AI films will ultimately return to the very essence of “film” itself.
AI movies have truly arrived.
The recently concluded Hundred Flowers Awards for Popular Films has set up a dedicated category for AIGC imagery for the first time, receiving 2038 submissions within just one month. Works such as Tales of the Paper Blade in the Wasteland and Hot Blooded Dachen Island have been successively launched on iQiyi, Tencent Video and Youku with official distribution permits. Sanxingdui: Future Past produced by Bona Pictures has obtained its official public screening license, just waiting for a good opportunity to hit theaters. Overseas, Hell Grind has just returned from its Cannes screening, and the team will release a new 114-minute feature film a few months later. Christopher Nolan’s Odyssey has just been released globally, while the AI-generated version Odysseus: Fallen has already been completed.
Tales of the Paper Blade in the Wasteland, Hot Blooded Dachen Island, Sanxingdui: Future Past, Hell Grind
Behind the buzz, however, most people vastly underestimate the complexity of AI filmmaking. It is by no means something that can be completed just by typing a few prompts, and its complexity is no less than that of traditional live-action films shot with real scenes.
“The price of using models has surged more than 20 times.” Li Muyang, a creator whose work Dream Agent was shortlisted for the AIGC Feature Film Unit of the Beijing International Film Festival, told Shenxiang that individual creators are finding it increasingly difficult to afford production costs.
Liu Haiyang, the AI director of Tales of the Paper Blade in the Wasteland, said that AI cannot meet the standards he expects, and the presentation of character performances, special effects and action scenes still has a stiff feel, so he can only wait for the next upgrade of the model.
Yang Liang, a director who has created many AI films including Confession, Hello, Big Head and Orange, also mentioned that AI generation produces diverse results for different users, and how to maintain the visual unity of the entire work during team collaboration has always troubled him.
Another industry practitioner who is working on the AI content part of a live-action drama also mentioned that current models have clear upper limits on capabilities, and it is very difficult to achieve a good combination of real people and AI, as real human images placed in AI scenes will inevitably undergo uncontrollable changes.
Obviously, AI can indeed “make films”, but it is far from reaching the level of “making films well”.
Three Uncertainties of AI Films: Uncontrollable Costs, Unfixed Construction Period, and Depreciating Experience
“Drawing cards has turned into drawing blood.”
Li Muyang, the creator of Dream Agent, described his current experience of making AI films to Shenxiang in this way.
He recalled that when he used Sora last year, the generation cost of a 15-second video was only a few cents. “At that time, I didn’t feel bad about the cost. Drawing cards was like playing a game of lucky draw: it’s better if you get a good result, and it’s normal if you don’t.” But as mainstream video models continue to upgrade, prices have risen significantly. Especially after the launch of the latest Seedance2.5, the price for a 720p 15-second video has reached 39 yuan, equivalent to 2 yuan per second. “One click, and the cost of a lunch is gone. At present, we can only use the relatively more cost-effective MiniMax H3 model as a substitute to keep costs under slight control.”
In addition, according to the data from the research report released by Orient Securities in May this year, the unit price of HappyHorse-1.0 and Keling 3.0 is about 1.63 yuan per second and 1.52 yuan per second respectively (converted from the API price for generating a 1-minute video). In less than a year, the prices of almost all mainstream models have risen to a new level. For individual creators without the support of companies and teams, the past method of “generating while brainstorming ideas and making mistakes while trying” has become increasingly unaffordable.
Prices are getting higher and higher, but the success rate has not increased proportionally.
The creator of Hell Grind, the AI feature film screened at this year’s Cannes Film Festival, used real data to show the core pain point of “drawing cards” — for just the first 22 minutes of content, 10,701 images and 16,181 video clips were generated, and only 253 qualified shots were finally included in the finished film. The prompt for a single shot can even be as long as 3,000 words, and every shot that is finally retained may have gone through more than 60 rounds of trial and error.
The uncertainty of drawing cards directly leads to the uncertainty of costs. Unlike traditional film and television production, creators cannot relatively accurately estimate the production budget based on the script, storyboard and shooting cycle before the filming starts. If great importance is attached to high quality and perfect details, the cost can be “unlimited”, and the project will exceed the affordable range in the middle of production due to the continuous accumulation of computing power bills.
When this uncertainty is amplified in a feature film with hundreds or thousands of shots, the problem of inconsistent pictures will surface.
The facial details of each character in the feature film, the fold direction of a piece of clothing, and even the relative position of a lamp in an indoor scene need to remain stable across different shots. However, AI-generated images are prone to de The “AI flavor” may be tolerable in short dramas, but it will appear particularly abrupt in feature films
Liu Haiyang, the AI director of Tales of the Paper Blade in the Wasteland, has a deep understanding of this. He mentioned that one of the main difficulties of making feature films with AI at this stage is “spatial consistency in multi-person scenes”: for example, when AI generates an ordinary scene where three people are sitting face to face chatting in a tavern, the positional relationship between the characters may change after a few seconds.
Tales of the Paper Blade in the Wasteland
Director Yang Liang also mentioned similar problems. In contrast, outdoor scenes have a certain margin for error, because the natural environment does not have many clear spatial reference points, and the audience may not immediately notice some changes between shots. But indoors, the building structure, the position of doors and windows, the spatial orientation and the position of the characters all have clear references, and any offset of a frame will be infinitely magnified in the eyes of the audience.
In order to solve these inconsistency problems, the workload of the post-editing process has to multiply. Creators must spend a lot of time comparing whether the character’s face shape, clothing details, light and shadow direction, and even the position of props between different shots match. Once a deviation is found, they either have to go back to draw cards and regenerate, or manually repair frame by frame through post-production retouching. This process will make the production cycle more unpredictable.
