Open Visual Program

July 29, 2026 ยท View on GitHub

Vex Auto Visuals use Open Visual Program v1 as the shared executable scene contract for HyperFrames and Remotion. The model controls composition, metaphor, geometry, visual primitives, and choreography. Vex controls evidence, validation, resource use, deterministic execution, and publication QA.

Pipeline

  1. Opportunity planning produces a signed VisualExplanationIR containing transcript evidence, facts, objects, relations, required labels, and forbidden content.
  2. Generative authoring asks the configured reasoning model for distinct complete scene graphs. Templates are examples and fallback macros, not the output format.
  3. Every candidate is normalized, signed, and checked against the packaged open_visual_program.schema.json authority.
  4. Semantic validation rejects unknown evidence bindings, invented copy, missing objects or relations, static semantic elements, unsafe asset references, invalid geometry, excess overlap, and resource-budget violations.
  5. Deterministic candidates are always generated. Invalid or unavailable model output therefore lowers creative variety without making Auto Visuals fail.
  6. The tournament scores grounding, object and relation coverage, semantic motion, mechanism fitness, novelty, layout safety, and resource cost.
  7. HyperFrames assigns distinct open programs to proof variants and renders them through its existing visual critics and final independent gate. Semantic fitness contributes to winner selection.
  8. Remotion compiles each valid candidate into signed SceneGraph v2, renders contact-sheet frames for all six treatments in one browser session, and signs the rendered winner selection before the final render.

Trust Boundary

Every non-decorative visual element has a binding to an exact fact_id, object_id, or relation_id. Visible copy must be supported by transcript or explicit title evidence. A valid JSON shape is not sufficient: the runtime also recomputes the program signature and evidence signature before execution.

Remote assets, executable code, dynamic imports, network requests, filesystem access, wall-clock animation, and unseeded randomness are not part of the Open Visual Program language. This gives the model design control without giving generated content process authority.

Scene Language

Open Visual Program supports normalized layout, semantic bindings, constraints, relations, repeated primitives, and motion tracks. Current element primitives are:

  • text, token, shape, group, path, connector, icon, metric, and chart
  • particle, mask, and evidence-bound local image

Motion tracks support opacity, translation, scale, rotation, progress, emphasis, and blur. Each track carries a semantic intent and ordered keyframes. Both renderers derive animation only from timeline progress, FPS, program inputs, and the deterministic seed.

Typed Repair

apply_open_visual_patch() accepts only bounded operations:

  • move or resize an existing element
  • replace evidence-grounded text
  • update an allowlisted style field
  • update an existing motion track
  • remove decorative elements
  • update the concept description

The patched program is re-signed and must pass the complete schema, evidence, geometry, motion, and resource validation suite. Unknown targets, arbitrary CSS, and unsupported operations are rejected.

Configuration

VariableDefaultPurpose
OPEN_VISUAL_PROGRAM_ENABLEDtrueEnables shared open-program compilation
OPEN_VISUAL_PROGRAM_LLM_AUTHORINGtrueAllows evidence-bound model authoring
OPEN_VISUAL_PROGRAM_CANDIDATES6Candidate budget, capped at six
OPEN_VISUAL_PROGRAM_AUTHORING_ATTEMPTS2Initial model attempt plus one repair
OPEN_VISUAL_PROGRAM_MIN_SCORE0.78Static acceptance floor before rendering

Artifacts

Auto Visuals manifests include the selected program ID, authoring mode, candidate count, semantic-fitness score, tournament signature, novelty fingerprint, and model rejection diagnostics. Renderer job directories preserve the complete program and the renderer-specific signed scene artifact. HyperFrames additionally stores one open_visual_program.json per rendered proof candidate.

Compatibility

Legacy Remotion scene families, HyperFrames Visual World programs, Scene Program V2, bespoke scenes, and templates remain available. They are compatibility fallbacks when Open Visual Program is disabled or cannot be compiled. Existing manual visual commands are not forced through model authoring.