AI video changed more in the last six months than in the two years before it. A model that led every comparison in spring has dropped out of the conversation by autumn. Clips that used to stop at five seconds now run to thirty, with sound. And the audiences watching the results have become sharper, not softer.

We make AI films for a living, for brands, streaming platforms and series. This guide is what we have learned from that work, set against what the rest of the industry is saying right now. It is long on purpose. If you commission, plan or approve AI video, it should save you a few expensive mistakes.

The short version: the model matters less than you think, the process matters more than you think, and the people directing it matter most.

The year the leaderboard stopped mattering

Start with how fast the ground moved. As of October 2026, public rankings put Alibaba’s Wan 3.0 and ByteDance’s Seedance 2.5 at the top. Google has made Gemini Omni Flash its default video model. MiniMax and Kandinsky have released strong open-weight models, with Kandinsky 6.0 published under an MIT licence this month. Kuaishou’s Kling 4.0 reached general launch in October with native 30-second clips, 4K output, 10-bit HDR and stereo audio.

Then look at who is missing. OpenAI closed the Sora app in April and switched off the Sora API on 24 September. Anyone who had built a process around it had to start again.

The lesson is not that one company failed. It is that any workflow tied to a single model has a shelf life measured in months. Studios that handle this well treat models like lenses: you choose the one that suits the shot, and you expect the set to change.

Five shifts brands should know about

Sound arrives with the picture. The newest models generate dialogue, ambience and music in the same pass as the video. That helps first drafts feel alive. It does not remove the need for a sound designer, because a clip that sounds fine on its own rarely sounds right inside a film.

Clips are getting longer. Thirty seconds in one generation is now possible. Longer clips are useful, but they also give a model more time to drift, so the review burden grows with the length.

Open weights are arriving. Strong open-weight video models are now available, which changes the options on cost, privacy and control for anyone willing to run them.

Platforms are automating ad creative. Tools from the large ad platforms can already generate simple ads from a prompt, and small advertisers are using them. That raises the floor. It also means a studio cannot compete on generation alone.

Audiences are watching closely. We cover the data below, but the headline is that viewers notice synthetic work, and the gap between how the industry feels about AI ads and how audiences feel is real.

What actually decides whether an AI film works

If the model is not the deciding factor, what is? Across our projects, five things separate a film that holds together from a clip that merely looks impressive.

1. Consistency is settled before anything is generated

The most common failure in AI video is drift. The hero is a little older in the wide shot, the jacket changes after a cut, the courtyard has different windows in episode three. Viewers rarely name the problem, but they feel it, and the story stops feeling real.

The fix is old filmmaking discipline. We build a character sheet for every main character, with turnaround views, expressions, a colour palette and approved costume looks, and an environment sheet for every location. The client approves them as images before any video exists. After that, nobody invents a face or a room again. Everyone works from the approved one.

For a long series the payoff is large. One character in our period series Veer Tejaji has four approved looks, from the wedding day to the final episode, and all 40 episodes draw from the same sheet.

Environments get the same care. We build an environment sheet for each location, so a courtyard looks like the same courtyard in episode 3 and episode 30.

2. Control comes from planning the shot, not prompting it

A prompt describes a shot. A shot plan decides it. We break every script into a numbered list of shots, each with a camera angle, the characters and location in it, and a length in seconds. We also grade each shot by difficulty so schedules stay honest. A close dialogue shot with several characters takes longer than a simple insert, and the plan says so.

Every shot then begins as an approved still keyframe and is animated only after that. For the hardest sequences, such as destruction or sweeping camera moves, we lock the layout first through a previs layer and generate the final imagery against it. Because the camera and the action are already decided, far fewer generation passes are needed. That is also why we can quote a fixed number of iterations.

Our From Previs to Final Pixel project shows one shot travelling from the previs layer to the finished image.

3. Hybrid is the answer when the story needs a real face

Some stories need a real performance. Others need a world that exists nowhere. Hybrid work cuts the two together so the audience experiences a single film. On Bayer’s Arize, Khurafati Studios led the scenes shot with real actors and locations while we built the other half with AI-generated cinematic visuals. For Nihar Naturals, we blended live-action footage with generative AI to make a regional campaign that went beyond what a conventional shoot allows.

The seam disappears when both halves share a colour palette, lighting references, camera language and one storyboard, and when everything is graded together at the end. Decide where the film moves between live action and AI before anyone shoots.

4. Finishing is what makes it a film

Raw generations look like raw generations. We edit in standard tools such as Premiere Pro and After Effects, grade in DaVinci Resolve and finish the sound. We add a light analogue grain and a cleaned-up image so the film feels photographic, while keeping brand colours and logos exact. Then we upscale for the screens it will run on and cut the versions a campaign needs: 16:9, 9:16 and 1:1, with subtitles and regional voice-over.

Sound deserves its own mention. Characters keep the same voice across a series, music is written or sourced to suit the film’s mood, and regional films are written and voiced natively rather than translated afterwards.

