A Conversation with 海辛: From Medical Student to Leading AI Artist

Summary

• AI video has no SOTA for now, so the window for startups remains open. 海辛’s view is that precisely because no single player has cracked every direction—可灵 is top-tier in realism, 海螺 wins on stylized animation and camera control, and PixVerse on speed and transitions—niche applications such as video typography, automated lighting, and effects templates “could all work.” Pika’s focus on video effects is a case in point.
• The demand side has shifted from “buying the AI gimmick” to “buying the result.” After Runway Gen-2, the Lenovo × GQ ad still sold itself primarily as “an AIGC film”; this year, 伊利金领冠 and 霸王茶姬 “came chasing the effect, not the fact that it was made with AI.” His conclusion: “AI has gradually evolved from a gimmick into a real tool”—and their quotes are rising, not falling, because AI creates incremental effects traditional production cannot deliver, rather than merely cutting costs or improving efficiency.
• Three technical breakthroughs are making character-driven narratives easier to produce. Character consistency has improved; image quality has improved; and style compatibility now extends to 2D illustration and collage. The remaining gaps are equally clear: scene continuity across reverse angles and large-scale camera movement—which is why 《故宫猫猫上班记》 was “basically all locked-off shots.”
• His spending is a real-world vote for the tools. His favorite, and the primary image base for his work, is Google’s free Whisk (pronounced “Visk”), an Imagen 3 product, which is “much better than the leading AI image tools in the industry today.” On video, he pays for the highest-tier plans from 可灵, 海螺, and PixVerse. 即梦 can generate accurate Chinese text and often outperforms open-source workflows.
• Legacy experience compounds in the AI era; it is not baggage. One of OpenArt’s most popular ComfyUI workflow authors, 瑞, has 15 years of game-art experience. People with film and television backgrounds understand storyboards and shot sizes; newcomers are more likely to stitch together “one wide shot after another.” AI companies are changing too—from pure algorithm engineers in 2022-23 to more people with art and creative-production backgrounds this year.
• His creative method has undergone a shift worth recording. “For the first 28 years, completion was always more important than perfection”; since last year, he has found that “fighting with myself is worthwhile.” 荣慧’s analogy to the DeepSeek R1 paper resonated with him: “The longer you think, the better the final result.” The Palace Museum cats video, which surpassed 100M views, was the product of that persistence: every shot had to add more emotional value than the last, with no more than 10 seconds produced per day over 2 weeks.
• The disappearing profession of motion-capture artists is a microcosm. A professional optical stage and motion-capture suit once cost about RMB2M; now “any camera with AI behind it can do the job.” But when the barrier disappears, “anyone can use their own movements to drive their own actors”—substitution and expansion at once. That logic is also the extreme version of his own ambition: “Maybe in the future I won’t need to ask people for help or ask for money anymore,” and one person will be able to make a film.

Deep dive

1. From Medical Student to Full-Time AI Filmmaker: The Work Is the Résumé

• 海辛 describes himself as a director trained in new technology. He began making AI shorts in 2023 and now works full-time. His signature projects include 《故宫猫猫上班记》, which surpassed 100M views across platforms, a city-promotion film for Beijing Daily made with 阿文, and 《枕着光的他》, the first AI-animated music video for the CCTV Spring Festival Gala in the Year of the Dragon.
• Recent commercial work includes an AI-produced Lunar New Year short for 伊利金领冠 that recreated the look of 3D animation, and a February TVC for 霸王茶姬 made after he and 阿文 joined the 3D art studio Sing Vision. Their job was to turn a 3D white-model style into an AI oil-painting look.

2. The Palace Museum Cats Method: Avoid Faces, Pay Homage to Satoshi Kon

• Choosing cats over people was a clear-eyed technical decision: “When we watch human performances, it’s easy to tell whether a person or AI is acting, because we’re extremely sensitive to human faces.” Avoiding human characters and making non-human animation may make it easier for audiences to perceive the work as high quality.
• The narrative structure directly pays homage to Satoshi Kon’s 《Millennium Actress》 and 《Paprika》. The female protagonist runs through changing settings, shifting with the shot size into different eras and films—“I rarely see storytelling this dazzling.” So when the kitten passes different landmarks on its way to work, the film switches into different styles.
• In pre-production, they used Shot Deck to search for classic shots as references and “replace the protagonist with a cat.” But the plan was repeatedly overturned during production: AI could handle some shots but not others. “The whole process was one of constant revision.”

