# 01 — Paper And Related Work

## Primary Paper

[CANVAS: Continuity-Aware Narratives via Visual Agentic Storyboarding](https://arxiv.org/html/2604.13452v1) frames storyboard generation as explicit world-state tracking. Its key claim is that long-form visual storytelling fails when systems only condition on local prompt text or recent frames. CANVAS instead combines:

- global continuity planning,
- persistent character/background/prop anchors,
- sequential generation with retrieved memory,
- candidate generation plus QA-based selection,
- memory update after each generated frame.

The paper reports gains on character, background, and prop consistency. Its appendix also notes a practical point relevant to this repo: training-free pipelines can keep heavy work in hosted multimodal/image APIs while local code performs orchestration, prompt construction, memory updates, and image I/O.

## CANVAS Mechanism

```mermaid
flowchart TD
  A[Input shot descriptions] --> B[Global continuity plan]
  B --> C1[Character appearance states]
  B --> C2[Location assignments]
  B --> C3[Object state transitions]
  C1 --> D[Visual memory]
  C2 --> D
  C3 --> D
  D --> E[Anchor retrieval]
  E --> F[Generate candidate frames]
  F --> G[QA selector]
  G --> H[Selected storyboard frame]
  H --> I[Memory update]
  I --> D
```

Core system idea:

- **Planning** decides what must persist or change.
- **Memory** stores reusable visual anchors.
- **Generation** is conditioned on retrieved anchors.
- **Evaluation** asks targeted continuity questions.
- **Memory update** writes back the generated visual state.

For TheRoad-I1, this maps cleanly to existing concepts: `scene_camera_flow`, `shot_staging`, `scene_consistency`, `RenderPromptCard`, `scene_reference_service`, `CharacterStateVariantStep`, and image validation. The gap is that these are not yet unified into an explicit story-world memory bank.

## Related Research Map

| Work | Main idea | Useful lesson for TheRoad-I1 |
|---|---|---|
| [StoryDiffusion](https://arxiv.org/abs/2405.01434) | Consistent self-attention for long-range image/video generation. | Good for subject style consistency, but it is model-side consistency, not a replacement for structured world state. |
| [Story2Board](https://arxiv.org/abs/2508.09983) | Training-free storyboard generation using latent panel anchoring and attention value mixing. | Useful reference-anchor direction; also highlights spatial composition and layout diversity, not just identity. |
| [AutoStudio](https://arxiv.org/abs/2406.01388) | Multi-agent interactive image generation with subject manager, layout generator, supervisor, and drawer. | Reinforces modular agents, but TheRoad-I1 should keep deterministic contracts between agents, not free-form handoff. |
| [ViStoryBench](https://arxiv.org/abs/2505.24862) | Benchmark for story visualization across narrative, character, style, and prompt alignment dimensions. | TheRoad-I1 needs continuity eval datasets/gates, not just unit tests and prompt residue checks. |
| [StoryMem](https://arxiv.org/abs/2512.19539) | Dynamic keyframe memory bank injected into multi-shot video generation. | Supports adding compact visual memory/indexing over past generated shots. |
| [VideoMemory](https://arxiv.org/abs/2601.03655) | Entity-centric dynamic memory storing visual and semantic descriptors for characters, props, backgrounds. | Closest match to this project's entity/reference architecture; suggests typed memory records. |
| [OneStory](https://arxiv.org/abs/2512.07802) | Adaptive memory and compact cross-shot context for coherent next-shot generation. | Useful for retrieval budget and compact conditioning design. |

## Research Takeaways

### 1. Explicit world state beats prompt-only continuity

CANVAS, VideoMemory, and StoryMem all converge on the same point: long-range consistency needs external memory. TheRoad-I1 should not keep trying to solve recurring location, prop, and state drift with longer scene_detail prose. The current RenderPromptCard direction is correct, but it should sit inside a broader memory loop.

### 2. Visual anchors need semantic labels

StoryMem-style keyframe memory alone can be ambiguous in multi-character scenes. VideoMemory-style entity memory adds semantic descriptors. TheRoad-I1 already has `EntityCanon`, short IDs, visible entities, outlooks, state variants, background refs, and prop refs. The next step is to persist those as a typed story memory rather than resolving them ad hoc per step.

### 3. Candidate verification should be targeted

CANVAS uses QA-based selection. TheRoad-I1 currently has validators and `verify_completion` gates, but most verification is contract shape, source provenance, or residue-oriented. It needs a post-image continuity judge that asks: did this generated image preserve the required character appearance, background anchor, prop state, and shot-specific contract?

### 4. Memory update must happen after generation

The repo has many pre-generation contracts. CANVAS emphasizes updating memory after seeing the actual generated frame. Without that, future shots condition on intended state, not observed state.

### 5. Prompt hygiene remains central

All these systems depend on strong structured inputs. TheRoad-I1's existing policy against semantic regex and scenario-specific prompt contamination is consistent with the research direction. The risk is not that the code lacks rules; it is that too many rules remain prose-only or scattered across prompt versions.

## What Not To Copy Blindly

- Do not add a monolithic "CANVAS agent" layer. TheRoad-I1 already has a step manifest, prompt loader, version registry, and checkpoint contracts.
- Do not push every world-state fact into `RenderPromptCard`; that card should stay shot-level.
- Do not introduce visual memory without source provenance and schema-version gates.
- Do not solve target identity by noun matching or regex over prose. Keep closed-world ID/enum validation only.
- Do not evaluate solely with global preference scores. Use fine-grained continuity checks aligned to contracts.