A team making a short product video may need ten attempts to settle on a camera move, then one polished version for delivery. Generation speed matters in the first stage. Reference limits and output format decide whether a model can handle the second.
Seedance 2.5 Flash (Flash, for short) is BytePlus's faster, lower-priced counterpart to Seedance 2.5. It accepts the same input types and the same 4–30 second clip lengths. The full model adds more reference capacity, 1080p output, and a dedicated preview-to-final workflow. Choosing between them starts with the shot you need to deliver.
What changes between Flash and 2.5
BytePlus lists Flash in its November 2026 model releases as dreamina-seedance-2-5-flash-261115. It supports text-to-video, first-frame and first-and-last-frame generation, multimodal references, video editing, and extension, but not Draft mode. BytePlus model releases
Capability | Seedance 2.5 Flash | Seedance 2.5 |
|---|---|---|
Input types | Text, images, video, audio | Text, images, video, audio |
Reference images per request | Up to 10 | Up to 30 |
Reference videos per request | Up to 5 | Up to 10 |
Reference audio clips per request | Up to 5 | Up to 10 |
Generated clip duration | 4–30 seconds | 4–30 seconds |
Output resolution | 480p, 720p | 480p, 720p, 1080p |
Draft mode | Unsupported | Supported |
The resolution gap goes beyond pixel count. Both models output 8-bit color at 480p and 720p, while Seedance 2.5's 1080p output uses 10-bit color and H.265/HEVC encoding. BytePlus notes that 10-bit preserves smoother tonal transitions and suits professional production, and that HEVC files may not play in every player. These are output specifications, not a quality ranking for every scene. BytePlus series tutorial
The input types match, but the creative brief may not fit
Both models accept text, images, video, and audio. Audio can be the only reference, or it can join images and video in the same request, so a workflow built around a music track does not require the full model.
Reference images and explicit first-and-last-frame inputs belong to different generation modes, and the API does not let you combine them in one request. If matching the opening or closing frame is essential, use the frame-based mode. Otherwise, use references.
Both models also cap total reference time. All reference videos together must run no longer than 30 seconds, and all reference audio clips have a separate 30-second limit, so five video slots do not mean five 30-second clips. Face uploads are restricted as well: ModelArk blocks direct uploads of real human faces and documents routes for authorized portraits, digital characters, and certain model-generated assets. BytePlus generation API
For example, a product shot built from four pack photographs, a camera-movement clip, and a music reference fits within Flash's limits. A brief that needs twelve separate product and location images exceeds its ten-image cap, even though Flash accepts every media type involved.
More references also call for clearer instructions. BytePlus's prompt guide describes assigning each material a role, such as subject, motion, style, or audio, and cautions that stability can drop as scenes become more crowded. A useful brief names the image that defines the product and the clip that supplies the camera movement. Uploading every available asset leaves those priorities unresolved. BytePlus prompt guide
How much faster is Flash?
BytePlus says Flash delivers results close to Seedance 2.5 at nearly twice the generation speed. The announcement gives no workload, queue conditions, or timing distribution, so treat it as a vendor claim to test, not a promise that every request finishes in half the time. Early independent coverage repeats the figure without timing tests of its own. BytePlus announcement
To test it, run the same brief on both models at 720p with matching duration, references, and audio settings. Measure submission-to-completion time across several attempts, then count how many clips meet the brief. A fast result that needs a second generation can lose its advantage before an editor downloads it.
Judge motion continuity, product geometry, and adherence to the requested camera move separately. For dialogue, add intelligibility and lip synchronization. Those results will show whether Flash's speed helps your workload, which the published specifications cannot.
What the lower price buys
As of November 16, 2026, BytePlus lists online inference without video input at $7.00 per million tokens for Flash and $10.70 for Seedance 2.5 at 480p or 720p, which makes Flash about 34.6% cheaper. BytePlus's own estimate for a five-second, 16:9, 720p clip without video input is $0.76 for Flash, or $0.152 per second, against $1.156 for 2.5.
Video input changes the arithmetic. The rates fall to $4.30 and $6.40 per million tokens, but input video duration counts toward token consumption alongside output duration, and minimum consumption rules apply. For the same five-second 720p output, BytePlus's estimates become ranges: $0.84–$3.25 for Flash and $1.244–$4.838 for 2.5. BytePlus pricing
For budgeting, divide spend by accepted clips and hold the acceptance standard fixed. If the cheaper model needs enough extra attempts, its nominal savings can disappear.
Choose the workflow before the model
Seedance 2.5's Draft mode is a two-step route: generate a 480p preview, then pass its task ID to request the final video. The final render reuses the Draft's model, prompt, input assets, seed, and other key parameters, and it currently comes out at 1080p. Flash does not support Draft mode, so teams that plan to approve composition on a low-resolution preview first should start with 2.5.
Flash is a sensible candidate for exploring camera moves and for deliverables that fit its resolution and reference limits. Start with Seedance 2.5 when the brief requires 1080p, a larger reference set, or the Draft workflow. Either way, choose the model for a requirement you can name and verify in the finished clip.
