Assemble a video from AI-generated media

Generate images from prompts, animate them, add narration, and cut the result together with crossfades. Everything runs locally; image and video generation need a CUDA GPU (see Install).

The shape of the job

Generation produces in-memory Video and Audio objects. Editing operations run only through the streaming engine, which reads from files — so save each generated scene, then assemble the saved files in one plan.

from pathlib import Path

from videopython.ai import TextToImage, ImageToVideo, TextToSpeech
from videopython.base.video import VideoMetadata
from videopython.editing import VideoEdit, SegmentConfig, TransitionSpec, Resize


def create_ai_video(output_path: str, workdir: str = "scenes") -> None:
    scenes = [
        {"image_prompt": "A serene mountain landscape at sunrise, photorealistic",
         "narration": "In the mountains, every sunrise brings new possibilities."},
        {"image_prompt": "A flowing river through a forest, cinematic lighting",
         "narration": "Nature flows with endless energy and grace."},
        {"image_prompt": "A starry night sky over a calm lake, dramatic",
         "narration": "And when night falls, the universe reveals its wonders."},
    ]

    image_gen, video_gen, speech_gen = TextToImage(), ImageToVideo(), TextToSpeech()

    Path(workdir).mkdir(parents=True, exist_ok=True)
    scene_paths = []
    for i, scene in enumerate(scenes):
        image = image_gen.generate_image(scene["image_prompt"])
        video = video_gen.generate_video(image=image)
        audio = speech_gen.generate_audio(scene["narration"])
        path = f"{workdir}/scene_{i}.mp4"
        video.add_audio(audio).save(path)
        scene_paths.append(path)

    # One segment per scene. Resize standardizes to 1080p; a 1s dissolve crossfades
    # each follow-on scene in (the first has no predecessor, so it carries none).
    segments = []
    for i, path in enumerate(scene_paths):
        meta = VideoMetadata.from_path(path)
        segments.append(SegmentConfig(
            source=path,
            start=0,
            end=meta.total_seconds,
            operations=[Resize(width=1920, height=1080)],
            transition_in=None if i == 0 else TransitionSpec(type="dissolve", duration=1.0),
        ))

    VideoEdit(segments=segments).run_to_file(output_path)


create_ai_video("ai_generated.mp4")

The generators

Step	Class	Local model
Prompt → still	TextToImage	Qwen-Image
Still → motion	ImageToVideo	Wan2.2-I2V-A14B (16 fps output)
Prompt → motion	TextToVideo	Wan2.2-T2V-A14B
Text → narration	TextToSpeech	Chatterbox Multilingual
Text → music bed	TextToMusic	MusicGen

Full signatures in AI generation.

Shape the narration delivery

generate_audio forwards three Chatterbox knobs. Each defaults to None, meaning "let Chatterbox decide":

dramatic = speech_gen.generate_audio(
    "We made it.",
    exaggeration=0.85,   # more expressive
    cfg_weight=0.35,     # slower pacing
)

Reframe a generated scene for vertical

face_crop is just another operation:

from videopython.ai import FaceTrackingCrop
from videopython.editing import VideoEdit, SegmentConfig

edit = VideoEdit(segments=[SegmentConfig(
    source="scenes/scene_0.mp4",
    start=0,
    end=5,
    operations=[FaceTrackingCrop(target_aspect=(9, 16), framing_rule="center")],
)])
edit.run_to_file("scene_0_vertical.mp4")

Notes

• Keep prompts stylistically consistent across scenes or the cut will not feel like one piece.
• Match narration length to scene length — ImageToVideo clips are short, so long narration will outrun the picture.
• Generation dominates the wall clock. Save intermediate scenes (as above) so a failed assembly does not cost you another generation pass.