Giving Seedance 2.5 and MiniMax H3 the same pavilion shot
We asked both models to push in from outside, then cut into the pavilion. H3 stayed outside; Seedance continued into the orbit.
CraftWorld · Updated
On 25 September, we tried Seedance 2.5 and MiniMax H3 with a wooden pavilion on a mountain. Someone plays a violin inside while rain falls outside. The camera moves along the path for five seconds, cuts to a low angle inside, then arcs around the person for four seconds.
We wanted to see how much of the camera work already placed in 3D would survive. These are the previews sent to the models: two lit exports of the same scene.
H3 stays outside
H3 changes things from the beginning. The straight path becomes irregular flagstones, and the rocks and trees change too. The camera does not reach the eaves as it does in the preview. Near five seconds, it is still looking diagonally at the whole pavilion from outside.
The cut should happen here, but it does not. At seven seconds H3 is still outside. It gets closer to the violinist near the end without reaching the planned angle, and waterfall-like streams appear. Those streams are absent from the preview.
Seedance keeps the cut
Seedance is closer to the planned shot. At 4.9 seconds it has reached the eaves, with the steps in front and the person near the center. At 5.1 seconds the view changes to the interior, with the right-hand column entering the frame.
Over the last four seconds the railing passes through the foreground, the person turns toward profile, and the pine shifts toward the right. Details change, but the low arc remains.
Five seconds is the useful point to compare in these clips. You can see whether the camera changed position. Looking at a single frame near the ending makes it easier to miss what happened at that cut.
For this shot, we would use the Seedance result because both planned shots are there. It also used the lower resolution: 480p against H3’s 768p. This is about whether this particular sequence was reproduced. With one generation each, it does not establish how the models compare across other scenes.
Settings and original records
Both requests specified 9 seconds at 16:9. The references are 1280×720 at 24 fps and exactly 9.000 seconds long. They are separate exports, but frames sampled at 0, 2.5, 4.9, 5.1, 7.0, and 8.8 seconds have a structural similarity score (SSIM) of at least 0.995. All four videos here are original files without recompression.
The project keyframes were checked on 29 September. The push-in runs from about 14.5 m to 6.6 m at a 48° vertical field of view. The arc has a radius of about 3.8 m, height of about 0.82 m, field of view of 36°, and roughly 90° of travel. The opening, cut, and ending of the previews exported on 25 September match those positions.
H3 was submitted at 15:57 Beijing time and Seedance at 16:04. The intervals from record creation to the last update are 105 and 192 seconds, both including queue time. Job IDs are vid_6riKm8qeWn0YnsO and vid_cbO_l0eMXVjKngTY respectively. Seedance enabled generated audio with seed at its default of -1; H3’s request has no generated-audio parameter. Both returned files have detectable sound. Their lengths are 9.425 and 9.056 seconds respectively.
Visual direction was the same, while system instructions differed. H3 was asked to reproduce the preview from first frame to last. Seedance was asked to follow camera and object motion without copying the mesh. These are the requests the product actually sent, without rewriting them to match for comparison.
Neither recorded job charged credits; that is not the public price. Based on the reference-video rate sheet dated 23 September 2026 and the current default route, those settings estimate at about 18 video credits for H3 and 28 for Seedance. Check the displayed quote when submitting your own request.