Skip to content

Roadmap

zhanglinghao edited this page Oct 29, 2026 · 5 revisions

English · 中文

This is the maintainer's plan (@ZLHad) as of 2026-10-29. It changes as work lands, and this page is edited to match. To suggest a change, open an issue. "Later" items are ideas, not promises.

What already works is on Project Status, with the pull request behind each part; what changed is in the CHANGELOG.

Now: in progress

  • The homepage, brought up to date in its own look (#168): today's counts and tools, a capability map, nothing loaded from other sites, a better phone layout. The READMEs were rewritten just before it (#166), and this wiki follows, so the three say the same thing.

On hold

  • Paid generation of 3D models and video clips. A 3D model generated from a picture or a line of text (Tripo), and generated video clips (MiniMax's video models), would join bin/vh image and music --provider lyria as paid calls. The budget, the estimate before each call and the asset ledger they would use are in place; the two commands are not being built for now.

Done since the last update (2026-10-03)

  • Edit lists (#157, #158, #159): a film as a row of pages, images, colours, slates, footage, other films and sound, with a chunk cache so a render after a change draws only what changed.
  • Material on record (#131, #133, #145, #149, #151, #152, #163): one asset ledger with a budget for every paid call; archive footage; generated pictures; a sound library; a generated music track; a project's cast.
  • 3D (#161, #162, #164, #165): canvas-film's Three.js layer, model3d check, hdri, Blender for assets only, and ManimGX as an optional backend for Manim scenes.
  • Rendering any page (#81, #82, #113, #125, #90, #148, #160): the page contract; resumable chunked renders, versions and patch; transparent films and motion blur.
  • The library (#154, #163, #143, #162, #123): 71 shot recipes with sample sheets, 24 shot-grammar cards, 5 motion components, a 32nd style.
  • Ideas and research before the film (#144, #137, #142, #140, #138, #146): know, scout, ref, breakdown, the concept-card extras, a topic card, and the next episode of a series.
  • Stops and review (#54, #57, #120): the review desk, caps on review rounds, a stop in one command (bin/vh gate), where a project stands (bin/vh status).
  • Effort and guidance (#64, #65, #105): effort sets iteration and review, not quality; taste written as advice, with a short list of floors.
  • Checks and delivery (#59, #93, #147, #68, #89, #92): check against the BRIEF and how the film moves, with a way out of each finding; text and fonts measured in the rendering Chrome; deliver and clean.
  • Sound (#69, #72, #115, #118, #153): the MiniMax voice, pace, qa reading the score, local word timing and places to listen.
  • The command line (#130, #132, #139): one command table behind every -h, an outside-module table behind doctor, HyperFrames 0.8.202.
  • Repo and install (#52, #53, #58, #167): a shallow clone in the installer, the videos in the media release and out of the history, CI on macOS in two shards.

Next

  • A release that carries what has piled up on main since v0.2.1: the CHANGELOG's Unreleased section is long.
  • Try the experimental parts on the real thing. Type 09 has never cut real footage. Its edit list now renders hard cuts and retimes captions from the clips' words, but the approval hash its docs describe isn't checked yet. A few of the Blender guide's smoke tests haven't been run, and the dashscope and elevenlabs voices haven't met their live APIs.
  • Every swatch reaches 8 or more on every review score (today none does; see Style Library).
  • 9:16 variants of the showcase films, and a Chinese-narration variant of the intro film.
  • A type doc for 3D and shader films. Today they follow the promo type plus playbook/08, with canvas-film's Three.js layer for 3D in a narrated film and the Blender guide (partly verified) for path-traced shots; the intro film's opening is the worked example.
  • Linux install polish. CI covers Ubuntu and a first round of fixes is in; the install has been tried on Ubuntu 24.04 only.
  • Community style submissions: a submission template and a review process.

Later: ideas

  • More engines and video types as the community asks for them.
  • A tempo map for bin/vh music (ritardando, accelerando): a score has one bpm today.
  • A desktop client that takes a film from one sentence to delivery, with the editing done by the local agent. The proposals are in docs/proposals; the client is on hold, and the command-line groundwork it needs is in (status, gate, versions, patch by shot).
  • Packaging the whole workbench as an installable skill or plugin. Today's skill is a thin pointer that installs the repo on first use.
  • Languages beyond Chinese and English, and English versions of the workflow docs.
  • Word-by-word highlighted subtitles (word timing already lands in captions.json when you use --align).

How to help

Pick an item, say so in an issue so work isn't duplicated, and follow Contributing.

Clone this wiki locally