AI Film Budget for a 2h20m Feature: What It Costs and How One Person Controls It

Based on actual measurements from "The First Avenue," one person using three computers over eleven days to complete a two-hour-and-twenty-minute AI film, splitting out a total of 1,029 shots, incurs costs mainly from AI generation channel calls and electricity. Excluding labor, the cost is on the order of a few thousand yuan. The exact figure depends on the per-call unit price of the model you use, whether billed per image or per video. This is not a fixed price but an engineering budget that can be calculated in advance, provided you first count the number of calls at each stage. Which stages actually cost money? Your money goes to five items. Character image generation uses drawrole via Volcano Engine synchronous direct call, billed per image. Location image generation uses the empty rolex branch of placeaiwords, which is pure text-to-image, billed per image. Scene image generation uses the character-present branch of placeaiwords, adding scenetopic for image-to-image compositing. This stage references two base images—fmpic is the character image and plcpic is the location image—so the cost of one scene image equals one image-to-image composite plus two base image references. Each shot produces two images, first-frame -a and last-frame -b, and these two images drive image-to-video generation, which runs through Jimeng and is billed per video. So the cost chain for one complete shot is: one character image, one location image, one scene composite, first-frame and last-frame images each once, plus one Jimeng video. Among these five stages, video generation is usually the most expensive, followed by image-to-image compositing, while pure text-to-image is relatively cheap. Based on experience, for a two-hour film, video generation accounts for more than half of the total cost. If your shots have many action scenes and each shot runs at full duration, this portion becomes very prominent. Electricity is actually a minor expense—three computers running at full load for eleven days, at residential electricity rates, amounts to only two to three hundred yuan. Can you run a test before spending at scale? Yes, and it is strongly recommended that you run a test before making a large investment. The specific approach is to first enter a book title in dbvideo_bookidx, establish a bookcode as the root node of the entire universe. Then enter only the full text of a single chapter in dbvideo_article, run it through funx.atctoscene, and the system will automatically extract characters and locations into dbvideo_roles and dbvideo_places, while also generating the dbvideo_process storyboard table. Run that chapter's storyboards end to end from image generation to video, and you will get a precise per-chapter cost. Based on experience, the first chapter's cost is roughly 5 percent of your entire film's total cost, because the first chapter requires establishing character designs, setting the main locations, and testing the art style. Subsequent chapters mostly reuse these characters and locations, so costs drop noticeably. If the first chapter run shows the unit price is too high, you have only wasted 5 percent of the budget rather than half the film's money. Many people go straight to entering all 93,000 characters of the full book, only to discover halfway through that the art style is wrong or the channel billing exceeds expectations. The entire bookcode is then void, and because atctoscene has an idempotency lock that prevents re-runs, they have to create a new root node and re-enter everything—that is the real waste. The correct order is: first test-run one chapter to see the cost, then decide whether to go full scale. How do you control the count of 1,029 shots? The number of shots is not determined by how much you shoot but by how characters and locations collide. Scene equals character plus location, and dbvideo_process equals dbvideo_roles plus dbvideo_places. In other words, events involving the same character at the same location count as one storyboard; a change in action does not split it into additional entries. So to control shot count, you do not delete rows from the storyboard table; you control scene concentration at the script stage. If a character only acts in three locations, their shots are at most the three scenes reused with minor variations, not one entry per action. Conversely, if you change the location with every action and introduce a new scene for every new character, the shot count expands exponentially. Based on experience, keeping the cast to five or fewer main characters and ten or fewer main locations is a comfortable range for an individual creator. Beyond that, image generation counts and video counts rise together, and both budget and schedule will spiral out of control. So when you find that among your thousand shots, the location count exceeds twenty, the best move is to go back to dbvideo_article and merge those incidental scenes from the original text rather than stubbornly generating images for all of them. Remember, the places field is not a location dictionary; it is one row per shot, and every new location means an extra full pass of image generation plus video through the entire pipeline. Does the final output specification affect cost? It affects it enormously, but it can be controlled through staged settings. The final output specification you want is 2560x1440, 30 fps, H.264 encoding, 7.7 Mbps bitrate, AAC stereo audio. The generation cost and time for this specification are far