Seedance 2.5 anime test: 20 s rooftop fight asked for as one unbroken shot, from a neon sign to a sky exit
Pink neon sign close-up, a katana tip sliding in, legs of an armoured fighter, a water tower, a two-fighter standoff, a clash with sparks and a final rise into clouds.
Task: A 2D hand-drawn action-anime fight on a rainy rooftop, asked to play as one continuous shot with no cuts.
Outcome: succeeded
0.84 s: Close-up of a buzzing pink neon sign.
2.51 s: A katana edge sweeps across the sign.
4.18 s: Wider: blade and sign, a wet rooftop below.
5.85 s: Legs of an armoured figure beside the sign, a water tower beyond.
7.52 s: Wide rooftop with a water tower at left; a distant figure.
9.20 s: Water-tower framework fills the frame.
10.87 s: Two fighters face each other, a pink sign behind.
12.54 s: The fighters clash, close together.
14.21 s: Armoured fighter with a gold crest against a dark fighter, orange sparks.
15.88 s: Heavy motion-blur streaks.
17.56 s: A tiny figure high among dark clouds.
19.23 s: A crouched silhouette in the clouds.
Intent vs result: The prompt asked for a single unbroken shot; the sampled frames move from the sign to the tower to the fight to the sky and the detector found 1 transition, but whether the clip is truly one take cannot be settled from 12 frames.
Publication update (2026-10-29): the founder has now cleared public use of this output, including incorporated input likenesses and soundtrack. Earlier rights/likeness holds described above are resolved by that confirmation. Reference-input files remain private. Audio remains not listened to; visual observations cover sampled frames only.
Structured fields
{
"run_at": "2026-08-08T04:31:08Z",
"metrics": [
{
"name": "output_duration",
"unit": "s",
"value": 20.064
},
{
"name": "resolution",
"value": "1280x720"
},
{
"name": "frame_rate",
"value": "24/1"
},
{
"name": "scene_detector_transitions_thr0.30",
"value": 1
},
{
"name": "black_intervals",
"value": "none"
}
],
"procedure": "Seedance 2.5, one reference still used as the look-and-world anchor, a long structured prompt (one subject, a setting and a shot-by-shot format), 20 s, 16:9, audio requested, no seed set. One of a run of tests on 9-10 August 2026 (UTC) that each pair the same kind of reference still with a different subject, genre and visual style. Sampled inspection: 12 timestamped frames from 0.84 s to 19.23 s of a 20.06 s file, plus ffmpeg scene detection at threshold 0.30 and blackdetect.",
"parameters": {
"seed": "unset",
"model": "seedance_2_5 (label in job metadata)",
"multi_shots": false,
"aspect_ratio": "16:9",
"audio_requested": true,
"resolution_label": "720p",
"requested_duration_s": 20
},
"limitations": [
"Sampled visual inspection from timestamped frames (contact sheets), not full real-time playback; motion smoothness, micro-timing and lip-sync are unrated.",
"Audio was not listened to; the file has an audio track but sound, dialogue and music are unassessed.",
"Prompt intent is taken from private job metadata and summarised here only in general terms; the founder's taste and intent for each test are unknown.",
"Single historical generation without a fixed seed; not a controlled benchmark and not validation of any current product or pipeline.",
"Whether the clip is one unbroken shot is not confirmed by sampled frames."
],
"artifact_ids": [
"ast_01M64W4D7HX95ZT4BCGS99R8XP"
]
}Evidence claims
- [sourced_fact] Timestamped observations above come from viewing 12 sampled frames of the output (sampled visual inspection, not full playback).
- [upstream_claim] The generation was run by the founder in an earlier test period; H01 did not run or reproduce it.
- [inference] A one-take instruction produced a continuous-looking sweep in the sampled frames, but only full playback can confirm it.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Original analysis written by H01 (an AI agent) for Veyrul from private viewing; no third-party text, frames or audio reproduced. |
| media | unknown | permitted | permitted | unknown | unknown | Founder confirmed public use of all these Higgsfield outputs on 2026-10-29, including incorporated reference inputs, likenesses and soundtrack. Audio has not been listened to; permission does not establish listening or independent review. |