Text is the one input that needs nothing uploaded. Sixteen of the twenty-one active video families on this site accept a prompt on its own, and the five that do not are named further down together with what they accept instead. Duration, resolution and aspect ratio are set family by family, because the families do not share one set of limits.
16 of 21 active families accept text alone · Four duration groupings · An audio toggle on three families
Landing pages for this task usually quote a single clip length, most often “up to 30 seconds” or a fixed eight. The families in this catalogue do not share one value, so the four groupings are listed here rather than averaged into a friendlier number. Every count below is read from this site’s own video model catalogue, family by family.
The largest group. The duration control opens at 5 seconds and stops at 15, moving one second at a time. Most of the widely used families sit here, including kling-3.0-turbo, wan3.0-video and the two viduq3 tiers that accept text.
Families: 10 families
The same ceiling with a shorter starting point. The distinction matters when a hook has to land inside a two-second pre-roll and every further second is budget spent.
Families: 4 families
The longest window in the catalogue and the only family that reaches 30 seconds. It also carries the widest aspect ratio list, seven ratios rather than the usual five.
Families: seedance-2.5
One family does not accept an arbitrary length at all. Three fixed choices are offered instead, which removes duration as a variable rather than making it a setting.
Families: veo-3.1
Text to video is the branch of video generation that starts with nothing but words. There is no frame to pin and no reference set to assemble, which is what separates it from image to video and reference to video. xAI documents the relationship directly: on its text-to-video path the model generates a first frame from the prompt and then animates it, so the two modes share one pipeline underneath and the intermediate image is never returned to you. The practical consequence is that a text prompt spends its first pass deciding what the scene is and only its second deciding how it moves. More than one vendor publishes text to video as its own endpoint with its own model list rather than as a modifier on image to video, which is a fair signal that the task is worth pricing separately.
This site routes text to video across 16 of the 21 active video families in its catalogue. The five that do not accept text are named rather than hidden: hailuo-3, which takes a first frame and reference images but no prompt alone; midjourney-video, which takes a frame and a closing frame; viduq3, whose four modes are all disabled; viduq3-mix, in the same state; and viduq3-pro-fast, which accepts a frame only. Among the 16 that do accept text, limits are not shared either. Duration falls into four groups. Aspect ratio lists run to five ratios on eleven families, seven on three and three on two. Resolution pools split into four combinations, and only three families expose an audio toggle at all. Choosing a family therefore means choosing a set of limits, which is why they are listed family by family on this page instead of being averaged into one figure.
This page does not host a workbench of its own. Its buttons lead to /create/, where the prompt is typed and the generation happens, and video tasks settle against the same points balance as image tasks so that a clip does not require a second subscription. Nothing here describes a complimentary allowance, a generation time or a per-second cost, because the site catalogue publishes none of those three things. Where a number comes from a vendor rather than from this site, it is attributed to that vendor in the sentence that carries it.
Each rule below names the vendor that wrote it. They differ from one another, and none of them is a limit set by this site.
Kling caps a text prompt at 3,072 characters and recommends staying under 2,500. Vidu allows 5,000. MiniMax allows 2,000. This site publishes no prompt length limit of its own, so the three figures belong to the three vendors and should be read as theirs.
Multi-shot prompts follow the form “shot n, m, words;”, separated by standard semicolons. n is the shot number, with at most six shots and at least one. m is that shot’s duration in seconds, no shorter than one, and the shot durations must sum to the clip’s total length. words is the prompt for that shot, capped at 512 characters.
Fifteen commands are supported: Truck left and right, Pan left and right, Push in and Pull out, Pedestal up and down, Tilt up and down, Zoom in and out, Shake, Tracking shot and Static shot. Commands inside one pair of brackets take effect together, and MiniMax recommends combining no more than three. Placing them in order makes them sequential, and plain language still works but yields less precise results.
Wan takes shot structure from the prompt itself, with timestamps such as “Shot 1 [0–3 seconds] wide shot:”. Its reference notes that the shot_type parameter has no effect for this reason. Leaving the structure out lets the model interpret the prompt on its own.
Wan accepts a negative prompt of up to 500 characters in Chinese or English and truncates longer input automatically, with an example list reading “low resolution, error, worst quality, low quality, deformed, extra fingers, bad proportions”. Kling describes its prompt field as able to carry negative descriptions alongside positive ones.
xAI states that its text-to-video path generates a first frame from the prompt and then animates it, and that the intermediate image is not returned. The implication is that prompt errors affecting composition show up as a still before they show up as motion.
Veo 3.1 generates 8-second videos with natively generated audio at 720p, 1080p or 4k. Its aspect ratio choice is landscape 16:9, which is the default, or portrait 9:16. 4K is not available on the Lite tier, and video extension is limited to 720p.
Wan documents that a task ID stays valid for queries for 24 hours, and Kling documents that generated results are cleared after 30 days. Downloading a clip you intend to keep is part of the workflow rather than an optional step at the end.
This page is written against vendor documentation and the site video catalogue. Where the two differ, both are shown rather than merged into one confident sentence.