3. No Excuses in the AI Era: Every Shot Must Add Emotional Value

• Traditional post-production could hide behind “that’s all the footage the director shot.” In the AI era, “there’s no such excuse—you can generate the footage you want yourself, so there’s nowhere left to hide.” For every shot, he kept asking: Does the next shot give the audience more feeling? Will they get bored?
• The cost was speed: over 2 weeks, they produced no more than 10 seconds a day. “Making 10 seconds was already something that made me extremely, extremely happy.” In AI circles, a 1-minute video can normally be produced in about a day.
• Beijing Daily, the client, imposed almost no requirements: just make it 1 minute 30 seconds and include plenty of Beijing elements. “That meant we could let ourselves go.”

4. Three Breakthroughs in a Year—and Two Remaining Gaps

• The first breakthrough is character consistency. In the past, uploading a photo and asking a character to smile or turn its head meant “it stopped being me”; a Chinese elderly person speaking could even become “an old American grandfather.” Now the character ID generally holds within a shot. “Only once characters are stable can you tell stories centered on characters.”
• The second breakthrough is image quality. Previously, footage became blurrier with each generation, and of a 5-second clip, “maybe only 1 or 2 seconds were usable.” Image quality has now improved substantially. The third is style compatibility: a 2D hand-drawn tarot card image that previously “probably couldn’t be made to move realistically” can now work in 2D, 3D, abstract, and collage styles, with fairly high quality.
• The gaps are equally concrete: complex camera movement and scene stability. AI can generate a shot from his camera position, “but it can’t generate the reverse-angle shot from my camera position.” His analogy is whether the front and rear cameras on an iPhone see the same space-time. Long takes like those in 毕赣’s 《路边野餐》 “still produce all kinds of errors.”

5. Sora at One Year: A New Coordinate System

• The host’s memory anchor is being jolted awake by a flood of posts at 4 or 5 a.m. on the third day of last year’s Lunar New Year: “It felt like something from outer space.” A year later, “everything seems to have become second nature”—but in retrospect, it was a year of seismic change.
• This episode adds the perspective of a frontline creator. Earlier episodes featured Luma AI product manager Barclay’s 20-question retrospective and 昌胤, founder of Vozo AI and a former Google X researcher whose product topped both the daily and monthly Product Hunt rankings and reached $1M ARR in 6 months. 海辛 brings the frontline creator’s view.

6. Voting with Money: Whisk First, No SOTA in Video

• His favorite product is Google’s Whisk, pronounced “Visk,” an Imagen 3 product. It is free, and its realistic generations are “basically indistinguishable from real photos.” Most of the image base for the Palace Museum cats was made with it. “It’s much better than the leading AI image tools in the industry today.”
• His exact assessment of video: “For now, there still isn’t anything at SOTA level.” Every model has its strengths and weaknesses: 可灵 is strongest in realistic scenes; 海螺 is better at stylized animation and camera-motion control; PixVerse has the smoothest transitions and is best at judging continuity of movement. He subscribes to all 3 at their highest tiers.
• An under-the-radar recommendation is 即梦’s image generation. Most people know it for video, but it can “directly generate accurate Chinese text,” making it useful for posters. It also incorporates ControlNet-style contour, skeleton, and edge references. “A lot of the time, it works even better than my open-source workflows.”

7. The Two New Features That Moved Him Most

• PixVerse’s Auto Sound initially did not interest him because he was used to adding sound after finishing a video. But for the growing volume of short, single-shot creations—“say, snow being poured over Mount Fuji, with only one shot lasting a few seconds”—going into an editing app afterward is actually cumbersome. The feature fits lightweight creation exactly.
• 海螺’s director mode uses professional filmmaking language to control shots. Upload an image of the Oriental Pearl Tower and say, “Move a little to the left, then lower a little, then rotate around it.” “I’ve always wanted to control the camera inside an AI model. It hit my needs exactly.”

8. No SOTA Means a Startup Window

• He keeps returning to the question: “Why hasn’t SOTA emerged after all this time?” The answer is also the opportunity: because no one company has fully solved every direction, “there may be room for startups.” There are too many video applications—text packaging, automated lighting, effects templates—and “I think all of them could work.”
• If he started a company, his early idea would have been “making storyboards on a canvas, then making video on top of the storyboard.” It failed at the time because generation was too slow and expensive, but “a video product like that will definitely appear in the future.” What he wants most now is to combine the real and virtual: shoot real city establishing shots, then use AI to add effects such as “being swallowed by a tsunami or hit by a meteorite.” “AI work combined with reality will attract more eyeballs, even if it may look less cinematic than a purely filmic piece.”

