On September 22, Volc Engine announced the launch of the Draft (Sample) mode for Seedance 2.5 API, splitting the video generation process into two stages: sample preview and final output. Creators can first generate a low-cost sample at 480P to check if the scene, composition, subject actions, dialogue and sound effects, and camera movement rhythm meet expectations. After refining or adjusting the prompt and reference materials, they can then initiate the 1080P final output based on the sample ID.

The model will perform secondary reasoning by associating the sample's prompts, reference materials, and key parameters such as seed, ratio, and duration, ensuring the final output maintains a high level of consistency with the sample in structure, composition, actions, and overall direction.

QQ20260923-102858.jpg

According to official data, using the Draft mode can save 57% of the cost and nearly double the generation speed when generating a 1080P, 5-second shot without input video and with four card draws. The more frequent the card draws, the more significant the cost savings, reaching up to 77% when drawing 20 times. Compared to the common approach of generating at a low resolution and then upscaling, this mode directly outputs 1080P native quality with richer details, more natural light and shadow layers, and color saturation. There may be minor changes in elements such as local textures, complex patterns, small text, and the number of people in high-density crowds, but the lens structure and motion performance confirmed in the sample can be fully maintained.

After testing, Runway found that the 1080P final output generated by the Draft mode has some differences in detail compared to mainstream upscaling strategies, but the effect is actually better; Magnific noted that it generates quickly and maintains stable consistency; Artlist stated that there is no AI feel commonly found in upscaling algorithms, and the motion imperfections in the 480P preview are fixed in the 1080P rendering.

QQ20260923-102919.jpg