5. A creative director reviews every shot

The review questions are the ones a continuity supervisor asks. Does the face match in the wide shot and the close-up? Is this the right costume for this scene? Is the light on the face the same as in the shot before? Are the buildings and props in proportion? A shot that does not match goes back before the client sees it.

Clients approve at four points: the script, the character and look frames, the keyframes or animatic, and the first cut. Because the biggest creative choices are made in the first three, the first cut is rarely a surprise.

Choose models per shot, not per studio

We use the leading video models and pick the right one for each shot. That sounds like a detail. It is the most important structural choice in the pipeline, because it is what lets the work survive a month like this one.

Our rule is to judge a model by the shot it has to deliver. A close dialogue shot, a product close-up, a wide environment and a stylised dream sequence each stress a model differently, and the best choice for one is often not the best choice for another. Rankings are a fair starting point. They are measured on generic prompts, not on your product, your character or your language.

Two practical points follow for a brand. First, ask any studio what happens to your film if a model it uses is withdrawn mid-project. A pipeline built on approved frames, sheets and shot plans keeps working, because those assets belong to the project and not to the model. Second, be wary of a studio that sells you a model name rather than a result.

Where AI video still breaks

Honest limits are part of expertise, so here are ours. Recent benchmarks agree with what we see in production.

Text inside the picture is unreliable. One benchmark published this month found that even the best video model had a word error rate of 0.25 when rendering on-screen text, which is roughly one word in four wrong. Our advice is to add titles, pack details and logos in the edit rather than generate them.

Precise, multi-step action still slips. A separate benchmark of goal-directed manipulation tasks found dropped steps and physics breaks. Anything where a hand must do something exact, in order, needs a shot plan and often several attempts, or a different approach such as live footage.

Long shots drift. Temporal stability over long durations remains a known weak spot, which is why consistency has to be enforced by process rather than hoped for from the model.

Rights and likeness need care. Know which tools are used, whether any real person’s face or voice is involved and with what consent, who owns the final film, and where it can run.

The trust gap is real, and it is a craft problem

Audiences are not rejecting the technology so much as the lack of care that often comes with it. A June 2026 Harris Poll, run with the 4As and Infillion, found that 78% of consumers say AI makes ads feel less authentic and 73% say they are less likely to trust an ad they suspect was AI-made. Research by the IAB and Sonata between October 2025 and January 2026 found that only 45% of Gen Z and Millennial consumers feel positive about AI-generated ads, while 82% of ad executives believed those consumers felt positive. That gap between the industry and the audience is where weak work gets made.

High-profile misfires have followed. McDonald’s removed an AI-generated Christmas ad after online backlash, and Coca-Cola’s AI holiday films have drawn repeated criticism.

What separates work that lands from work that gets mocked is rarely the model. In our experience it comes down to five things: start from one human insight, lock the look before you generate, make the film natively in the audience’s language, treat sound as half the film, and be clear about rights and claims. We go through each of them in our post on avoiding the AI slop trap.

What this looks like in practice

Theory is cheap, so here is what the process has produced.

For the IPL 2025 auction, Rajasthan Royals made their first AI-generated campaign film with us. It earned 17 million organic views with no paid media behind it. Our read is that it arrived when fans were already looking, gave them something to share, and carried the team’s own character. That is our view, not a measured study.

For Instamart’s Switch campaign, we made six GenAI films in the second quarter of 2026. The board below shows the approved keyframes for one of them. Every shot started as one of these frames, and nothing was animated until it was signed off.

For Marico’s Parachute Advansed, we made a launch film for Tamil Nadu television, in Tamil, to a fixed air date. It aired in August 2026. Every face and room was approved before any shot was made, which is why the film held together under a deadline with no second chance.

In series, our work for STAGE on titles such as Veer Tejaji Maharaj shows the same method holding across a long run of episodes.

Eight questions to ask before you commission

Can I see released work for brands like mine? Ask for films that have run, not tests, in your category and your languages.

What do I approve, and when? You should see the script, the look frames and an animatic before the final film is generated. Changes are cheap then and expensive later.

How do you keep characters and products identical from shot to shot? Ask to see a character sheet or a product reference, and how it is enforced.

Who writes and directs, and can I meet them? The tools are open to everyone, so the people using them are the difference.

Which models do you use, and what happens if one is withdrawn? The honest answer involves shot-by-shot choice and project-owned assets.

What is generated and what is added in the edit? Logos, packs and on-screen text should be exact.

How are rights, likeness and claims handled? Ask for the answer in writing.

What do I get at the end? Formats, languages, subtitles and the source assets you will need for the next film.

What we expect next

Expect clips to keep getting longer, sound to keep improving and costs to keep falling. Those are real gains, and we use them. But none of them answers the question a brand actually has, which is whether the film feels like it belongs to them. That depends on a story worth telling, a look that stays fixed and someone with taste checking every frame. The tools are available to everyone. The direction is what is scarce.

If you are planning a film, a launch or a series and want to see how this would run on your brief, write to namaste@teevra.co.