30-Second AI Video Clips: What Seedance, Wan and Kling Change for Architecture
30-second AI video clips from Seedance 2.5, Wan 3.0 and Kling 4.0 explained for architecture studios, with a decision table and a test protocol.
In this article
Three AI video models started producing 30-second AI video clips in a single generation between late July and late September 2026, which changes the basic unit a studio works with when it turns renderings into motion. Seedance 2.5, Wan 3.0 and Kling 4.0 all now generate one continuous clip roughly the length of a full social video, not a five-second fragment. This guide explains what changed, when a single long generation beats stitching several shorter clips together, where it adds risk, and how to test one on your own project before you commit to it.
Key takeaways
- Seedance 2.5 (July 31, 2026), Wan 3.0 (August 2026) and Kling 4.0 (announced September 28, 2026) all moved their maximum single-generation clip length to 30 seconds, up from roughly 5–15 seconds in each model's prior version.
- A single 30-second generation suits one continuous shot with one lighting state, such as a slow push through a lobby; it's a poor fit for a project that spans several rooms or times of day.
- Longer single-pass generations carry more geometry drift, since the model has to sustain straight lines and proportions over six times the duration of a five-second clip.
- Keyframes and multi-image references let you anchor specific renderings to specific moments in a long clip, rather than leaving the whole 30 seconds to one still image.
- Test any long-clip model on a short, cheap generation first, and watch the second half of the clip specifically, since drift tends to compound toward the end.
Three models, one change: native clips just got six times longer
Until mid-2026, most image-to-video models generated in five-second units; a 30-second film meant six separate clips edited together, each starting fresh from its own reference image. That ceiling moved three times in nine weeks.
ByteDance announced Seedance 2.5 on July 31, 2026: "up to 30 seconds per generation, with multi-round extensions," accepting "up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass." Its public API opened August 7, 2026 on Volcano Ark and BytePlus ModelArk; EvoLink's third-party testing confirmed 480p, 720p and 1080p output, and cautions against planning around 4K.
Alibaba followed with Wan 3.0, in public beta on Alibaba Cloud Model Studio and Qwen Cloud from August 10, 2026, with clips "up to 30 seconds, double the previous 15-second maximum" of the prior Wan model, since "the longer format allows for complex camera movements and continuous unbroken shots." Its formal launch on August 24, 2026 added resolution up to 1080p and per-second API pricing — $0.05 per second at 480p, $0.10 at 720p, $0.20 at 1080p — alongside a shift to closed model weights, a change from Alibaba's earlier open-source Wan releases.
Kuaishou's Kling followed on September 28, 2026, announcing that Kling 4.0 "doubles its maximum output to 30-second clips", up from 15 seconds, with "over a dozen reference formats" and up to 10 keyframes. As of this writing only the lighter Kling 4.0 Flash tier is in closed beta for annual subscribers, capped at 20 seconds and 720p; the full 30-second, up-to-4K model is scheduled for October 2026, per Morphic's model page. Check which tier you're testing before planning around its 30-second ceiling.
The pattern: longer native clips, more reference inputs per generation, and a cost or access tradeoff attached to the longer format. None of this is architecture-specific — these are general-purpose video models — but it changes the unit a studio plans a project film around.
When one 30-second generation beats stitching shorter clips
The appeal of a single long generation is continuity: one model, one pass, no seam where two separately generated clips meet. That's a real advantage in a specific situation — a continuous camera move through one space, in one lighting condition, where nothing in the shot needs to jump between rooms or times of day.
Stitching several shorter clips is still the better fit whenever a project film needs to cover ground a single continuous shot can't: several rooms, an exterior-to-interior sequence, or renderings made at different times of day that shouldn't be forced into one uninterrupted take. A six-image project — an approach shot, an entry sequence, two interior views and a material detail, say, for an adaptive-reuse warehouse — usually tells a better story as six distinct beats than as one long take trying to connect spaces the source images never actually show together.
| Situation | Better approach | Why |
|---|---|---|
| One continuous shot, one lighting state (a slow push through a courtyard garden) | Single 30-second generation | Nothing has to jump; one model sustains one consistent scene |
| Six renderings spanning several rooms or times of day | Several shorter clips, edited together | Each image anchors one beat; lighting and setting don't have to resolve inside one generation |
| A hillside house exterior needing camera movement across two facades | Two or three shorter segments, each keyed to one reference image | Long single-pass camera movement raises geometry drift; shorter anchored segments limit it |
| Interior clinic corridor with signage or fine railing detail | Short clips (5–10 seconds), keyframe-anchored | Fine detail is more likely to warp the longer a single generation runs |
Where longer single-pass clips get riskier
A five-second clip only has to hold a straight line and a proportion steady for five seconds. A 30-second clip has to hold the same things steady six times as long, with no drawings or measurements to check itself against — it's predicting plausible motion from a still image, the same way any image-to-video tool does; see what image-to-video AI can and can't do. Errors barely visible in a five-second clip, such as a mullion drifting a few pixels, have more time to compound into something visible by the back half of a 30-second shot.
