Master AI video generation with this complete guide: prompt structure, reference images, camera language, credit optimization, and Seedance 2.5 settings. Best practices for better AI video results.
Master video generation with this complete guide. Learn prompt structure, reference image best practices, camera language, credit optimization, and the Seedance 2.5 settings that improve every render.
Everything here is written against Seedance 2.5 as it actually runs on 智擎云 - 4-to-30-second generations at 24 fps, up to 1080p, seven aspect ratios, up to 30 reference images and 10 reference clips per request, native audio, and credit costs quoted before you generate. Choose Seedance 2.0 up to 4K when a shot needs a higher-resolution finish. Where a limit is mentioned it is the real limit, enforced by the generator rather than rounded for the marketing copy.
Every successful generation starts with a clear prompt, the right references, and appropriate settings.
A clip runs 4 to 30 seconds. One clear action with one clear camera behaviour is still what fits best, and the longer end buys continuity rather than extra beats. Multi-beat narratives are built from several generations joined afterwards, not requested in a single paragraph.
Unspecified camera work becomes a slow drift. Stating the move and its speed - "slow push-in", "fast whip pan" - is the single highest-value addition to a prompt that is nearly working.
Explore at 480p on Seedance 2.0 Mini, then reuse the selected prompt and references in a new Seedance 2.5 generation at 1080p - or on Seedance 2.0 when the shot has to finish at 1080p or 4K. The maximum reservation appears before submission, and the new result may differ.
A reliable prompt has five parts, and you can check them off: subject, action, camera, light, mood. "Slow push-in on a chef plating scallops, tungsten key from the left, shallow depth of field, steam rising" contains all five in one sentence. What it deliberately contains none of is praise - "cinematic", "stunning", "8K", "masterpiece" describe how you feel about the result rather than what should be in it, and the model cannot act on any of them. Replacing every such word with a concrete noun, a lens, or a lighting term is usually the fastest single improvement available.
When a result is close but wrong, resist the urge to rewrite. Change one variable and regenerate, so you learn which word was doing the work. If the composition is wrong, adjust the camera and shot size; if the feeling is wrong, adjust the light; if the subject is wrong, add a reference image rather than more adjectives; if the motion is wrong, adjust the speed qualifier before anything else. And if the model keeps ignoring part of the prompt, it is usually a sign that the prompt is asking for more than 30 seconds can hold - cut it down and generate the rest as a second shot.
Subject, action, camera, light, mood. If a prompt is underperforming, one of the five is usually missing entirely.
Composition wrong, change the camera. Feeling wrong, change the light. Subject wrong, add a reference. Motion wrong, change the speed.
Verified model data
A source-linked snapshot generated from the same model catalog and pricing rules that validate real jobs. Download the data, cite the anchored sections, and check the stated failure boundaries before choosing a workflow.
Updated Sep 19, 2026
Compare text, frame, and reference input modes across every Seedance model in 智擎云, including output limits, input constraints, audio behavior, and a parameter-matched credit reservation.
| Model / provider ID | Output | Duration | Input constraints | Audio | Example quote |
|---|---|---|---|---|---|
Seedance 2.5workspace.mode.textdoubao-seedance-2-5-260628 | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–30 seconds | — | Native audio output | 799 credits |
Seedance 2.5workspace.mode.framesdoubao-seedance-2-5-260628 | 480p / 720p / 1080p · Auto | 4–30 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 799 credits |
Seedance 2.5workspace.mode.referencedoubao-seedance-2-5-260628 | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–30 seconds | 30 images · 10 videos · 10 audio files | Native audio · audio-only reference | 799 credits |
Seedance 2.0workspace.mode.textdoubao-seedance-2-0-260128 | 480p / 720p / 1080p / 4K · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | — | Native audio output | 525 credits |
Seedance 2.0workspace.mode.framesdoubao-seedance-2-0-260128 | 480p / 720p / 1080p / 4K · Auto | 4–15 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 525 credits |
Seedance 2.0workspace.mode.referencedoubao-seedance-2-0-260128 | 480p / 720p / 1080p / 4K · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | 9 images · 3 videos · 3 audio files | Native audio · reference needs image or video | 525 credits |
Seedance 2.0 Fastworkspace.mode.textdoubao-seedance-2-0-fast-260128 | 480p / 720p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | — | Native audio output | 420 credits |
Seedance 2.0 Fastworkspace.mode.framesdoubao-seedance-2-0-fast-260128 | 480p / 720p · Auto | 4–15 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 420 credits |
Seedance 2.0 Fastworkspace.mode.referencedoubao-seedance-2-0-fast-260128 | 480p / 720p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | 9 images · 3 videos · 3 audio files | Native audio · reference needs image or video | 420 credits |
Seedance 2.0 Miniworkspace.mode.textdoubao-seedance-2-0-mini-260615 | 480p / 720p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | — | Native audio output | 263 credits |
Seedance 2.0 Miniworkspace.mode.framesdoubao-seedance-2-0-mini-260615 | 480p / 720p · Auto | 4–15 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 263 credits |
Seedance 2.0 Miniworkspace.mode.referencedoubao-seedance-2-0-mini-260615 | 480p / 720p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 | 4–15 seconds | 9 images · 3 videos · 3 audio files | Native audio · reference needs image or video | 263 credits |
Wan 3.0workspace.mode.textwan3.0-video | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 | 4–30 seconds | — | Native audio output | 300 credits |
Wan 3.0workspace.mode.frameswan3.0-video | 480p / 720p / 1080p · Auto | 4–30 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 300 credits |
Wan 3.0workspace.mode.referencewan3.0-video | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 | 4–30 seconds | 10 images · 5 videos · 5 audio files | Native audio · audio-only reference | 300 credits |
Wan 3.0 Primeworkspace.mode.textwan3.0-video-prime | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 | 4–30 seconds | — | Native audio output | 450 credits |
Wan 3.0 Primeworkspace.mode.frameswan3.0-video-prime | 480p / 720p / 1080p · Auto | 4–30 seconds | workspace.frame_label.required · workspace.frame_label.optional | Native audio output | 450 credits |
Wan 3.0 Primeworkspace.mode.referencewan3.0-video-prime | 480p / 720p / 1080p · Auto / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 | 4–30 seconds | 10 images · 5 videos · 5 audio files | Native audio · audio-only reference | 450 credits |
Comparable example: 5s, 720p, 16:9 where selectable; frame inputs use their required adaptive ratio. No video reference. This is a maximum reservation, not a flat price; successful jobs can settle lower.
