Runway Launches Draft Mode for Seedance 2.5, Faster Generation and Lower Points Consumption in Exploration Phase, Enhance Clips to Full Quality After Selection
Runway announced the addition of draft mode for Seedance 2.5: faster generation in the exploration phase and lower points consumption, with clips enhanced to full quality after selection. Seedance 2.5, developed by ByteDance, was launched on Runway in August, supporting up to 50 references for text, images, videos, and audio, with a maximum native length of about 30 seconds, allowing for multiple shots, dialogues, and sound design. It offers four working modes: reference, keyframe, editing, and extension, aimed at paid packages.
Volcano Engine also launched a draft for the Seedance 2.5 API in the same week: first producing 480p samples to validate scenes, composition, actions, dialogues, sound effects, and camera movements, then initiating 1080p final output using the sample ID, reusing prompts, references, random seeds, ratios, and durations for secondary inference. The official example shows that a 5-second 1080p without input video, with 4 draws, can save about 57% in costs and nearly double the speed; savings can reach about 77% with 20 draws. The final output is native 1080p rather than upscaling from 480p, with potential differences in local textures, small text, and high-density crowds, while aiming to lock in shot structure and movement direction. Runway, Magnific, and Artlist have participated in effect comparisons, claiming details are superior to common upscaling, with motion flaws corrected in the final output.
Video generation is billed by seconds and resolution, with repeated outputs in the exploration phase being the main cost driver. The draft separates "finding the right shot" from "delivering usable pixels," aligning with platforms like Higgsfield that turn generation into a marketing pipeline, but differs as Seedance places control over the number of references and a single 30-second long shot. Runway also rents out its own Gen-4.5 and other third-party models, with the draft extracting costly model trial and error from full resolution.
Mechanically, what is purchased is low-resolution trial and error counts, while the old billing deducts points for each exploration at 1080p. The event-driven aspect is the platform integrating the two-stage inference of model vendors into the creative interface. Funding shifts from wasting points on unusable footage to final enhancement; benefiting are studios that still need to draw a lot under pay-per-use billing, while under pressure are billing models relying on full-resolution trial and error and tools that only upscale without secondary inference. Samples and final outputs are not pixel-for-pixel copies, and brand review may still require another version to be run.
Source: Public Information
ABAB AI Insight
The billing for generated videos occurs during the "drawing cards" phase, not on the final output. Runway's integration of the draft mode for Seedance acknowledges that the model with 30 seconds and 50 references is too expensive, preventing users from using final resolution to find creativity. The 480p version is responsible for determining shot validity, while the 1080p version is for delivery, with seeds and references locked in to reduce "low-res looking good, high-res failing." The 57% to 77% savings provided by Volcano are a function of the number of draws: teams that struggle with prompt writing find drafts more valuable.
The capital path involves tiered points. Paid packages sell monthly points pools, with drafts reducing the speed at which the pool runs out, improving retention, and allowing 30-second long shots to shift from demonstration to an iterative process. Runway acts as a distribution layer, packaging ByteDance's API into buttons, profiting from the channel price difference; model vendors use drafts to lower the enterprise access threshold and increase usage. The upscaling route has been pointed out as inferior, indicating that secondary inference aligns better with motion consistency than mere upscaling algorithms.
Analogous to proxy previews in 3D software, raw exports in photography, and draft models in encoding assistants: all separate exploration and delivery into two price points. The industry phase shifts from who can produce 30 seconds to who can make 30 seconds affordable for testing. Whoever establishes consistency from draft to final output as a committed metric can enter advertising review; those who only reduce resolution may lose composition in the final output.
Structural judgment belongs to technological substitution. Pricing power shifts from "per second full resolution" to "separate pricing for exploration seconds and delivery seconds." The mechanism indicates that the cost of failure in video generation is far higher than the number of successes, with two-stage inference leaving failures on cheap computing power. Points remain points, but unusable footage is no longer sold at final output prices.
ABAB News · Cognitive Law
Money for generated videos is spent on drawing card counts, not on the final output.
Locking in shots before delivering pixels is cheaper than delivering pixels first and hoping for the best.
Once drafts can reuse seeds, upscaling shifts from a process to an option.