What is even more frustrating is that even if creators finally “tame” a certain model, once the model is upgraded, all the accumulated experience will disappear.
Every model upgrade may invalidate the prompt skills, parameter experience and the entire workflow accumulated by creators in the past, and the production cycle will be extended accordingly. Creators cannot predict whether they will just catch up with the node of major model progress, but once they do, they have to face a difficult problem: if they use the old version, the clips with backward visual effects mixed in will be very abrupt; if they switch to the new version, they not only have to spend extra money, but also have to rework.
Li Muyang mentioned that the finished film of his Dream Agent mixes shots from multiple generations of products, from the early version of Keling, Sora, to the later Seedance. “It is actually a combination of many types of shots,” so there are differences in the clarity and refinement of many clips. This mixed use of cross-version materials will directly affect the overall smoothness and viewing experience of the film. But for Li Muyang, if all the previously accumulated video materials are reworked and regenerated, the cost will rise sharply, and it is impossible to predict how much it will increase.
This is probably the most “contradictory” part of AI films at present — AI is expected to “liberate productivity”, but this productivity at the current stage is neither stable nor cheap. It lowers the threshold for ideas to be realized, but pushes the industrialization of real AI feature films to a new level of difficulty with cruel drawing card probabilities and computing power bills.
The Greatest Certainty Still Lies in “Human Beings”
When the three layers of uncertainty in AI films are superimposed, the only thing that can be confirmed is that we still have to rely on “people”.
First of all, individual creators are using their long-accumulated experience to figure out the “temperament” of various AI tools.
Li Muyang has accumulated experience through a large number of tests, established his own prompt toolkit and skill plugins, trying to turn the originally probability-dependent generation process into a reusable personal workflow. This kind of experience cannot be achieved overnight. Only after enough attempts can creators know under what conditions a certain model is prone to failure, what prompt structure is more likely to generate relatively stable results, and how to combine several models to improve the success rate.
Similar practices are very common in the AI creator community. Many AI creators spend a lot of time studying prompts and model characteristics, and even use AI tools dedicated to writing prompt templates. Essentially, they are all doing the same thing: narrowing the distance between “what I want” and “what the model can give” as much as possible.
Prompt template sharing on Xiaohongshu
Secondly, when upgrading from individual creation to a professional team, the collaboration model of “traditional director + AI director” has become a feasible solution.
“Creators should not be led by the nose by the ‘good pictures’ randomly given by the model.” As Yang Liang said, the model is very good at providing all kinds of “wonders”, but creators must be clear about what they really want to express, and then make choices from these possibilities. If you only accept the composition, light, characters and actions that the model is easiest to generate, and constantly repeat the proven visual routines, the works will naturally tend to be stereotyped.
The model is responsible for providing possibilities, and people are responsible for deciding which possibilities are worth retaining. The better AI is at generating, the more creators need to know what should not be generated. This requires creators to have both the artistic aesthetics of traditional film and television and the ability to apply AI technology.
Tales of the Paper Blade in the Wasteland adopts the collaboration mode of “traditional director + AI director”. The former is in charge of artistic expression and performance guidance, while the latter is responsible for turning creativity into executable instructions for the model. This division of labor logic also extends to the entire team, which naturally divides into two camps under the coordination of the two directors: one side is traditional film and television positions such as directors, screenwriters and concept artists, and the other side is emerging technical types of work such as AI asset effect artists and prompt engineers. The two sides have their own responsibilities, forming a new cross-border complementary collaboration structure.
Tales of the Paper Blade in the Wasteland
Whether it is an individual or a team, it ultimately needs to land on a set of executable workflows. The core of building this workflow still relies on people to mend and bond.
Jiang Cheng, founder of Dameng Innovation, producer and chief producer of Tales of the Paper Blade in the Wasteland, told Shenxiang about their operation logic: before AI is involved, the team first locks the story worldview, script logic, character and monster assets. After AI is introduced, a strict human-machine relay assembly line is formed: people first establish assets and standards, then AI conducts production, and then people carry out secondary repair and fine polishing. For example, when facing AI problems such as multi-person panoramic shots and inter-frame consistency, the team will use traditional film and television production processes, such as manual frame-by-frame retouching, to assist in production.
Although there are many products on the market that focus on “one-stop AI video platforms” designed to help teams establish standard workflows, in the eyes of feature film creators, these products are in an awkward situation. They can indeed reduce operational friction and are suitable for fast-paced, large-scale production of AI short dramas and manhua dramas. But for feature films, they cannot make audiovisual language decisions for creators, nor can they automatically guarantee extremely high quality control standards. In addition, after adding general functions such as director’s console and infinite canvas, such platforms will also increase usage fees, which is not cost-effective for cost-conscious producers.
There are no universal shortcuts, nor ready-made templates. At this stage, every team that aspires to make AI films has to build its own “AI film and television production line” by hand, just like geeks.
In addition, what ultimately determines whether an AI film project can be launched online is production management — using rationality to hedge against uncertainty and using quantification to lock in output.
Many mature AI production teams will lock in character and scene assets in the early stage, establish fixed reference images, multi-camera visual standards and strict shot review mechanisms in the middle stage, and bring all the materials scattered in different models, different members and different time periods into the same visual system in the later stage.
The Tales of the Paper Blade in the Wasteland team also built their own workflow to precisely control the prompt design, drawing card cost, responsible person and even output compliance rate of each shot, and manage the entire production process with a refined data dashboard.
