How Much Does a 2h20m AI Film Actually Cost?

A two-hour-and-twenty-minute AI film, based on the measured scale of "The First Avenue"—three computers, eleven days, 1,029 shots, two frames per shot—has a total cost of roughly between 3,000 and 6,000 yuan based on experience. The bulk of this money isn't computer electricity or labor; it's the API call fees for AI image generation and video generation. You can treat this film as an engineering project: every character image, every location image, and every video generated burns API quota, and quota is money. The 3,000-to-6,000-yuan range depends on how many retries, how many failures, and how much rework you incur. Is video generation more expensive, or image generation? Video is more expensive—far more expensive. Character images use drawrole via Volcano Engine synchronous direct call, and based on experience, one image costs a few mao to one yuan. Location images use the empty-rolex branch of placeaiwords for pure text-to-image, at a similar price. But video generation is billed per second—even a shot that only needs five or six seconds is charged by the second, and the unit price is far higher than for images. Under the measured specifications in the material, the entire film has 1,029 shots, with one video generated per shot—this single line item is the biggest chunk of the cost. More precisely, the image-to-video generation driven by first and last frames is the most money-burning part of the entire project. The material mentions the full film runs two hours and twenty minutes, which converts to 8,400 seconds. If each second of video generation costs anywhere from a few fen to one mao, the video cost for the entire film lands at 2,000 to 3,000 yuan. So when calculating costs, look at video generation first—it determines your total budget ceiling. Do three computers require extra spending? Based on experience, if you already own three computers, you spend nothing extra—electricity is negligible. But if you don't have three computers and want to rent cloud servers, that's a separate bill. The material explicitly states the entire workflow runs in the Yuncao Academy system with the dbvideo_ table prefix; theoretically, you can deploy the program on the cloud. Three cloud servers rented by the day, at a few dozen yuan per day for a low-spec machine, would cost several hundred to a thousand-plus yuan over eleven days. But the actual practice for "The First Avenue" used three physical computers. Based on experience, bringing your own computers is cheaper because computers are fixed assets—they're still there after you're done. Cloud servers are consumables—shut them down and they're gone. So if someone asks whether to buy dedicated equipment for making AI films, my answer is: first get the workflow running on the computers you already have, then consider adding more machines. Can you cut costs further? Yes. The key to saving money lies in idempotency locks and the retry mechanism. The material specifically emphasizes that atctoscene has an idempotency lock and cannot be re-run, and that first and last frames must be generated as a pair, otherwise secondpic keeps the STCODE.png placeholder. What these two statements mean underneath is: every re-run burns money. If you don't implement idempotency control and run the same shot twice, you've paid for two videos, used only one, and the other half is money straight down the drain. Likewise, if you generate only the first frame and forget the last frame, that video generation fails, and you have to re-run the whole thing—effectively burning an extra video generation fee. So the correct way to save money is not lowering image quality but reducing unnecessary re-runs. Based on experience, a disciplined engineering workflow can keep the retry rate under five percent—meaning a 3,000-yuan base can be held to about 3,200. But if the workflow is messy and the retry rate hits thirty percent, that same 3,000-yuan base shoots up to 4,500. Are there counterexamples? Yes, and they're quite representative. Someone tried to save money by skipping video generation entirely, thinking they could just stitch the 2,058 first-and-last frames together, add music and subtitles, and call it a film. The result: audiences watched a sequence of static images flipping back and forth with no motion whatsoever—it looked like a slideshow. The material clearly states that the full film does include video generation, but vdurl contains no real video addresses across the entire film—the videos exist only in Jimeng. This actually proves that the video generation step cannot be omitted—it's not a nice-to-have; it's the skeleton of the whole film. Another counterexample: someone tried to save on image generation by generating characters and locations together in a single scene, only to find the character and background styles didn't match, causing repeated failures in the scene compositing step and burning several times more money. The material explains this clearly: image generation follows three separate paths—characters through drawrole, locations through placeaiwords' empty branch, and scenes through placeaiwords' character branch. The three paths each go through their own doors and cannot be mixed. Based on experience, keeping the three paths separate is actually the cheapest approach, because each path can be independently controlled and retried without a character failure dragging down the entire scene in a wasted run. How do you calculate your own project's cost? Here's a three-step method. Step one: count the number of shots. Suppose you're planning a ten-minute short film. At about five seconds per shot, you'll need roughly 120 shots. This step determines the number of video generations. Step two: count the number of images. Two images per shot—first frame and last frame—meaning 240 images for 120 shots, half of which are character images and half location or scene composite images. This step determines image generation costs. Step three: add the video generation cost and the image generation cost, then add a 10-to-20-percent budget for failed retries—that's your total cost for the film. Based on experience, the AI generation cost for a ten-minute short film runs roughly 100 to 200 yuan, and a two-hour-and-twenty-minute feature runs 3,000 to 6,000 yuan, in yuan, not ten-thousands. This is where the cost gap between AI films and traditional films is most dramatic. Is there any way to push costs down further? Yes. Three details can be dug out of the material. First, the imgstyle field switches between realistic, ancient, anime, and modern styles with a single value, meaning you can reuse one template across four projects without rewriting prompts for each style—this saves development and debugging time, indirectly saving money. Second, three prompt templates cover the three image generation paths—character, location, and scene. With fixed templates, error rates drop, and retry counts fall. Third, the database table design makes records enumerable, re-runnable, and replaceable—swapping an actor changes one row; adjusting a shot doesn't require redoing everything, which directly cuts rework rates. Based on experience, rework is the biggest hidden cost in AI filmmaking. Many people blow their budget not because the unit price is expensive but because the same shot ran five times before succeeding. So if you truly want to save money, design your database table structure well first, avoid all the traps like idempotency locks and CRTOR-bound uids, and get it right the first time—nothing saves more than that. What's the immediate next action? Take a three-minute short film project you have on hand and build a cost sheet: count the shots, multiply by two to get the image count, multiply by the per-image price; then multiply the shot count by the estimated seconds per video and the per-second price to get the total video cost; finally add a 10-percent failed-retry budget. Add the three items together and see which one is highest. Based on experience, 80 percent of people find that video generation accounts for over 60 percent—and then they realize that to save money, they need to improve video generation success rates, not nickel-and-dime the image generation over a few mao. That's how you do the accounting for AI filmmaking: it's not about how much money you have, but about which link each yuan goes to and whether it's worth it.