How an 11-Day 2-Hour AI Film Keeps Faces Stable and Scenes Consistent

Completing a two-hour-and-twenty-minute AI film in eleven days without face collapse or scene crossover relies on a highly disciplined system: three templates plus one switch. Three prompt templates control characters, locations, and scene compositing. The imgstyle field globally controls four art styles—realistic, ancient, anime, and modern—with a single value. All storyboards share this fixed structure, with only three variable points; everything else is locked down. In measured testing across four chapters and over two thousand first and last frames, not one face collapsed and not one scene drifted to the wrong place. How are characters kept consistent? Through the character information segment in the eight-part skeleton. The process.aiwords Jimeng video prompt template is fixed at eight segments: character information, lip-sync consistency, accent and tone, shooting technique, scene location, on-screen action, aspect ratio, resolution, and duration. Within these eight segments, the character information segment is a description refined and finalized for each character. For the same character, no matter which shot or which chapter they appear in, this text never changes. For example, if you've settled on "black hair, high ponytail, scar on the left brow, black leather jacket, sharp eyes," then every shot featuring this character across the entire film copies that description verbatim. Character traits won't get lost because the description itself is hard-coded in the data. The material states clearly: there are only three variable points, and character information isn't one of them—it's an anchor, welded into the template. How do art styles stay unmixed? Through the imgstyle field, locked globally. This field is carried in from article.atctype. When you create a chapter, you fill in the value matching that chapter's style. The style switch has four values: xs for realistic, gz for ancient, dm for anime, xd for modern. If you define the entire novel as gz ancient, then all storyboards automatically apply all the visual constraints of the ancient style when generating images and videos. The model has no chance to drift off and draw randomly, because you never give it an opportunity to switch styles. The material explains this plainly: the style switch is a single field, not a set of prompts, not a per-generation parameter you tune in the moment—it's a global, unique value. Change it once, and the entire film is unified. In other words, you only need to decide the style once at the start of the project and never touch it again, because the fixed style description in the template travels with every shot. Do more shots cause collapse? No, because every storyboard generates images and videos independently. The template is redundant but stable. With 1,029 shots, if one shot fails to generate, you only re-run that shot's image generation and video generation tasks; the other shots are completely unaffected. This is what the material means by enumerable and re-runnable—every shot generates with the same template paired with that shot's unique character, location, and action parameters. The template is the foundation; the parameters are the variables. When the foundation doesn't break, the variables can't cause confusion no matter how complex. In practice, "The First Avenue" ran all four chapters completely, and characters didn't undergo identity shifts across chapters, nor did locations become type-confused due to the number of scenes. Why is this achievable? Because at generation time, the character information segment locks the character's appearance, the scene location segment locks the location's features, and the action segment is only responsible for that shot's dynamic description. The three variable points each do their own job without overstepping. Based on experience, ninety percent of face collapses and scene crossovers happen because template variables weren't locked down—you treat what should be fixed as a free variable, and one wrong step leads to everything going wrong. What mistake do beginners most often make? Messing with template variables. The material explicitly states there are only three variable points—that's the proven limit. But beginners can't resist tinkering. For example, in a few shots they add "wearing glasses" to a character, in the next few they write "not wearing glasses," or they casually rewrite the entire character information segment. The result is the same character in the final film alternately wearing glasses and not, with a face shape that fluctuates between chubby and thin, and clothing flipping between ancient and modern. Based on experience, even changing a single adjective makes the model treat it as a high-weight feature fused into the generation result, directly polluting the entire video. So the three variable points mean: you may fine-tune within the dimensions of action, scene, and shooting technique. Character appearance, accent, and lip-sync consistency are absolute forbidden zones. The material deliberately placed "only three variable points" in the process.aiwords description to remind users: don't treat this as a free creative entry point. It's an assembly-line control panel—which buttons can be pressed and which can't is marked at the factory. How are locations protected from crossover? Through the places table's one-row-per-shot design plus the placeaiwords empty-rolex branch to lock scenes. The material specifically states: places is not a location dictionary; it's one row per shot. What does this mean? Every shot's location is an independent record that doesn't reuse other shots' scene descriptions. Shot A is at the palace, shot B is at the city gate—the two records don't interfere. At generation time, unmanned locations go through placeaiwords' empty-rolex branch for pure text-to-image; scenes with people go through placeaiwords' character-present branch plus scenetopic for image-to-image compositing. In the compositing prompt, fmpic is the character image and plcpic is the location image; the two combine into the scene's last frame. The key to this workflow: each shot's location prompt is its own—no sharing, no borrowing—so there's no such thing as scene crossover. If you took the palace's location description and used it for the city gate, the city gate would generate palace pillars. That's data pollution, not an inadequate template. Based on experience, the best practice is to write all location prompts uniformly *before* starting the run. Don't start, discover a location description is bad halfway through, and change it—changing one means regenerating a whole batch of images, and costs rise exponentially. Are there counterexamples? Yes, more than one. The most typical: treating imgstyle as an optional switch, deciding to try a different style in chapter four—and generating a mess of incoherent visuals: the first three chapters are ancient, chapter four suddenly jumps to modern, and the audience is thrown straight out of the story. The material clearly states imgstyle is carried in from article.atctype—it's a global variable fixed per chapter, not something you change mid-way. Another counterexample: someone, to make the visuals richer, took it upon themselves to expand the eight-part skeleton to ten parts, adding two segments. The model over-weighted those two segments during generation, they conflicted with the original eight segments' information, and the generated visuals literally split apart. Based on experience, the template's segment count is a floor, not a ceiling. You should only fill in words within the existing eight segments—never add segment count. Because each segment's order, position, and logical relationship were tuned through extensive testing; adding or removing one can change how the model interprets the entire semantic structure. What does this film's stability ultimately rest on? It rests on the underlying logic of turning creation into engineering. The material's line—"the essence of this design isn't the prompts; it's breaking a film into enumerable, re-runnable, replaceable records"—translates to: you don't rely on inspiration to maintain art style; you rely on structure. The style can't collapse because structure won't allow it to collapse. Faces can't drift because data won't allow them to drift. Three templates plus one switch are the skeleton of this structure. They aren't magic; they're like molds. Every generated frame passes through this mold, and what comes out naturally shares the same shape. What's the immediate next action? Take the character information segment of a character you use frequently and save it as a standalone template. Then in the drawrole pipeline, generate ten consecutive images of this character in different poses. Based on experience, you'll see that across all ten images, face shape, skin tone, hairstyle, and clothing details remain completely consistent—the only changes are angle and expression. That's the effect of the character information segment anchoring the invariants. Whether it's coherent—you'll know in your gut after seeing those ten images. If one face in the middle collapses, don't go tweak that generation's image parameters. Go back and inspect the character information segment; most likely, ambiguous wording crept into your description. Make it precise to the specific shape, regenerate ten images, and compare again. These ten images are the first cornerstone of art-style consistency across your entire film.