# CANVAS / Storyboard Continuity Research Pack

Date: 2026-05-17  
Repo snapshot inspected: current working tree, `HEAD=b9ee5f1`  
Primary paper: [CANVAS: Continuity-Aware Narratives via Visual Agentic Storyboarding](https://arxiv.org/html/2604.13452v1)

This folder translates CANVAS and adjacent long-form visual storytelling research into concrete development direction for TheRoad-I1. The emphasis is not paper summary alone. It maps research ideas to the current pipeline, prompts, contracts, reference image services, semantic-debt cleanup, and verification strategy.

## Documents

1. [01-paper-and-related-work.md](01-paper-and-related-work.md)  
   CANVAS core ideas, related papers, and what each contributes to a practical storyboard system.

2. [02-project-architecture-map.md](02-project-architecture-map.md)  
   Current TheRoad-I1 pipeline map: prompt packs, StepRunner manifest, RenderPromptCard, reference image service, state/gaze work, scene_detail, and existing continuity contracts.

3. [03-gap-analysis.md](03-gap-analysis.md)  
   Gap analysis between CANVAS-style world-state modeling and the current codebase.

4. [04-development-roadmap.md](04-development-roadmap.md)  
   Proposed development direction: near-term, medium-term, and later work, with module-level ownership and risks.

5. [05-prompt-and-eval-guidelines.md](05-prompt-and-eval-guidelines.md)  
   Prompt hygiene, structured SOT rules, evaluation gates, and benchmark direction.

6. [06-analysis-method.md](06-analysis-method.md)  
   How this analysis should be repeated: code/prompt audit method, research-reading method, and decision framework.

7. [07-claude-handoff-no-vlm.md](07-claude-handoff-no-vlm.md)  
   Claude execution handoff: how to apply CANVAS ideas without trusting VLM as a truth source.

8. [sources.md](sources.md)  
   Source list and local evidence anchors used in this analysis.

## Executive Takeaway

CANVAS is most useful here as an architectural signal: long-form visual consistency improves when the system maintains explicit world state, retrieves anchors from memory, and updates state across shots. The VLM/QA portion of the paper should not be copied as a trusted verifier because current VLMs are not reliable enough for closure gates in this project. TheRoad-I1 already has many pieces of the useful non-VLM part, but they are spread across prompt packs, RenderPromptCard, reference services, scene_detail, state variants, background binding, and semantic cleanup specs.

The development direction should be:

- Keep the current structured-SOT migration direction. Area #1 and Area #2 are aligned with the paper trend.
- Treat RenderPromptCard as the shot-level contract, but introduce a higher-level scene/story memory layer rather than stuffing all long-range world state into one card.
- Add an explicit world-state memory bank for character appearance state, location/background anchors, prop state, and recent frame summaries.
- Keep evaluation contract-first. Generated-image review can exist later as advisory triage, but not as a source of truth or closure gate.
- Avoid semantic regex and scenario-noun prompt rules. Use structured producer emit, helper validation for closed-world IDs/enums, and fail-fast gates. Model-based image observations, if added later, stay advisory.

## High-Level Direction

```mermaid
flowchart LR
  Script[Script / selected shots]
  Plan[Global continuity plan]
  Memory[World-state memory bank]
  Shot[Shot-level staging]
  Card[RenderPromptCard contract]
  Prompt[scene_detail prompt]
  Image[Image generation]
  Eval[Contract/provenance review]
  Update[Memory update]

  Script --> Plan
  Plan --> Memory
  Memory --> Shot
  Shot --> Card
  Card --> Prompt
  Prompt --> Image
  Image --> Eval
  Eval --> Update
  Update --> Memory
```

## Project-Level Interpretation

TheRoad-I1 is already moving from prose-heavy prompts to deterministic contracts:

- `shot_staging v13` separates gaze direction, target ID, and subject state.
- `scene_detail v25` treats RenderPromptCard as the primary contract.
- `render_prompt_card.py` collects render strategy, ID policy, background binding, continuity elements, asset requirements, and render contracts.
- `scene_reference_service.py` already manages visible entity refs, previous-shot background refs, source provenance, state variants, and image-ref rewriting.

The missing piece is a persistent, typed, post-generation memory update loop. CANVAS and related memory papers suggest that is the next major architectural layer.