Video pipeline selection · v0.1.1 experimental
Framewright
Help an AI agent choose and execute a suitable video pipeline, based on documented experiments.
Not a new video engine. Tested recipes, limitations and result checks for existing tools.
Reproduce the example
Task: introduce Framewright in a 30-second horizontal film.
Selected chain: authored Remotion scenes + original locally synthesized music → native MP4 and PNG still.
Result: watch the Framewright film.
Checked: H264 1920×1080, 30 fps, 900 frames, stereo AAC, native PNG and sampled encoded transitions. Human listening and audience performance are separate.
Open reference project: authorised assets, pinned dependencies, exact agent request, commands and output checks. Pilot elapsed time and compute cost were not retained; no speed or cost benchmark is claimed.
Choose tools by scene and layer
- Software launch. Brag full + Hyperframes for a short introduction grounded in the working project.
- Narrated book trailer. Existing artwork + local Kokoro narration + sound-layer checks for the final encoded file.
- Live React interface. Flute camera/focus around existing components; FFmpeg can add music and extract a thumbnail where the portable snapshot failed.
- Authored motion graphics. Remotion for frame-based typography, native audio and stills; Anatomy can provide object SVG assets within its prototype scope.
- Common finish. A film can combine tools per block and layer. Declare timing, canvas, colour/alpha and sound handoffs; validate the assembled file.
Eight-tool capability matrix distinguishes tested, documented, failed and unknown features. Style is an authoring direction, not a fixed engine limit. The offline JSON planner proposes combinations; new chains still need integration tests.
Watch the published experiments
- Jeanne la bataillonneuse — French trailerLocal narrated montage · approximately 28 seconds
- Good Dogs of the Apocalypse — English series trailerThree-book sequence · approximately 34 seconds
- CoqPi — Flute React camera pilotReal components with synthetic inputs · 23 seconds · FFmpeg audio finish
- Framewright — Remotion identity filmAuthored graphics, original music and native still · 30 seconds
Video URLs were supplied by the owner; public playback references are separate from retained master hashes.
Part of the ABVX ecosystem
- Agent Skills
- General reusable workflows; Framewright stays an external companion.
- SET
- Optional planning around the selected recipe.
- agentsgen
- Lean repository context, not a video runtime.
- Framewright
- Source of truth for video methods and retained experiment evidence.
No default installation, publishing hook or video job. New tools enter through bounded experiments.
Start small, keep the evidence
Read the canonical README, follow the open reference route, or install the optional router from the experimental v0.1.1 release. Model weights, media and upstream tools are not bundled.
The next planned experiment evaluates agent selection against frozen scenarios. Deterministic planner tests and completed films do not qualify routing behaviour; it remains structural_only.