I built an AI pipeline that takes a production idea and turns it into generated images and video sequences. In my local project, the app is called Reel Forge.
It starts with a production spec. From there, it plans references and shots, creates images, animates them, reviews the clips, and assembles a vertical video with captions and narration.
What makes it interesting is how the shots relate to each other. A sequence has to carry enough information that one image can plausibly follow the one before it. If an object changes shape or resets between shots, a set of individually good-looking images can still fail to tell the story.
Making the spec usable by the app
The first stage turns the production spec into structured data. The planner identifies reusable prompt blocks, reference images, individual shots, timing, voiceover, captions, counters, and any shots that need a composed layout, like a split screen.
That structure gives the later stages something concrete to work with. A shot points to specific references and describes its starting state. The assembly stage uses the timings and overlay instructions without having to reread the whole original request.
The plan also gets validated. One requirement is that reusable prompt blocks keep the exact wording from the spec. That stops important instructions from drifting while the plan is rewritten into another format.
Approving references before producing a sequence
Reference images get their own review step, and some depend on other references. The project docs use a restored watch, generated from an approved damaged watch, as an example. The point is to keep it the same underlying object.
That dependency matters. Two images generated separately from similar descriptions can still show two different objects. Passing an approved reference into a later generation gives the workflow a specific visual starting point.
I can regenerate a reference with a note, or supply one myself. Approval is a checkpoint, so the system doesn't produce every dependent shot until I've signed off on the reference.
Carrying state between shots
For each shot, Reel Forge combines the relevant references with continuity notes. The first frame can depend on earlier frames and on a description of what has already changed in the scene.
The docs give examples like a case that's open or closed, parts lying on a workbench, or an object that's partly cleaned. Those details are simple, but a viewer notices them. They describe the state of the story, not just how an image looks.
Generated stills get checked for problems like design drift, unwanted text, or other breaks of the production rules. A rejected result can be retried with the issues fed back in as corrections. That's a real review step. It doesn't guarantee every error gets caught.
From still images to edited footage
The approved still becomes the first frame for an image-to-video task. Reel Forge sends those jobs to MiniMax and tracks the results. The video prompt is built from the planned blocks, the visible scene, the action, and the relevant constraints.
After generation, another review step looks at the clip and picks a trim window that fits the planned shot length. A generated clip is often longer than the part that belongs in the edit, so choosing that window is an editing decision as much as a quality check.
FFmpeg handles the final assembly. The app trims clips, burns in captions and labels, adds counters where the spec calls for them, builds the split-screen shots, and places the narration where the plan says. The output is one vertical MP4, not a pile of separate generation results.
Keeping work alive and revisions affordable
The app saves its workflow state in Firestore and uses a lease around the production work. A scheduled continuation can carry the process forward after the browser closes. That matters, because video tasks take time and shouldn't depend on a tab staying open.
Pause and resume are part of that state. Video tasks that were already submitted still need to be collected, and the app has to remember where it left off.
Revisions can be narrower than starting over. I can regenerate one shot or clip with a note. A change to the captions or the assembly can reuse the media that already exists. Keeping those options separate gives me better choices as an editor.
Reel Forge is my attempt to build the production system around an AI video sequence: a plan, references, continuity, review, and an edit. Whether the result is good still comes down to watching it. The engineering work is making the process something I can repeat and revise.