This is the same geometry-drift risk covered in camera moves that keep buildings straight: static shots and slow push-ins carry the least risk, while pans, orbits and any move that reveals area outside the original frame carry the most. A longer generation doesn't change which moves are risky — it gives a risky move more time to go wrong, and one continuous 30-second orbit is a harder request than the same orbit broken into three shorter passes a person can check individually.
Cost adds a second constraint on the models with per-second pricing. At Wan 3.0's published rate, a 30-second clip at 1080p runs about six dollars in API calls; a shorter test clip at a lower resolution costs a fraction of that, one more reason to test before committing a full project to one long generation.
Using keyframes and reference images to anchor a long clip
Kling 4.0's up-to-10 keyframes and Seedance 2.5's up-to-30 image references exist for the same reason: a single still image gives a model very little to hold onto across 30 seconds, so both let you feed in more anchor points along the way, rather than free-associating everything between a starting image and an ending point. For a six-image project set, that maps onto six checkpoints spread across 30 seconds — an approach shot early, an entry a few seconds in, then interior views and a closing detail — each keyed to one rendering instead of one image doing all the work. The tradeoff is setup time: choosing and ordering reference images per checkpoint takes more planning than uploading six images and letting a tool sequence them for you. Choosing the six best images for a project video covers the same selection discipline these models now reward through keyframes.
Arch2Video's own project films are built the other way: its AI director studies six uploaded images together and plans six connected five-second scenes as one roughly 30-second story, deciding the order itself rather than relying on you to place reference images along a single generation's timeline. That's a deliberate design choice for studios who don't want to learn keyframe placement on a foundation model — see how Arch2Video's six-step process works. A director-planned six-scene film gives up the fine-grained control keyframes offer in exchange for not having to set that control up yourself; compare it against building your own pipeline if you're weighing which approach fits your studio.
A test protocol before you commit a project to one long generation
Run a small test before spending a full project's budget on one long single-pass clip.
- Generate a short clip first. Most of these APIs charge per second, so a five- or ten-second test at a lower resolution shows whether the model handles your materials and lighting before you pay for the full 30 seconds.
- Watch the back half specifically. Drift compounds over a shot's length, so a clean first five seconds doesn't guarantee a clean twenty-fifth. Play the full clip slowed down and check railings, window grids and any straight edge.
- Compare cost against your shorter-clip workflow. At Wan 3.0's rate, a 30-second clip at 1080p costs roughly six times a five-second clip at the same resolution; weigh that against the editing time saved by not stitching clips together.
- Keep a fallback plan. If a long single-pass generation breaks down partway through, fall back to two or three shorter, keyframe-anchored clips covering the same ground, the way you'd redo one bad scene rather than a whole film.
- Review before publishing, like any AI-generated clip. A longer clip still interprets your source images rather than measuring them, so someone who knows the project should watch it before it goes out. Arch2Video's example films are concept studies built from published project photography and renderings, reviewed scene by scene the same way.
Frequently asked questions
Do these 30-second models replace stitching shorter clips together?
No, not for most projects. They add a real option for one continuous shot in one lighting state, but a project spanning several rooms or times of day still tells a clearer story as several distinct scenes edited together than as one long take the model has to invent transitions inside of.
Is a 30-second single-pass clip more expensive than several 5-second clips?
It depends on the model and resolution. Wan 3.0's published per-second pricing means a 30-second clip at 1080p costs about six dollars in API calls, roughly six times a five-second clip at the same resolution; a finished-film subscription tool prices per project rather than per second, so the comparison isn't direct.
Can I use Kling 4.0's 30-second clips right now?
Only in a limited form as of late September 2026. Kling announced the 30-second capability on September 28, 2026, but only the lighter Kling 4.0 Flash tier — capped at 20 seconds and 720p — is in closed beta for annual subscribers; the full 30-second, higher-resolution model is scheduled for October 2026, so confirm which tier you're testing before planning a project around it.
What is geometry drift and why does it get worse over 30 seconds?
Geometry drift is when a straight line, proportion or material shifts gradually as an AI video model generates motion it has to invent, since it's predicting plausible movement from a still image rather than tracking real geometry. A 30-second clip asks the model to hold that same invented motion steady for six times as long as a five-second clip, so small errors have more time to accumulate before the clip ends.
Do I need to switch tools to use a 30-second generation?
Only if your current tool caps clips at five to ten seconds and you specifically need one continuous long shot. For most six-image architecture project films, a director-planned sequence of shorter scenes, like the one Arch2Video builds automatically, covers more ground in a project than a single long take can, since it can move between rooms and lighting states that one continuous generation isn't suited to.