ai-video-gen

The scope is ambitious on purpose: image generation, video synthesis, voice-over and FFmpeg assembly in one entry point, with multiple providers named per stage. That is convenient, but it also means a single request inherits the cost, latency and failure characteristics of three or four upstream services, and the slowest one sets the wall-clock time. Because provider selection is per stage, pinning choices rather than letting the script choose makes runs both cheaper and more reproducible. FFmpeg assembly is the most likely place for a run to succeed on every API call and still produce nothing usable, so verify the output file, not just the exit code.

Use case: Producing a complete narrated video from a text brief where multi-provider orchestration is worth the latency.

# Running the video pipeline in stages The skill chains four independent stages, image generation, motion synthesis, voice-over and editing, and each stage has its own provider. Treating it as one make-a-video command is where runs go wrong. Generate and inspect stills first, then animate, then narrate, then assemble. ## Provider choice Stills can come from a hosted image model or a diffusion provider. Motion comes from a video model, and voice from a text-to-speech provider. Only the image key plus one video-provider key are strictly required for a first cut, so configure the minimum and add narration later rather than wiring every provider up front. ## Keep the intermediates Every stage takes explicit input and output paths. That is the entire reason the pipeline is worth scripting: you can regenerate one stage without paying for the others again. ```bash # stage one: stills, stage two: motion, stage three: narration python3 generate_video.py --prompt "harbour at dawn, slow pan" --output cut.mp4 python3 images_to_video.py --images shot1.png shot2.png --output motion.mp4 python3 add_voiceover.py --video motion.mp4 --text "narration line" --output final.mp4 ``` ## Tips - Generate stills and review them before animating. Motion models amplify a bad frame rather than repairing it. - Never delete an approved intermediate; regenerating it wastes spend and changes the look between revisions. - Match narration length to the motion clip before assembly. Stretching video to fit audio costs more than rewriting the line. - gotcha: keys are read from the environment or a dotenv file, so a missing key surfaces as a provider error partway through the pipeline rather than at start-up. Validate the keys you plan to use before the first paid call. - Use a single voice provider for a project instead of mixing voices across revisions.