi thriller in which the entire visual track is AI-generated based on the author's own script. This is one of the first feature-length films made entirely from prompts to video models without a film crew, and the script and full list of tools are open for verification.

image
image

What happened

The film tells the story of a clone of an aging tech magnate who finds an exploit in his own security code and takes control of the company. The feature-length version appeared on YouTube on September 5, 2026. Production used character and location references created in Midjourney and Nano Banana, the majority of shots were generated in Google Veo, the final act was made in Seedance 2.5 via Higgsfield, editing was done in iMovie, and sound in ElevenLabs. The script is published in PDF format under the MIT license along with the project website, and on September 12 the author presented the work on Hacker News.

Context

“Deviant” is not a scientific result or a benchmark, but a demonstration that one person can close the full cycle from script to 84 minutes of finished film, relying only on commercial APIs. For “made by AI” projects, this is rare: the author opened the script, the full list of tools, and the project itself, so the claimed pipeline can be verified and in principle replicated. Additional context is provided by the film's structure itself: the transition from Google Veo to Seedance 2.5 is visible on screen, and the author directly notes that the final act “looks noticeably better” than the first 60 minutes, which clearly shows the gap between generations of video generators.

Why this matters for the industry

For the industry, this is a real benchmark: feature-length content has become available to a single developer through commercial APIs, and the open pipeline is a ready-made reference framework for teams planning generative video pilots. The case simultaneously identifies bottlenecks that remain the main problems of the current generation of models: character consistency, multi-character scenes, voice matching. The signal that Veo stopped reliably generating scenes and it was necessary to switch to Seedance 2.5 indicates the instability of generation via API and pushes toward designing production pipelines with model orchestration, fallback, versioning, and automatic quality control.

Why this matters for users

The film can be watched for free and in full on YouTube, and the script can be downloaded in PDF under the MIT license. For the reader, this is not just content, but a documented working pipeline of one person: from references in Midjourney and Nano Banana to generation in Google Veo and Seedance 2.5, editing in iMovie, and sound in ElevenLabs — along with an honest list of tasks that still have to be solved manually.

What is still unknown / limitations

This is a practical demonstration, not a controlled comparison of models: the first 60 minutes were made in Google Veo, the last 20 in Seedance 2.5, the scenes are different, and the quality assessment relies on the author's subjective judgment. Accurately replicating the result is hindered by dependence on third-party commercial APIs: the author had to switch generators when Veo stopped generating scenes reliably. Finally, the author himself names the unresolved limitations of the current generation of tools: acting, frame composition, character consistency, voice matching, and coordinated scenes with more than two characters in one frame.

Sources

Author

Look at AI, editorial team