AI video generation just became free — what that changes for film-makers

In a single day, three companies film-makers rely on stopped treating video generation as the expensive part. Google now gives high-definition generation away, Runway moved inside the edit suite so a clip can be restyled without leaving the timeline, and YouTube showed an edit held as a conversation. The generation is free or nearly free; the judgment about what to generate is not. Singapore is already a useful test case, because its AI-made work has been judged on story, symbols and performance rather than on how the frames were produced.

On 23 September 2026 Google opened Gemini Omni 1.1 in Google Vids to anyone with a Google or Workspace account at no cost. It generates scenes in 1080p, lets you set exact clip durations so a shot lines up with a voiceover, and extends a scene while holding the lighting, the characters' appearance and the environment consistent. Every clip carries a SynthID watermark. The same day, Runway shipped a plugin for DaVinci Resolve, and YouTube showed conversational editing: instruct the edit in plain language and step back into the timeline when you want control. YouTube promises it for Shorts and the Create app in early 2027.

The Runway plugin says more about the industry than any model release. It works on the clip under your playhead: set in and out points, restyle the anchor frames or set up to five keyframes, and the model re-renders the clip at the same duration as the source. Before and after sit side by side, and the cost of each generation is shown before you commit. The plugin is a free download, needs DaVinci Resolve Studio 19 or later, and generating runs on paid Runway plans using credits you already have.

Until now it meant exporting the clip, finding a browser tab, uploading, waiting, downloading and re-importing. Free generation plus a panel inside the edit suite says where the value now sits: not in making frames, but in deciding which ones are right.

Singapore is further into this than most markets. Its first AI-hybrid drama, Crooks, has actors performing on live-action sets while AI builds the environment around them; more than 50 crew across production, technical and editing roles are being retrained for it, and showrunner Boi Kwong says the work has shifted towards post-production. Two AI-made local films released over the National Day weekend drew criticism instead of applause, one of them for inaccuracies including a Singapore flag with six stars.

The sentiment data is more interesting than the headlines. Analysis by media intelligence firm CARMA of 271 tracked conversations around AI found 54 per cent positive, 37 per cent neutral and only 9 per cent negative. But 92.2 per cent of the negative conversation referenced the National Day Parade, and the complaints were about visible errors, poor creative quality and a perceived lack of human oversight. Nobody was objecting to the generation.

Google's own AI and Economy ATLAS data, published on 15 September 2026, points the same way. Outside the wealthiest economies, arts, design, entertainment, sports and media occupations are among the heaviest users of AI at work, and in India that group accounts for 19 per cent of work-related AI usage, 1.6 times the global average. Generating is no longer a specialist act, which is why the shot list matters more than the model: what each shot is for, how long it lasts, what must not change inside it.

Four habits matter more now than they did with paid tools. Write the shot list before generating anything. Lock the look once as a single approved reference and attach that same reference to every shot rather than describing it again in words. Approve a still before you animate it, because a still costs a fraction of a clip and it is the last point at which a wrong face is free to fix. State durations and the things that must not change — lighting, wardrobe, environment — instead of re-describing them and hoping.

Free generation is good news, and we will use it — mostly for the first pass, the shot we are not sure about, the idea we want to see before spending anything on it. What we would not conclude is that any of this changes the craft. It moves a cost, and the decisions that make a film hold together still have to be made by somebody, faster.

We are less convinced than the coverage that a conversational edit changes how serious work gets cut. A conversation is an excellent way to try an idea and a poor way to hold a rhythm; the reasons a cut works are things you compare, not things you describe. The Singapore evidence is what we would hand to anyone planning around these tools: audiences are not rejecting AI. They are rejecting work nobody checked.