Seedance 2.5 draft mode renders videos at 480p before HD conversion, Techhalla tests

• The workflow starts with a Magnific One still, animates it at 480p in Seedance 2.5, then renders the chosen take in HD, as demonstrated in techhalla's demo.
• Branding and lettering are reasons to use the HD step, according to techhalla's reply about preserving specific details.
• The creative work happens inside timed prompts: the cenote script specifies camera movements, dialogue and a six-part prank across 30 seconds.

A single draft_task_id requests the HD pass in Unifically's API documentation, carrying over the original prompt and references. Seedance 2.5 accepts up to 30 reference images per generation, according to ByteDance's launch post, while techhalla's cenote scene assigns distinct jobs to just two.

For the Maracanã scene, techhalla's image prompt establishes the starting composition before animation:

• Viewpoint: a high-angle shot across the stadium crowd toward the pitch.
• Foreground: fans' heads and shoulders, with several people holding phones.
• Lighting: bright midday sun, hard shadows and a red flare drifting through the stands.
• Look references: Leica M6, a 35mm lens and Kodak Portra 400.

Techhalla describes three steps in his workflow post:

1. Generate the starting image in Magnific One.
2. Use it as a reference in Seedance 2.5, with Draft mode producing a 480p video.
3. Select a result, then use Generate in HD (1080p).

He calls the 480p stage “super cheap” in his demonstration, without giving a per-clip credit total.

Unifically describes both outputs as native generations from the same task in its draft documentation. The HD render keeps the draft's duration and aspect ratio; the provider lists 720p and 4K renders from drafts as coming soon.

Techhalla says the feature retains small details better than upscaling, particularly branding and text.

Magnific also describes a separate 4K delivery route using its video upscaler.

Techhalla divides the stadium sequence into seven timed beats, all inside one continuous take:

1. 0-1.5 seconds: a wide view establishes the bouncing crowd and match below.
2. 1.5-5 seconds: a handheld swing and hunting zoom find a silver hatchback above the fans.
3. 5-10 seconds: the crowd passes the car overhead; passengers add a red flare and Brazil flag.
4. 10-17 seconds: the camera follows the car across the stands, gradually pulling back.
5. 17-23 seconds: successive rows push it toward the camera.
6. 23-27 seconds: the car approaches close enough for its TECHHALLA plate to be readable.
7. 27-30 seconds: the car passes overhead, exposing its underside before the phone gets knocked skyward.

In his second script, techhalla separates the scene reference from the character reference:

• Image 1: the first frame, preserving the woman, clothes, arm tattoo, rocks, wooden boat, manta rays and crowd.
• Image 2: the small man's face, build, velvet tailcoat and curled-toe shoes. Its background is explicitly excluded.

The camera brief is pleasingly scruffy: an iPhone 13 with a sunscreen-smeared lens and blocky video compression.

His requested sequence gives the model six timed actions:

1. 0-7 seconds: the woman warms up while the crowd chants and claps.
2. 7-11 seconds: the camera tilts down to a shark fin; the person filming warns her.
3. 11-16 seconds: she performs a twisting backflip into the water.
4. 16-22 seconds: the fin approaches and she swims toward the boat.
5. 22-26 seconds: the fin is revealed as a rubber prop strapped to the small man.
6. 26-30 seconds: he climbs onto the bow and dances as the crowd laughs.

The cenote stunt and fin reveal

The audio brief requests only sounds within the scene, with no background music or subtitles.

Magnific's official site offers a macOS desktop download and says the desktop version uses the same account, credits and tools as the platform.

Techhalla also links a Space containing more than 100 Magnific One prompts in his workflow post.