Seedance 2.5 weak spots: physics and crowded scenes, a test plan

ByteDance says Seedance 2.5 can still get complex physical motion and scenes with several interacting subjects wrong, so test those two cases before you trust the model with them. The Seed team states in its launch post that there is room for improvement in the physical plausibility of complex motions and in the stability of scenes involving interactions among multiple subjects. This page turns that sentence into a short test plan you can run on Sume at 480p, with frame checks that show where a clip breaks.

What exactly does the vendor admit?

The closing section of the Seedance 2.5 launch post, dated 2026-07-31 and read 2026-10-03, names two areas: the physical plausibility of complex motions, and scene stability when multiple subjects interact. It also says the team will keep working on coherence, editing and real-world physics.

The same post lists strengths: 30-second clips, up to 30 images, 10 video clips and 10 audio clips as references, and timestamp-level editing. A strength list and a limits note can both be true; the test plan checks the second one for your content.

How do you run the test on Sume?

Submit the same three prompts to seedance-2.5 at 480p, once at 10 seconds and once at 30 seconds. The Video Router doc lists 4 to 30 seconds and 480p, 720p and 1080p for this model. Sume does not accept a seed, so each run is a different take; run each prompt twice and note whether the failure repeats.

Keep the prompt short and literal, one subject and one action for the physics cases, and add references only for the crowded-scene case, so you can tell which input caused a failure.

How do you look at the frames?

Video frames extracts stills from one Sume-hosted clip at times you name, as durable image artifacts at source size. Send at with 1 to 24 seconds, or fps up to 2, and choose png when you want a lossless frame. Each time must be inside [0, duration) or the job fails with frame_time_out_of_range.

Video inspect is the quicker first look: with no frames field it returns eight mid-bin stills at a default long edge of 768, plus probe facts. Its seek: "fast" option snaps to the keyframe at or before each instant, so use precise (the default) when the exact moment matters, such as the frame where the ball hits the wall.

What if the test passes?

A pass at 480p means the idea works, not that the final will be flawless. The higher resolutions are separate requests that produce a new take. Run the approved prompt at 720p or 1080p, repeat the same frame checks on that output, and keep both outputs until the final is signed off. Seedance 2.5 draft mode explains why a final is not a re-render of the draft.

Also test the content you will really ship. A prompt about a generic ball tells you little about your product, your character or your location. Put your own reference images in the crowded-scene test, and use the clip length you plan to order, because a failure that appears only in the last ten seconds of a 30-second clip will not show in a 10-second run.

What do you do with the result?

Score each clip pass or fail against one sentence, such as "the ball keeps its size". If a prompt fails twice at 480p, a higher resolution is unlikely to repair it, so rewrite the shot as simpler: one subject, one motion, a shorter clip. If it passes, rerun only that prompt at the resolution you need. Record the prompt, the request body and the verdict, because a later model update can change the result.