Two-Person Team Spends 12 Days Making a Six-Minute AI Short Resembling 'Dune 3' With Seedance 2.5 and Codex

The team divided the work into four stages: polishing the script, building an asset library, generating video, and post-production. In traditional film terms, that meant writing a script, casting actors, choosing costumes, building sets, shooting, and editing. The tools changed, but the team said the production logic remained similar.

Script work took one day. The team studied the original novel, films, trailers, and public setting materials, discussed a story framework, and gave the framework and materials to Codex to expand into a standard screenplay with plot and dialogue. They then manually checked each scene for character actions, dialogue, and emotional changes. The account said a useful technique was to ask the AI not only what happens but also how characters do it, including the order of actions, subtle facial changes, tone, and spatial relationships. The goal was not literary polish but clearer execution.

The asset library was the most critical step, covering character appearance, costumes, props, and core scene settings. It determined the film's aesthetic and whether a character would suddenly change in the next shot. The team encountered a copyright-related review issue after uploading movie screenshots as references. The initial results were close to the film's visual tone, but the AI video platform triggered copyright protection review and blocked further generation. The team rebuilt the assets, keeping abstract visual features such as low-saturation color, desert texture, and spatial scale while remaking character silhouettes, costume structures, and prop details. They removed highly recognizable elements such as the Fremen blue eyes and the stillsuit filter plugs. Generating image assets took two days, followed by another day of centralized adjustments. In their multiple revisions, images generated by the web version of GPT became increasingly blurry, so later adjustments were mainly done in Codex.

Video generation used Seedance 2.5, which the team had previously tested in an internal beta on Xiaoyunque. The operation itself was simple: upload the relevant character, scene, and prop materials into a dialogue box, attach prompts, and @ the materials to be referenced. The gap in results came from prompts. With Codex and a dedicated Skill, the team prepared prompts for each video segment. The prompts had three layers: an overall summary of what happens, the cinematography style, and characters' emotions; a time-line segment explaining where characters stand, what they do, how the camera moves, and how performance rhythm changes; and constraints on what must not appear, what actions should not happen, and what must remain consistent.

Seedance 2.5 can generate a 30-second video at once, but longer duration makes prompts less forgiving. If camera movement, blocking, or performance rhythm is unclear, characters may suddenly change positions, actions may reverse order, and restrained performances may become exaggerated. The team therefore split videos into second-by-second timelines and specified character positions, expressions, body movements, and dialogue tone. They added negative prompts to tell the AI what not to generate. The account said this improved control and reduced repeated generation, saving credits. For larger material sets, they tried Xiaoyunque's free canvas, which places characters, scenes, props, and generated results together and lets users extend from selected key clips. The team still relied mainly on Agent dialogue and used the canvas as an auxiliary tool. Agent was better for revising prompts and follow-up questions, while the canvas was better for managing materials and organizing many shots. For beginners, the account suggested starting with Agent mode's canvas switch and moving to a new free canvas when more characters and shots are needed.

For flaws, the team used Seedance 2.5's local editing. In one shot, they wanted to remove an extra spaceship. Text descriptions alone might not make clear which ship, where it was, or whether other content should change. By framing the area and adding a revision request, the model limited the task to the selected location and produced more stable results than text-only instructions.

Post-production remained human-led. AI short films are rarely generated in one pass; they are a process of generation, selection, and recombination. A 30-second video might contain only a few good seconds, but those seconds can become an effective shot. The team kept every generated result, moved them into editing software, selected the best clips, adjusted rhythm, and recombined them. They then added music, sound effects, and subtitles. The generation model supplied material; human editing decided whether it became a film.

The team's two members were both first-time AI short film makers. Twelve days and more than 3,000 yuan in credits produced a six-minute work. The account also cited Hell Grind, a 95-minute AI-generated feature that screened at a Cannes Film Festival event. That project was made by a 15-person team, used 14 days for generation after extensive asset preparation, and had a total budget below $500,000. Its team later opened the full project, including prompts, assets, and generation records for each scene. The ifanr team said its scale was far smaller, but after studying that project, it found the underlying workflow was not essentially different: prepare scripts, characters, and scenes as reusable assets, choose tools to generate shots, and use human selection and editing to turn material into a work. The team's main takeaway was that tools can change, but the production logic is shared. Rather than chasing every new tool, creators should spend more time on scripts, asset management, shot scheduling, and post-production judgment.

Editor's Summary The account describes a 12-day, two-person effort to make a six-minute AI short film with Codex and Seedance 2.5, including script development, asset building, second-by-second prompting, local edits, and human editing. It also notes copyright review risks when using film references and compares the workflow with Hell Grind, a 95-minute AI-generated feature made by a 15-person team. The core message is that AI video tools still require conventional filmmaking discipline in scripts, assets, shot control, and post-production.