AI news

Seedance 2.5 flash launches; Zhidx tests a 15-second cat parkour video

By 孙小雯1 viewsSource: 智东西
Seedance 2.5 flash launches, and Zhidx tests a 15-second video of a cat doing parkour

ByteDance reports faster generation and lower prices. Zhidx tests shot transitions and text stability in a 480P clip, finding that some background signs still lose parts of their characters.

Article illustration

Supports video generation of up to 30 seconds per clip.

Zhidx reported on November 12 that ByteDance today launched the Seedance 2.5 flash video generation model, describing it as “the model currently closest in quality to Seedance 2.5”. It generates videos twice as fast as Seedance 2.5 and costs less. Taking 720P text-to-video and image-to-video generation as examples, Seedance 2.5 flash costs approximately RMB 0.99 per second, about 34% less than Seedance 2.5. Generating a 10-second video costs approximately RMB 9.9.

The company says that Seedance 2.5 flash outperforms the previous flagship generation, Seedance 2.0, in overall quality, maintaining a high standard of instruction following, physical motion, and audio performance. It also supports video editing, allowing users to adjust video content and produce different creative versions.

How do these capabilities perform in real workflows? Zhidx spent RMB 15 to generate a 15-second video, trying it out on readers’ behalf through LibTV. Judging by the test results, Seedance 2.5 flash largely carried out the given instructions, rendering fast movements in detail and accurately while maintaining good consistency in the characters and scenes. However, garbled text still occasionally appears in fast-paced footage.

Seedance 2.5 flash currently supports clips of up to 30 seconds, offers two resolution options, 480P and 720P, and accepts up to 10 images, five video clips, and five audio clips in a single input.

The Seedance 2.5 family currently includes the flagship Seedance 2.5, a 4K version, and the newly released flash. Seedance 2.5 first appeared at this year’s Volcengine conference in June and launched on platforms including Jimeng in July, supporting resolutions up to 1080P. On October 23, ByteDance officially announced Seedance 2.5 4K for high-quality production, initially making it available for trials by selected enterprise customers, including film and television production and postproduction companies in China and abroad. ByteDance previously said that the 4K version had learned human expressions, subtle movements, and visual details more thoroughly, making it the strongest model in the Seedance 2.5 generation.

The company disclosed that the new model would roll out to platforms including Jimeng, LibTV, Updream, Creao AI, and TapNow, with its API scheduled to launch on Volcengine Ark on November 15.

01.

Testing cat parkour: smooth transitions between rapidly alternating wide shots and close-ups

But garbled text still appears occasionally

For this test, we tried Seedance 2.5 flash through LibTV, using settings of 480P, 15 seconds, and 16:9, with sound enabled. We focused on testing whether text remained stable and shots connected smoothly when switching rapidly between wide shots and close-ups.

We uploaded a street scene image and a reference portrait of an orange-and-white cat, asking the cat to do parkour along a commercial street crowded with signs, knock over a basket of apples, and then use wooden crates, a counter, and an awning to leap up onto the eaves. The prompt called for rapid cuts between wide shots of the street, close-ups of the cat’s paws, and side-on tracking shots, creating the rhythm of a chase in an animated film.

Article illustration

▲ Reference images of the cat and scene uploaded to Seedance 2.5 flash

This test also included some specific requirements for text. We added signs reading “Zhidx Fruit Shop,” “Afu Noodle House,” “Good Luck Bun Shop,” and “Shunyi General Store” to the street image in advance. We required each sign to retain its original wording across different shot sizes and after being obscured by the cat.

Article illustration

▲ Some of the prompt instructions given to Seedance 2.5 flash

As can be seen, “Zhidx Fruit Shop” remains recognizable in the opening wide shot, the shot close to the sign, and the final pullback. The words “Apples: five yuan per jin” on the price sign are also preserved. After the camera pushes in, the main sign’s character shapes remain relatively intact, without garbled text appearing.

But a closer look reveals that the more complex background characters in the distance are still missing parts of their shapes.

Article illustration

▲ Parts of the characters for “fragrant, sweet, crisp, and refreshing” are missing. The top image is a video frame; the bottom is the reference image.

