Which Step in the 5-Step Book-to-Storyboard Formula Produces Most Bad Frames?

In the formula where scene equals character plus location, the step that most often produces bad frames is scene compositing. Because the character-present branch of placeaiwords uses fmpic and plcpic—two images—for image-to-image compositing. Once the character image and location image have mismatched styles, the composite image falls apart. Based on the practical work of "The First Avenue," among all 1,029 shots, the step with the highest bad-frame rate is this scene compositing step—not the earlier character-image or location-image generation. How is scene compositing specifically triggered? Through the placeaiwords character-present branch. The condition for this branch is simple: is there a person in the scene? If yes, use the scene prompt plus scenetopic for image-to-image compositing. In the compositing prompt, fmpic is the character image and plcpic is the location image. The system automatically pulls the character ID and location ID linked to the storyboard record, retrieves the character image from the dbvideo_roles table and the location image from the dbvideo_places table, feeds both to the generation interface, and outputs a merged scene image. After this step completes, what you get is the actual visual foundation this shot will use—the first and last frames are refined from this image later. So it's the hub of the entire image generation flow. The character and location being qualified separately is just the baseline; the composite result determines whether this shot is usable. Why does it collapse? The core cause is style conflict. Say your style switch imgstyle is set to gz ancient. Then the character image should be an ancient-costume figure and the location image should be an ancient-style scene. But if the image stored in the character table was generated with xs realistic parameters, or the location image is a background generated with dm anime parameters, the model doesn't know how to stitch these two completely incompatible visual languages together at compositing time. The result is either the character's face gets swallowed by the background, the background style gets dragged off by the character, or the model produces an undefinable monster. The material explicitly states: image generation follows three paths—characters go through drawrole via Volcano Engine synchronous direct call, locations go through placeaiwords' empty-rolex branch as pure text-to-image, and scenes go through placeaiwords' character-present branch plus scenetopic for image-to-image compositing. These three paths must obey the same style logic simultaneously. If even one path lags behind, the flaw shows up at the compositing step. Based on experience, even when the style difference isn't as extreme as realistic versus ancient—just a subtle tonal difference within the same style, like the character image leaning warm and the location image leaning cool—the composite result develops a gray hazy film, and the image looks visibly dirty. How do you avoid it? First make sure the character and location images are generated separately under the same style, then run scene compositing. Concretely: when generating the character, confirm drawrole carries the current chapter's imgstyle value, say gz ancient. When generating the location, confirm the placeaiwords empty-rolex branch also carries the same gz value. Once both images are out, do a quick visual check of color tendency and light direction. If the character image has left-side lighting and the location image has right-side lighting, regenerate the location image immediately—don't wait until after compositing to fix it. Based on experience, spending one minute before scene compositing to check the two images' styles saves three times the time compared with re-running the whole chain after compositing. Because once compositing produces a bad frame, you have to redo tuning, regeneration, and re-compositing—the entire chain—and every step in between involves an extra large-model call and minutes of waiting. The material mentions that the large model channel is bound to uid via CRTOR, precisely to keep styles of images generated in the same time window close. So some style drift issues can be mitigated by uid binding, but it can't fully replace manual inspection. How high is the bad-frame rate? Based on experience, with a reliable template and style set, one or two composite images out of ten being bad frames is normal. Don't expect zero collapse—that probability isn't realistic. What do you do when it collapses? Re-run. Re-running doesn't mean changing the prompt; it means submitting the exact same compositing task again. Every model generation has randomness. Re-running once will most likely yield a much better result than the last. The material emphasizes the system's re-runnability, and its core scenario is precisely here: adjusting shots doesn't require redoing everything; re-running a single shot doesn't affect other records. So a high bad-frame rate isn't the problem—being able to redo at low cost is the design's core. Based on experience, a shot will output a usable composite image within at most three re-runs. If it still collapses after three, go back and check whether the character image or location image themselves have problems—don't keep pushing forward into video generation, because video generation amplifies the compositing stage's collapse and burns more money. What mistake do beginners most often make? Some think scene compositing collapsing means the prompt wasn't written well enough, so they start making the action segment in the eight-part skeleton elaborate and flashy, trying to "save" the image with text descriptions. The result is exactly what the material warns: there are only three variable points. You turn the action segment into meaningless flashy descriptions, the model treats it as absolute weight, and the result goes even further off the rails. The correct approach is to keep the template structure untouched, first check whether the two source images' styles match, then consider re-running. Others skip scene compositing entirely and manually stitch images—compositing the character and location in PS and sending that into video generation. This can look passable for the first frame, but when the last frame is generated a second time, because the semantic fusion process is missing, the character and background look like they're not properly attached. When the character moves in the video, the background edges show. So don't take shortcuts. The image-to-image process at the scene compositing step exists to teach the model the logical relationship of "this person in this place." Skip it, and you're making the model guess everything. Are there counterexamples? Yes. Someone, to make scene compositing more "textured," swapped in a high-definition portrait-style headshot, with a blurred photo-studio background, and composited it into an ancient palace location image. The result: the face is tack-sharp, but the model re-rendered half the clothes in ancient style while the lower half remains modern fabric—like someone who was pushed into a palace wearing half an ancient costume. This is the classic case of the character and location samples being too different—the model can only take half of each and force them together. Another counterexample: someone reused the same character image with different locations, thinking this would keep the face consistent. But because that image's lighting and angle were too distinctive, it carried an unnatural floating feel in every location. Based on experience, for a character image, prepare three to five versions at different angles and lighting conditions, and select by the scene's lighting needs. Don't rely on one image to work everywhere. What's the real meaning of this formula? It's using determinism to fight randomness. Character plus location equals scene—the formula itself doesn't generate imagery, but it generates a storyboard record. The record is deterministic, enumerable, and re-runnable. When the compositing stage collapses, you re-run the same record—you're not re-creating a scene from scratch. The material explains this thoroughly: the design's essence is breaking the film into enumerable, re-runnable, replaceable records. Swap an actor with one row edit, adjust a shot without redoing, parallel machines grab tasks. So bad frames aren't an accident—they're the daily routine of this system. What you should pursue isn't zero failure; it's ensuring every failure can be redone at low cost. What's the immediate next action? Find a shot you're working on that collapses often. First check whether the current imgstyle of the character image and location image is the same value. If not, unify them to the same style, regenerate the character image and location image separately, then run the scene compositing interface once. Then visually inspect the composite: is the character's silhouette fully embedded in the scene? Is the light direction consistent? Did the background swallow the character's face? Get this right, and the bad-frame rate drops by about half. The remaining one-in-ten-or-two randomness gets handled by re-running. Based on experience, a shot re-run two or three times will ultimately yield a clean, usable scene foundation—from there, proceeding to first and last frames and video generation is stable.