9. Two Relationships with Vendors: Power Creator and Advisor

• He is a power creator for multiple video-model companies and submits requests directly. The two issues he has repeatedly raised across the industry are character consistency and 1080P. “They may put those requests on the roadmap.”
• The deeper relationship is advisory. He is now an advisor to 美团妙刷, a product built around ComfyUI-style innovation, contributing to everything from creative direction and execution to some promotional work. Outputs include turning the plush toy Shirley Cat into a drawn-over image and making an animated Harry Potter school-admission poster.
• The changing makeup of teams is a signal. The domestic AI teams he encountered in 2022-23 were “basically all developers or algorithm engineers.” This year they are becoming “increasingly mixed,” with many people from art, creative production, and video-effects backgrounds joining. He cited 归藏 as an example. “The field still needs to absorb people with deep professional knowledge.”

10. A Full Model Ranking: The Frontline Creator’s Scorecard

• The top 3 image models are Midjourney, which has been popular since 2022-23 and is “extremely aesthetically strong”; Google Imagen, which came from behind—“people are realizing that its output is actually better than Midjourney’s”—with more realistic texture; and 即梦, which fills the gap with Chinese text and reference controls.
• In video, 可灵’s realism is “absolutely first-tier”; 海螺 is first-tier for stylized animation and camera control and is “definitely the best”; PixVerse is the fastest, with the best transitions and keyframe animation.
• Luma’s website UI is “in a class of its own,” and its animation feels exceptionally natural. Overseas creators have used it for Miyazaki-like work that is “really quite close to what Studio Ghibli would make.” Runway has been quiet for a long time, “but it was still the first heavyweight in the top tier.” Pika took a different route by focusing on effects—turning people into balloons and flying them away, or crushing them. “Those effects were pioneered by Pika.”

11. The Market Is Learning to Watch the Story: From Mood Pieces to Narratives

• His view last year was that AI worked best for trailers that emphasized atmosphere and did not require strong logic. His updated view is that films centered on characters are “gradually becoming easier,” although still not as easy as shooting with real people.
• Evidence came from the Narrative track of the global AI creative competition Project Odyssey. A Christmas skeleton scare story that placed near the top made him realize, “I was actually paying attention to its story, not whether it would break.” But he retains his earlier conclusion: atmosphere pieces, teasers, and concept films “may still be where AI has the longest runway today.”
• The Palace Museum cats’ narrative quality was partly achieved through avoidance. They deliberately did not use the same cat and kept switching styles, so viewers would not say, “Why does this cat have an extra stripe on its face in the next shot?”

12. The First Job: Lockdown, Street Dance 5, and Two Moons Generated on Set

• His first commercial AI film was a teaser for 《这就是街舞》第五季, made at Shanghai Media Group in May or June 2022, with Midjourney used for production design. AI accounted for roughly 80-90% of 李承铉’s background; 刘雨昕, 王一博, and 韩庚’s backgrounds were more uneven, at 10-30%.
• The trigger was the pandemic. The project was scheduled to start in March or April and launch in August, but 2 months were lost to production shutdowns and the budget was insufficient. They did not reach Hangzhou until June, leaving AI as the only way to create concepts and deliver quickly.
• The real validation was communication efficiency. Previously, an art director would hear a brief, go back to draw a draft, and return for feedback. Now, when the director said, “Could we have 2 moons?”, 海辛 could generate them on the spot and quickly finalize the concept with the performers and director. The client did not care how much AI was involved. “Everyone was really looking at the result.”

13. Three Client Phases: From Gimmick to Tool

• The first phase was game companies, using AI to improve the efficiency of art-asset production. They approached it experimentally and “didn’t have especially strong resistance.” The second began after Runway Gen-2, when advertising agencies started selling AI as the attraction. The Lenovo × GQ collaboration was promoted primarily as “an AIGC film.”
• The third phase is now: “Everyone is chasing the result, not the AI.” A landmark turned into a plush toy went viral; the client wanted that effect, “and AI happened to be what made it.” 伊利 wanted a 3D feel, while 霸王茶姬 wanted an oil-painting style. His summary: “AI has gradually evolved from a gimmick into a real tool.”
• On the argument that AI is making vendors compete themselves into ruin, his counterexample is blunt: “Our quotes have actually gone up.” AI is not just about efficiency and cost. “AI creates incremental value.” In a body-wash ad made with a British friend, the male lead moves between piles of oranges, the seafloor, and a forest. The transitions would be “basically impossible with traditional production.”

