Generative AI: The 2026 Field Guide

The third frontier is video length. As of 2026, the longest coherent clips that frontier video models can produce are around 25 seconds (Sora 2 Pro), 15 seconds (Kling 3.0 cinematic mode), and 8 seconds (Veo 3.1, Wan 2.5, Seedance 2.0). Beyond those lengths, the models lose temporal coherenceTemporal coherenceThe property of video where elements remain consistent across frames. The central challenge of video generation., characters drift, objects appear and disappear, the scene morphs in ways the promptPromptThe text description you provide to a model to specify what you want it to generate. did not request. The technique that produces a coherent 8-second clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. does not generalize to a coherent 8-minute clip without something fundamentally new.

ANALOGY

Writing a tweet and writing a novel feel like the same task (you are putting words in order) but they are fundamentally different. A tweet has no scenes, no chapters, no character arcs, no foreshadowing. A novel has all of those, and the techniques you need to write a good tweet are useless for keeping a novel coherent over 300 pages. The same is true for video generation. An 8-second clip is a tweet, you can hold the entire thing in working memory and the generator never has to remember what happened ten seconds ago because there is no ten-seconds-ago. An 8-minute clip is a novel, the model has to remember that the character in scene three was wearing a green coat, and that the door in scene one was closed, and that the lighting was dusk, and bring all of that forward when generating scene seven. Today's video models do not have the architectural equivalent of a novelist's outline, which is why scaling current approaches will not get us to long-form. Something new has to be invented.

There are several research threads attempting to break this barrier. The first is hierarchical generation: produce a low-resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. sketch of the entire long video first, then progressively refine it into the full video. This is the approach used by some long-form video research at Google and Meta, and it gets to 60-second clips with reasonable coherence but the quality is meaningfully worse than the short-form frontier models.

The second is autoregressiveAutoregressiveGenerating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion. video generation, where each new clip segment is conditioned on the previous segment. This is computationally cheaper than generating the whole video at once and lets you produce arbitrarily long videos in principle. The problem is error accumulation, small inconsistencies between segments compound over time, and after a few minutes the video has drifted away from the original premise. Recent work like StreamingT2V and CogVideoX-Flash addresses this with sliding-window attention and explicit memory mechanisms, but the results are still worse than short-form generation.

Fig.diagram

■Why long video drifts. Autoregressive generation conditions each new segment on the last, which is cheap, but small errors compound segment over segment until faces morph and color creeps. Holding coherence across minutes, not seconds, is the unsolved problem.

The third is the one I think is most likely to win, which is to treat long-form video as a fundamentally different problem from short-form video and build dedicated architectures for it. The key insight is that long-form video has a hierarchical structure that short-form does not, there are scenes, shots, and beats, and the right level of compression is different at each level. A research direction called 'scene-aware video generation' is trying to encode this structure explicitly, generating videos as a tree of scenes and shots rather than as a flat sequence of frames. This is closer to how a human filmmaker would think about a long video, and it may be the architectural insight that breaks the length barrier.

The practical timeline for usable long-form video generation is probably 18 to 36 months. Sora 1's preview release in February 2024 already showed clips up to a minute long with rough coherence; the difference between that and a usable production tool is mostly polish and reliability. By 2027, generating coherent two-to-three-minute clips with named characters and consistent settings should be feasible. By 2028, longer narrative pieces. By 2029 or 2030, the question of 'can AI generate a feature film' will have been answered (probably with: 'yes, but the quality bar for theatrical release will still require significant human curation').

Long-form video and the temporal coherence problem

Get the week's moves in your inbox.