higher than for a lower one. But you do not need to run every shot at this specification from the start. Based on experience, first run every chapter's first-frame and last-frame images and video previews at 1280x720 resolution. After you confirm the art style, composition, and color are all satisfactory, re-generate the key shots at 2560x1440. In actual practice, many shots' backgrounds and static portions are already adequate at low resolution; only character close-ups and action scenes need the high specification. Another cost-control method is the art style field. imgstyle has only four values: xs for realistic, gz for ancient costume, dm for anime, and xd for modern, carried in by article.atctype. For the same shot, the realistic style typically costs more than the anime style because realism demands more from detail and lighting. If your budget is tight, first run the entire film in dm or gz, then consider upgrading the main chapters to xs. The trade-off is that you must accept subtle differences in art style between chapters, but from a cost-control standpoint, this saves far more than regenerating everything. The output specification also directly affects the per-video price in the video generation stage. Tools like Jimeng typically price by resolution tier, and 2560x1440 may cost more than double the price of 1280x720. So starting low and upgrading later is the hard rule for saving budget. What is the most common mistake beginners make? It is not that video generation is too expensive; it is repeated calls in the image generation stage. Many people do not know that first-frame and last-frame images must be generated as a pair. If you only generate the first-frame -a and forget the last-frame -b, the secondpic field will retain the STCODE.png placeholder, and that shot's video is ruined. You have to re-call image-to-image generation, and re-calling means paying for another image-to-image composite. Another frequent error is the misuse of the places field. It is not a location dictionary; it is one row per shot, and every row must be independently called. Many people treat it as a deduplication dictionary, causing a single machine to produce a large number of empty rows and burn money for nothing. There is also the point that the large model channel is bound to the uid via CRTOR. If you switch machines to run the same bookcode, the channel does not automatically migrate. You must re-bind it on the new machine, or you will be billed twice. Based on experience, when running multiple machines in parallel, it is best to assign fixed chapters to fixed machines and avoid switching back and forth; that is where the mess happens. Also, vdurl contains no real video addresses for the entire film—the videos exist only in Jimeng. Do not spend time hunting for video files locally; that will waste huge amounts of time. Additionally, atctoscene has an idempotency lock. Re-running the same chapter does not overwrite old data; it simply stalls. If you think it has not finished and run it again, you only waste channel usage. Are there counterexamples that show budget runaway? Yes. Based on experience, the most common runaway case looks like this: one person took a 93,000-character novel, entered all chapters into dbvideo_article on day one without testing a single chapter first, and then on day four discovered that the character image direction was wrong and needed to adjust the imgstyle field. But the art style switch is carried by article.atctype, and changing the field means re-running the storyboard split. Since atctoscene cannot be re-run, they had to create a new bookcode and re-enter the entire book. At that point, they had wasted four days of generation costs and all the entry labor. Another counterexample is shot count runaway. For the same chapter, one person split out 35 storyboards while another split out 9. The difference lies in scene concentration. The former had a character walk from the office to the corridor to the elevator to the lobby, changing location with each action and creating four storyboards. The latter had the character converse, stand up, pick up a document, and leave the office, all within the office, counting as one storyboard. The cost of a single storyboard multiplied by a 4.5x difference, and across an entire film, that becomes a several-fold difference in budget. So the first cut for cost control is reducing scenes, the second is controlling quality, and only the third is comparing prices. Putting all of this together, for one person making a two-hour-and-twenty-minute AI film, the cost floor can be pressed below 3,000 yuan, the normal range is 4,000 to 8,000, and if you pursue extreme 2560x1440 full-film high quality with loosely designed shots, it could exceed 10,000. But regardless of which figure it is, you can calculate it in Excel before you start. The next action is very concrete: open a blank spreadsheet with three columns. The first column is drawrole call count, estimated by your number of characters multiplied by the average number of shots each character appears in. The second column is placeaiwords call count, estimated by the number of locations multiplied by roughly one per shot. The third column is the number of Jimeng videos, estimated by the total storyboard count. Multiply each column by the actual unit price of the channel you plan to use, add 30 percent margin, and the resulting number is your budget floor. If you come in below that figure, it means you missed a stage. If you come in above it, it means you should go back and control your shot count.