14. Legacy Experience Compounds: 瑞’s 15 Years and Shot-Size Awareness

• There are 2 pieces of evidence. First, 瑞, one of the most popular workflow authors on the ComfyUI workflow platform OpenArt, has 15 years of game-art experience.
• Second, people with film and television backgrounds are more likely to make good AI video because they understand storyboards and shot sizes. Inexperienced creators are “very likely to make an entire film by stringing together one wide shot after another,” with no sense of rhythm. He added a hedge: that gap “may be filled over time as everyone spends more time making things.”

15. AI Generation Is the Final Cut: What Clients Find Hardest to Accept

• The recurring friction point is that clients love the overall atmosphere but want to adjust details: “Make the horse walk a little more slowly; don’t have motion trails during the transition.” In a traditional workflow, you could reopen the project file. AI video has no layers—“once you make a change, you have to go back and do it again.”
• A new workflow convention has emerged. They still prepare creative proposals and storyboards as usual, but explain upfront that “some things probably can’t be changed; if you want to change them, you can only replace them.” Based on experience, “it’s something that will keep changing flexibly.”

16. Reproduction Is Not Enough: Collage Experiments and the Ontology of Imagination

• He and 阿文 are making a collage-style animation. Halfway through, they asked themselves: “If we’re only using AI to reproduce the look of collage animation, what’s the point of making this short with AI?” They then added things traditional collage could not do: elements moving between different collage paintings, breaking through the 2D plane into 3D and returning to 2D. “We have to do more than reproduce effects from traditional production; we also have to create new ones.”
• The underlying motivation goes back to 《Up》. The scene where balloons lift the house “didn’t move me because of the story; I was simply moved by the imagination.” His explanation of why imagination moves us: “It lets us instantly leave the life we’re living… to feel free, as if the outside world contains infinite possibilities.”

17. The First Crossroads: Five Years Forced into Medical School

• After watching 《Up》 in high school, his dream was “to become a screw in the Pixar animation company. I felt that would be enough to make me happy.” After the gaokao, he wanted to study directing, but his family forced him into West China School of Medicine at Sichuan University. “That may have been the first argument I ever had with my family.”
• His retrospective on compromise is candid. The lack of confidence had 2 layers: “One was uncertainty about whether I could support myself as an artist. The other was a lack of confidence in the field itself—in my era, art was something people with poor grades went to study, although I no longer think that way at all.”
• The turning point came from a question he asked himself during those 5 years: “Did I want to study art because I wanted to succeed?” The answer was happiness, not success. “Even if I don’t succeed, this is what I want to do.” He applied to art school in his fourth year and received an offer in his fifth. This time, he argued with his family but did not compromise.

18. The Use of the Useless Is the Greatest Use

• His list of “useless” activities during medical school includes founding a radio-drama club and joining 梁中和’s Plato Academy at Sichuan University, where they read Plato every 2 weeks. “The time I actually spent making my portfolio was basically just that winter break in my fourth year.”
• His decision rule is not to choose activities from a utilitarian perspective. He realized he had multiple core needs: beyond grades and work, he wanted “to feel that life was interesting, to create, and to feel that I was a decent person.” “Those scattered things actually made me more three-dimensional and more changeable.”
• The host connected this to a constructive use of high sensitivity: “Sensitive isn’t necessarily a bad word.” Sensing what you need and then meeting that need “is also a way of developing yourself.” 海辛 added: “I’m trying not to let labels or professions determine which direction I’m supposed to develop in.”

19. Advice for Beginners: Completion First, Then Learn to Persist

• The first step is always to start making things: “Do something small first. Don’t let the difficulty you imagine scare you off.” Then post frequently on social media to document the process. “It doesn’t matter if it isn’t good enough yet.”
• But his method changed last year: “For the first 28 years, I think completion was more important than perfection. Since last year, I’ve found that fighting with myself is worthwhile.” He keeps asking: Is this really what I want to express? Is it honest?
• 荣慧 compared this to the DeepSeek R1 paper, and 海辛 said it might apply to him too: “The longer you think, the better the final result.”
• Persistence has a cost, so balance matters. The Palace Museum cats was “the most painful of all the AI shorts, but also the one we’re happiest with now.” Still, “80-90% of the time I’ll take the other route.” His current rhythm is to alternate between “running lightly for a while, then holding back a big move.”