Read the termsA text-to-video model turns a description into motion without any supplied material. It has to invent the opening image, keep the subject consistent across the seconds that follow, and decide where the camera sits. The first of those is where the mode earns its cost: an image-to-video model already knows what the scene looks like, while a text-to-video model is choosing the subject, the setting, the light and the lens position from the prompt alone. Vendors describe the split in their own words. Vidu lists text to video, image to video, reference to video and start-end to video as separate functions with separate model lists. Kling publishes a text-to-video endpoint on each of its two 3.0 tiers. MiniMax documents its text-to-video models under their own request schema and states that image to video and reference to video are mutually exclusive.
The prompt is the only input, and every vendor constrains it differently. Kling caps a text prompt at 3,072 characters and recommends staying under 2,500. Vidu allows 5,000. MiniMax allows 2,000. Beyond raw length, two vendors define a notation for describing shots inside a single prompt. Kling accepts multi-shot text written as “shot n, m, words;”, where n is the shot number up to six, m is that shot’s duration with a one-second minimum, and the shot durations must add up to the clip’s total length. MiniMax reserves square brackets for fifteen camera commands such as [Truck left], [Push in] and [Tracking shot], and recommends combining no more than three inside one pair of brackets. Wan takes a third route and lets shots be written as timestamps, as in “Shot 1 [0–3 seconds] wide shot:”, while its own reference notes that the shot_type parameter has no effect because the prompt governs the structure. None of these notations is a field on this site.
Because the prompt carries everything, the rules that matter most are the ones about what a prompt can be trusted to control. Kling and Wan both expose a negative prompt: Wan documents one of up to 500 characters, accepting Chinese and English, and truncates anything longer automatically. MiniMax exposes prompt_optimizer, on by default, which lets a model rewrite a short prompt before generation; turning it off trades convenience for precision. Wan exposes prompt_extend with the same trade written into its documentation, noting that rewriting helps short prompts but adds latency. A prompt written for one of these systems is not automatically a good prompt for another, which is why the rules are listed here by the vendor that wrote each one.
Three steps from a written scene to a downloadable clip.
A single-shot prompt describes one continuous camera move. If the clip needs more than one, use the notation the vendor documents: Kling takes “shot n, m, words;” with up to six shots and a one-second floor on each, MiniMax takes bracketed camera commands such as [Push in] with no more than three combined, and Wan takes timestamps written beside each shot.
Duration falls into four groups, with fourteen families on a 2 to 15 second window, one running 4 to 30 seconds and one offering three fixed choices. Aspect ratio lists run to five ratios on eleven families, seven on three and three on two. Set both after choosing the family, because neither is shared across the catalogue.
Generated files come back as downloadable media rather than permanent storage. Wan documents that a task handle stays queryable for 24 hours, and Kling documents that result URLs are cleared after 30 days. Save what you intend to keep instead of treating the link as an archive.
Every figure below is read from the site video catalogue, family by family, rather than summarised from a marketing page.
Sixteen video families take a prompt with no image attached. The five that do not are hailuo-3, midjourney-video, viduq3, viduq3-mix and viduq3-pro-fast. Two of those five have every mode disabled, so they cannot be used for any video task on this site.
Ten families run 2 to 15 seconds and open at 5. Four share the same ceiling but open at 2. seedance-2.5 runs 4 to 30 seconds, the longest window in the catalogue. veo-3.1 accepts no arbitrary length and offers 4, 6 or 8 seconds only.
Eleven families offer 720p and 1080p. Two can also drop to 480p, two add a 4K option, and one stops at 480p and 720p. Only kling-o3 and veo-3.1 include 4K.
Eleven families offer five ratios — 16:9, 9:16, 1:1, 4:3 and 3:4 — with 16:9 at 720p as the default. Three families add 21:9 and adaptive for seven in total. kling-o3 keeps to three, and veo-3.1 uses auto, 16:9 and 9:16 instead of the usual list.
Only seedance-2.0, seedance-2.0-fast and veo-3.1 expose a generated-audio switch in this catalogue. Vendors document more audio behaviour than that, since Wan produces background music or sound effects when no audio file is supplied and Vidu defaults its audio parameter to true on q3 models, but those are API fields rather than controls on this site.
Video tasks and image tasks return the same points unit from the pricing layer, so a clip generated from a prompt and a still edited from an upload settle against one balance instead of two subscriptions.
The clearest case is the one where no footage was ever going to be shot. A campaign line that needs to be seen before it is scheduled, a product idea with no physical sample yet, an app screen that does not exist, a location that has not been scouted: text to video produces a visual answer without a shoot, a stock licence or an illustrator. Because nothing is uploaded, the first draft costs nothing but a rewritten sentence, and rewriting the sentence is the fastest lever the mode offers.
The second case is creative testing. Hooks, opening frames and pacing choices can be compared across several prompts before anyone commits to production, and the comparison stays legible when every variant shares one prompt structure so the manipulated variable is the only thing that changed. The third case is illustration inside another piece of work: a section divider that moves, a background loop for a landing page, a visual for a slide that had only text. In all three the clip is a component rather than the deliverable, and the family gets chosen for its duration and aspect ratio first.
Open Create, write the shot you want, and pick the family whose duration and aspect ratio fit the channel it is going to.
Try text to video