Article illustration

▲ Parts of the characters for “Bun Shop” and “Shunyi General Store” are missing. The top image is a video frame; the bottom is the reference image.

The shot transitions are relatively smooth. A wide shot of the street cuts to close-ups of the cat hitting the basket, apples rolling down, and its paws stepping down, before following the cat upward. The actions before and after the cuts largely connect, without suddenly leaving viewers unable to tell where the cat has gone. In the final pullback, the fruit shop and neighboring stores are in largely the same positions as at the opening. The apple basket is still tipped over, and the scattered apples remain on the street.

The background music creates just the right sense of tension, and the sound effects are realistically matched as well, making it feel like watching an animated film.

Article illustration

02.

Official examples: making ads and performing circus acts

Multi-shot sequences show a high degree of completeness

In addition to the test above, ByteDance also provided several official examples, including a telescope ad, an earbud ad, and a circus performance.

The first example is a telescope ad. A young man in the clip walks from the cabin of a boat onto the deck, takes out and extends a telescope, looks toward a distant island shore, then waves to interact with visitors on shore. Close-ups of the character, product detail shots, and views through the telescope are interspersed throughout.

The ad has a high degree of overall completeness. The character’s expressions as he looks into the distance, lowers the telescope, and smiles in response are relatively natural. The transitions between close-ups, the view through the telescope, and the final pullback are also relatively smooth. Reflections on the sea, the wooden cabin, and the brass telescope body all have convincing textures. The character’s appearance and clothing show no obvious changes after multiple cuts.

Article illustration

▲ A telescope ad generated by Seedance 2.5 flash (video source: Volcengine)

The second example is an earbud ad. A woman opens an earbud case, puts on the earbuds, and then dances amid the crowd. The clip alternates between product close-ups, the character’s expressions, and full-body movements.

The character’s performance in this ad is relatively nuanced, with natural changes as she closes her eyes, opens them, and smiles; facial close-ups show no obvious stiffness. Putting on the earbuds, raising a hand to touch them, and dancing connect smoothly. Even when the shot size changes rapidly, there are no obvious continuity errors in the character’s movements or clothing. Close-ups of the earbuds and earrings are interspersed relatively naturally with wide shots of the street. The background detail of a man scattering papers and continuing to pick them up at the end also carries through. The clip as a whole already feels like a fairly complete advertisement.

Article illustration

▲ An earbud ad generated by Seedance 2.5 flash (video source: Volcengine)

The third example is a circus performance. Continuous circus acts place particularly high demands on shot transitions.

The shot transitions in this video are also relatively natural: it moves from a high-wire act to juggling on the ground, then follows a lion through a ring of fire. The switches between acts are tight, and the spatial relationships on stage are relatively clear. The lion’s takeoff, movement through the air and the hoop, and landing on the platform form a continuous sequence. Its body and limbs adjust with the jump, and its mane and tail change accordingly. Warm lighting, flames, and stage shadows maintain a unified atmosphere. The entire sequence conveys the speed of following the action while avoiding obvious abrupt changes in the scene.

Article illustration

▲ Tightrope walking, juggling, and a lion jumping through a ring of fire, generated by Seedance 2.5 flash (video source: Volcengine)

03.

Conclusion: the new model’s quality is a pleasant surprise

Seedance 2.5 and 2.5 flash target different use cases

Seedance 2.5 flash retains good short-video generation capabilities while lowering prices. Judging from our experience in this test, its output is a pleasant surprise: it can handle AI video production for streaming platforms and also has the potential to make commercial ads and action films, giving creators seeking value for money a new option.

According to ByteDance’s positioning, the flagship is intended for tasks with higher demands for audiovisual quality and creative expression, while Seedance 2.5 flash targets frequent creation and batch content production. Creators can also allocate their budgets according to the characteristics of the different models: generate segments that require repeated experimentation with Flash first, then choose the flagship for shots with higher image-quality and detail requirements, putting production spending where it is needed more.

This article comes from “Zhidx,” and was written by “Sun Xiaowen.”

Related topics

After reading

Join the AINRK community

Read how to join the community and connect with people working in AI.

Community information (opens in a new tab)