20. Studying in the US: Packed Schedules, Burnout, and Accent as Personality

• His schedule was classes Monday through Thursday and the set from Friday through Sunday. In his second year, he became the lead production designer for the entire class’s short films, until obvious burnout set in. “This was probably the first time in my life I discovered the limits of my body.” After slowing down to lead only 1 film a month, “the rhythm became much better, and the quality of the work improved a lot.”
• His first lesson in freedom was boundaries: “The people around me didn’t care what I was doing anymore.” Since every class was related to creativity, “it felt as if my life had no dead time.” The second lesson was tolerance. His best friends were an Iranian Muslim effects artist and a Brazilian Jewish friend. “I was encountering people, not a particular title.”
• The small thing that affected him most was accent. An American friend told him, “There is no standard English. My accent is part of my personality.” From then on, “standard” stopped mattering to him, giving him more room to be free.

21. From a Mary Short to SMG, Then AI: How the Bridge Was Built

• The one-line premise of his graduate short was: “If the Virgin Mary lived in the contemporary world and learned she was pregnant with Jesus, might she choose to have an abortion?” It was a black comedy, because “in the biblical stories, no one ever asked whether she wanted to be pregnant with Jesus.” His current analysis is that he was wrestling with whether feminism or other people’s expectations of him mattered more.
• The short took him into the Shanghai International Film Festival’s project-development training camp. He decided to stay in Shanghai and joined Shanghai Media Group as a full-time director focused on new technologies: the metaverse, 3D scanning, digital humans, and XR virtual production. Unreal Engine ran on Nvidia GPUs, “so when AI appeared, it came from exactly the same soil. I encountered AI immediately.”
• The final leap came during pandemic lockdown. He saw artists using AI to make NFT collectibles and wondered, “Can this thing also be made with AI?” After taking a course, he “realized AI was much more compelling than NFTs,” and eventually focused exclusively on AI.

22. The Fantasy of Making a Film Alone—and the Disappearing Motion-Capture Artist

• His underlying need is unambiguous: “I want to make a film by myself.” In the US, he discovered that filmmaking required “a lot of people, a lot of money, and a lot of communication, with compromise built in… None of that is entirely related to creation.” So every technology that lowers the cost of filmmaking is his technology. Virtual production lets you shoot a sunset “in a game environment that stays at sunset for 24 hours”; AI is the ultimate form. “These technologies gave me the fantasy that perhaps one day I won’t need to ask people for help or ask for money anymore, and can create without any compromise to the work.”
• But he does not romanticize working alone. The pleasure of a set is real. While holding up a flat set wall 2 or 3 times his height, “I realized for the first time that I was creating something much bigger than me… My act of creating gave my life a huge extension.”
• The profession clearly disappearing is motion-capture artist. A top studio once cost “around RMB2M per stage”; now “any camera with AI behind it can do the job.” But he sees expansion rather than a zero-sum loss: “Anyone can use their own movements to drive their own actors.” More people entering may mean more possibilities entering with them.

23. An Inspiration System and “Scientists of AI Art”

• He deliberately keeps his inputs separate from the industry: swing-dance events in Shanghai, evening classes in spices and fish-lantern making. “Not thinking about the thing I’m working on is a very strong form of relaxation, and everything that enters me eventually gets digested to some extent and becomes part of my work.” A trip to Japan last November directly produced a series of videos, including animals jumping out of electronic screens and Kirby with Mount Fuji.
• He and 阿文 describe themselves, in 大聪明’s words, as having an industry-leading role: “scientists of AI art—continually discovering where new forms of expression in art might be.” Even now, “we feel that the potential of AI work at every stage still hasn’t been fully realized.”
• His self-assessment of engineering versus art is surprising. He considers himself stronger at engineering, having studied science in high school. “Artistic creation is harder. It isn’t something I can make good simply by building a framework. There are many intuitive and reflective moments.” So while creating, he consciously tries to “put down the plan” and avoid arriving at the most obvious, easiest answer.

24. Women Creators and the Energy Field of Demo in

• There are still few women speaking publicly in the industry. His observation is that the women around him often impose more demands and reflection on themselves: “I only made this small thing—can I stand on stage?” His encouragement is direct: “The people standing on stage aren’t actually that much better than them.” They should have the courage to step up. “I have this self-doubt too, and I’m gradually overcoming it.”
• He and AJ run an “editorial department” called 整活编辑部 at 通往AGI之路, exploring new effects and publishing tutorials, while community remixes feed new ideas back to them. He also runs the community Demo in, which held its first AI Hacker House at Caohejing last month, attracting 50 or 60 people. The entry requirement was that everyone had to demo their own work.
• The rules repeated by the host are themselves part of the product: no one may be called “teacher” or “expert”; introductions cannot mention one’s university or company; and demos should be met with encouragement rather than criticism. That night, “half of the engineers who demoed were women—and, speaking frankly, I couldn’t understand them. The level was extremely high.”