ByteDance has launched a Draft preview mode for the Seedance 2.5 API. The feature lets users first generate a lower-cost 480P sample clip, review elements such as framing, motion, dialogue, and camera movement, and then turn a selected draft into a 1080P final video.
Jimeng has also added 「Seedance 2.5 (Draft Mode)」.
1080P generation is tied to the original draft ID
The workflow differs from the common approach of testing a prompt in low resolution and then rerunning the same prompt to get a high-resolution result. When generating in 1080P, the model links back to the original draft ID and reuses parameters including the prompt, reference materials, seed, aspect ratio, and duration for a fresh inference pass, with the goal of keeping the original structure, composition, and motion as consistent as possible.
Still, local textures, small text, and the number of people in a crowd may change. That means it is not a frame-by-frame reproduction, and it is not a simple 480P-to-1080P super-resolution process either.
Official estimates on cost and speed
According to the official estimate cited in the source, for a 5-second 1080P video without video input, Draft can reduce cost by 57% and nearly double overall generation speed if it takes an average of four attempts to get a satisfactory result.
If it takes 20 attempts to select one usable clip, cost savings can reach as high as 77%.
Similar setup had appeared in Seedance 1.5 Pro
A similar Draft setup had previously appeared in Seedance 1.5 Pro. This update brings that lower-cost trial-and-error workflow into Seedance 2.5 and allows the selected draft clip to continue into a corresponding 1080P final render.
The source says the most expensive part of AI video generation is often the repeated "gacha-style" trial phase at the front end, and this update is aimed directly at reducing that portion of the cost.