Seedance 2.5 Real-World Test | Is It Really Worth Paying a High Price to Use?

When the production cost of a 30-second 1080P video approaches 156 yuan, availability rate becomes the real threshold that determines whether it can be integrated into the workflow.

As long as the character orientation in the generated video is wrong, the entire clip needs to be recreated. Calculated based on the current premium membership tier of Jimeng, a 499-yuan package can only generate three complete 30-second 1080P videos. As the model's capability extends by the second, creators have also started to calculate costs on a per-generation basis.

The 36Kr AI Evaluation Team invited 27 professional evaluators to conduct a one-week high-pressure test centered on *Journey to the West*. This test covers complex plots, multi-character dialogues, actions, group scenes and ultra-long videos. We also conducted a full review of video interviews and written feedback. This time, we focused on observing how it adapts to real workflows: how much effort needs to be invested in material preparation; how many usable shots can be retained from a single generation; and how much cost has to be borne after a single failure.

36Kr AI Evaluators: Covering AIGC content creators, heads of AI manhua drama teams, creative directors of 4A advertising agencies, CG R&D specialists, tech industry observers, top Bilibili UP owners, and practical experts in AI video production implementation across various industries.

First, the conclusion:

● The local editing feature effectively reduces error costs.

● To make the instructions more compliant, more work is required in the early stage. Characters, positions, lines and error-proof conditions all need to be clearly specified.

● The efficiency dividend first goes to professional users; ordinary users are still constrained by their prompt engineering skills and trial-and-error budgets.

01

30 Seconds, the First Time It Looks Like a Proper Scene

When a narrative director from an AI manhua drama team was producing the *Daughter Country* segment, his focus has shifted from "whether a single frame is visually appealing enough" to "whether a plot segment is logically valid".

He set the queen, Tang Sanzang, Sun Wukong, Zhu Bajie, the palace and the city gate as reference materials respectively, assigning clear responsibilities to each material. When writing prompts, he would specify the characters' pauses, gestures, prayer beads and line of sight.

The 30-second generation duration is exactly what changes this sense of delivery. In the past 15-second mode, creators often had to wait for the previous segment to finish generating, then continue making the next segment through feedback. The 30-second duration allows more plot details to be retained in a single generation, so character relationships, lighting and space do not need to be edited and spliced in subsequent steps.

A short drama creator said frankly that in the past, to make a 1-minute draft, he had to splice multiple 15-second materials, but now he can cut out the prototype with two or three 30-second outputs, and the production process is much smoother.

Another source of this smooth experience is that after the number of reference materials increases, creators can separate and constrain characters, scenes, props and transition frames, the number of variables that the model needs to complete on its own is reduced accordingly. It does not make all shots pass at one go, but increases the proportion of shots that can be retained.

A film concept director put it plainly: "There are slightly fewer discarded shots, and slightly more usable shots."

The return on generation duration is most stable within 30 seconds. When too much information is packed in, the model will automatically cut down plots, or use unnatural acceleration to squeeze the content into 30 seconds. Another professional creator tested an 87-second video, and the final estimated shot availability rate was only 40%, with character image drift appearing in the second half of the video.

For short dramas, manhua dramas, concept films and advertising samples, 30-second videos are already close to a usable production unit; while ultra-long generation is currently more like a technical demonstration.

02

There is a Clear Threshold Between Being Compliant and Being Truly Understood

Multiple evaluators mentioned that version 2.5 is more compliant now: if you specify the shot duration for a few seconds, most of the generation results can be executed according to the rules; as long as the instructions are specific enough about when the character pauses, where the hand is placed, and who speaks first, the execution stability is significantly improved. A high-frequency creator stated that when he uses detailed Chinese prompts, the success rate in his personal samples exceeds 90%.

But being compliant does not make prompt writing easier. To avoid errors in multi-person plots, a director of an AI manhua drama team will clearly specify who says which line, that the lines must be fully delivered, and where the action stops. He calls these contents error-proof conditions. An animation creator also found that: the more the model can match the input, the more creators need to fully predict the shots, actions and risks before generation, and the workload in the early stage increases accordingly.