cny-fen-2026-09-19-1to1Seedance 2.5 expands time and reference capacity; Seedance 2.0 retains the higher-resolution ceiling. Choose from the constraint that matters for the shot.
Longer clips and heavier reference sets
1080p and 4K output
Both workflows create a new generation. Extend continues from video reference material; Transition generates the bridge between two required endpoint images.
Extend video with AI by uploading a source clip and describing what happens next. 智擎云 creates a separate, reference-guided continuation rather than a frame-perfect native extension.
An AI video transition generator uses a required start image, a required end image, and a motion prompt to create new in-between video. It does not add a preset fade or wipe; review the generated middle for artifacts.
Model specifications
Published model limits, not marketing rounding. Every number below is enforced by the generator before your credits are reserved.
The model is ByteDance's. What 智擎云 adds is what makes it usable at work: honest limits, honest billing, and output you are allowed to ship.
Text, up to 30 reference images, 10 reference videos, and 10 audio references can guide one Seedance 2.5 generation instead of being stitched together afterwards.
When audio is enabled, Seedance 2.5 can generate footsteps, room tone, impacts, dialogue, and score with the video. Review timing and content before publishing, or switch audio off when the clip needs separate sound design.
One prompt covers a 9:16 vertical cut for Reels, a 16:9 master for YouTube, and a 21:9 anamorphic frame for a title sequence. Resolution runs from 480p to 4K, so the same take holds up on a cinema screen and on a phone.
The maximum credit reservation is calculated from the selected model, duration, resolution, ratio, and video-reference input, then shown before submission. A successful video job can settle lower from provider-reported usage; confirmed failed jobs return the reservation.
Use Mini for cost-efficient prompt exploration, Fast when you want a balance of generation speed and cost, and Seedance 2.5 when its longer duration or richer reference inputs fit the shot. Switching models starts a new generation, so results may vary.
Content generated with a paid subscription or purchased credit pack may be used commercially under the Terms of Service. Content generated solely with free sign-up or promotional credits is for evaluation and personal use only.
Three phases, in the order that wastes the fewest credits.
Write a short shot list - three to five beats, each with a shot size and an action. Knowing what you need is what keeps the credit spend proportional to the result.
Run each shot on Seedance 2.0 Mini at 480p, changing one variable per attempt. Save the prompt and reference set you want to reuse, while remembering that they do not lock the final output.
Reuse the selected prompts and references in a new Seedance 2.5 generation at 1080p, or in a Seedance 2.0 generation when you need 1080p or 4K, then join the shots in an editor or with Video Extend and Video Transition. Each result is a new generation and can vary.
They all run on the same model family. What differs is what you hand it and what you are trying to hold constant.
Use Text to Video to explore, and Motion Control to make the camera deliberate. This is the right starting point when you are still finding the idea.
Use Image to Video for a single subject, or Reference to Video when a character, product and art direction all have to hold at once across several shots.
Use Video to Video to restyle footage and Video Extend to continue it. Create an AI video transition between two images to bridge existing shots, or use a first and last frame video to guide how one new shot begins and ends.
Writing a story instead of a shot. A 4-to-30-second clip holds one clear action with one clear camera behaviour. Multi-beat sequences are built from several generations joined afterwards, not requested in one paragraph.
Draft at 480p on Seedance 2.0 Mini and change one variable per attempt, so each generation teaches you something. Only re-run at full resolution - Seedance 2.5 at 1080p, or Seedance 2.0 for 1080p and 4K - once the prompt and references are locked.
Usually because the prompt asks for more than the clip length can hold, or because it is full of words like "cinematic" and "stunning" that carry no actionable information. Cut it to one action and replace praise with concrete detail.
Whenever a specific real thing has to be accurate - your product, a particular face, an established art direction. Text will invent something plausible; only a reference pins the actual thing.
Build the sequence from several generations. Video Extend continues a shot past its final frame, and Video Transition generates the movement between two shots - both work better than asking for one very long take.
A confirmed failed generation returns its reserved credits automatically. 智擎云 shows the maximum reservation from your settings before submission, and a successful video job can settle lower when provider-reported usage is lower.
Each tool uses Seedance 2.5 for a different job. Move between them freely — your credits, history, and exports are shared.
Choose paid access when you need commercial-use eligibility, predictable generation costs, and credits restored after a failed or rejected job. One-time packs are available with no subscription.