Seedance 2.5 has become more compliant, but it still does not truly understand human intentions. It can already complete action and time requirements more accurately, but when facing abstract expressions such as sense of privacy and commercial atmosphere, deviations in understanding may still occur.

When professional creators describe private lighting, the model may directly interpret "privacy" as heavier shadows; when users in the commercial analysis field input abstract commercial concepts, the picture tends to be scattered; negative constraints occasionally fail, and compliant fighting and flame shots may also trigger secondary review. To put it more plainly: its language comprehension capability is not good enough.

This is the alignment problem between language and shots. Specific instructions can be split into time, position, action and object, so the model has a clear execution target. Abstract expressions usually contain emotions, narrative purposes and visual metaphors at the same time. If any layer is processed literally, the picture will deviate from the intention. If multiple lines of dialogue, continuous actions, shot switching and negative restrictions are added within 30 seconds, the prompt will quickly become a small-scale shooting plan.

The 5000-character input upper limit further amplifies this burden: for a single shot, 5000 characters are sufficient, but for a complex 30-second plot, creators often need to make trade-offs among performance details, scene continuity and error-proof conditions. The ultra-long mode still uses the same upper limit, making the problem more obvious. The model provides a longer canvas, but users do not get a proportionally expanded control space.

03

You Start Spotting Flaws After the First Glance

In the first evaluation, picture texture, movement and camera movement were the core dimensions to judge the progress of the model, but in the actual use process, creators' viewing patterns have changed. An AI action director summed up this change as shifting from "whether it looks realistic" to "whether the flow is smooth".

In the action test between Sun Wukong and Erlang Shen, sparks, air currents, follow shots and motion blur can easily create a strong first impression. When you slow down to observe, there is occasionally a missing segment between the stick swing, block, force bearing and rebound. A CG action pre-production head reminded that "looking fast" and "the action is really fast" are two different things. Special effects can connect the visual perception, but cannot replace the causality of actions. For editors, shots missing intermediate links are difficult to integrate into the final film.

This is also a key turning point in user experience. In the early stage, as long as AI videos had one or two amazing shots, they were enough for demonstration. After entering the workflow, the evaluation unit becomes the availability rate of the entire segment. Whether the lines are wrong, whether the props are deformed, whether the environment is consistent, which door the character walks into, these details together determine whether a 30-second video can be retained. Seedance 2.5 makes the first glance more satisfying, and also makes the second round of review more rigorous.

04

How Much Does a Failed Shot Cost

Pricing is the most consistent consensus in all interviews. Professional users are willing to bear higher costs, on the premise that the shot utilization rate is high enough, and the saved production time can cover the generation cost.

Calculated based on the premium membership tier adopted in this round of test, 499 yuan corresponds to 6160 points. A 30-second 720P video consumes 780 points, which is about 63.19 yuan per generation; the newly added 1080P tier consumes 1920 points, which is about 155.53 yuan per generation. The costs of the two tiers are as follows:

Figure

This accounting actually affects the operation mentality. Lightweight users once mentioned that even if the character orientation is only slightly wrong and other contents are perfect, the entire result may lose its use value. After Seedance 2.5 launched the local editing feature, the situation that small problems lead to the scrapping of the entire video has been alleviated, and creators can continue to modify the problematic segment. The problem has also changed from "can it be modified" to "can it be corrected in one attempt and how much additional modification cost will be incurred".

Interviewees also put forward improvement directions: some hope to generate low-resolution previews first, and pay for rendering after confirming the camera movement and plot; some hope that points can be refunded for obviously failed results; both demands point to the same gap: the generation cost has entered the price range of professional tools, but error handling is still stuck in the stage of full-clip random generation.

05

Professional Users Get Dividends First

Seedance 2.5 presents a seemingly contradictory experience: the product interface is easy to get started, and the community provides a large number of case references, so new users can quickly generate a good-looking video. But the threshold to stably reproduce