Source-linked AI summary
Live Artifacts: Authoring Dynamic Media via Live Layers Encapsulating Generative Specifications
Leixian Shen, Haotian Li, Hugo Romat, Fanny Chevalier, Nicolai Marquardt, Nathalie Riche
TL;DR
Conventional GenAI outputs become static files because their generative logic is discarded, limiting context-dependent evolution across modalities. The paper introduces Live Artifacts and LiveCanvas, which preserve generative specifications in live, linked visual layers within a media-first canvas. Examples and a qualitative study with six professionals suggest a shift toward authoring responsive artifacts while retaining familiar authoring practices, with clear limits where deterministic outputs are required.
Problem
Conventional GenAI creation treats generation as ephemeral and discards the logic needed for persistent, context-responsive, cross-modal media.
Method
LiveCanvas embeds generative specifications in static or live visual layers connected by constrained dependencies that govern propagation within a visual authoring workflow.
Results
A gallery and qualitative study with six multimedia professionals found that LiveCanvas supported reasoning about persistence, dependencies, and change in a visual canvas.
Takeaways & Limitations
Live Artifacts let media retain generative behavior while preserving recognizable structure and cross-modal coherence without shifting authoring into programming abstractions.
Takeaways & Limitations
Nondeterministic regeneration remains unsuitable when strict predictability, precise reproduction, or fully specified outcomes are essential.
Abstract
from arXiv · showhide
We frame Live Artifacts as a class of persistent generative media between static assets and interactive software. Unlike conventional generative outputs that collapse into static files, Live Artifacts retain their generative logic as a persistent media property, enabling continuous context-dependent regeneration. Time, location, or live data become part of their generative specifications, initiating coordinated updates across modalities (e.g., adapting text, visuals, and audio together) while preserving composition, semantics, identity, and cross-modal coherence. To facilitate experimentation with this medium, we present LiveCanvas, an authoring system that reconceptualizes visual layers as live generative specifications with explicit mutability and constrained dependencies. Creators orchestrate dynamic behaviors and manage generative persistence within a visual canvas rather than through programming, defining what remains stable, what can change, and how changes propagate. We evaluate Live Artifacts through a gallery of responsive examples and a qualitative study with six professionals, finding that LiveCanvas facilitates a shift from composing static outputs to crafting responsive generative artifacts while remaining aligned with familiar authoring practices.
1 Introduction
Live Artifacts address the gap between static media and full-fledged software by preserving generative logic within media and enabling context-responsive, cross-modal evolution. LiveCanvas supports this through live visual layers and constrained dependencies in a familiar media-first workflow.
- Motivation: Current GenAI workflows usually discard generative logic after one-shot rendering, leaving static outputs unable to adapt to changing context.Natural-language tools can produce multimodal content, but the underlying generation process is not retained.
- Live Artifacts: Live Artifacts are dynamic media whose persistent specifications respond to external context while preserving identity, composition, semantics, and cross-modal coherence.Examples include location-aware promotional displays and social posts that regenerate visuals, captions, and audio around a recognizable structure.
- Authoring challenge: Authoring Live Artifacts requires controlling what remains stable and what evolves, a boundary historically handled by moving from the design canvas into software development.The paper frames this as a central tension involving identity, layout, and continuous change.
- LiveCanvas: LiveCanvas keeps a media-first abstraction by embedding generative specifications in visual layers, aligning authoring with spatial composition, hierarchy, grouping, and dependencies.Authors can designate layers as static or live and establish constrained links through which context- or time-driven changes propagate.
- Evaluation: The paper presents a gallery and a qualitative study with six multimedia professionals examining how creators reason about persistence, dependencies, and change.The study focuses on shifts from composing static outputs to orchestrating generative behaviors rather than task efficiency.
- Contributions: The contributions define Live Artifacts, introduce LiveCanvas, and report examples and preliminary findings about authoring live generative media.Together, they outline a design space for media with capabilities traditionally associated with software.
2 Related Work
Prior systems separately support spatial media composition or persistent generative logic, leaving their combination difficult. Live Artifacts embed generative specifications in visual layers to unite direct manipulation with ongoing re-evaluation.
- Interactive media: Interactive-document systems preserve executable or reactive relationships across code, prose, data, and visualizations for analysts and end-users.Examples include Jupyter, Observable, Living Papers, CrossData, Mavo, Varv, Streamlit, and Webstrates.
- Authoring gap: Dynamic media authoring must balance spatial arrangement with generative behavior that changes content over time, while most existing tools emphasize one side.This motivates a media-native approach that retains both composition and persistence.
- Media-centric systems: Media-centric tools emphasize direct manipulation and composition, but generative operations are typically one-time and flattened into static pixels.Constraint-based predecessors demonstrate that persistent declarative relationships can instead propagate changes within documents.
- Logic-centric systems: Logic-centric systems foreground declarative pipelines or systematic exploration of generative design spaces rather than spatial composition.Examples include node-based environments, structured creative tools, and canvas-based systems that blend layout with iterative generation.
- Live Artifacts: Live Artifacts embed persistent generative specifications directly within spatial layers, combining direct manipulation with re-evaluation triggered by context changes or timed events.They extend complementary strengths of media-centric and logic-centric systems rather than simply reproducing either approach.
- Layers: Layers organize media as separable, spatially ordered elements and provide a cognitive and computational bridge to rendering pipelines.Prior work applies layers across image compositing, animation, immersive painting, text, and generative video.
3 Live Artifacts
Live Artifacts are dynamic media that preserve generative specifications and relationships, allowing context- or time-based regeneration while maintaining an underlying identity. They occupy a space between static media and software through bounded cross-modal propagation rather than open-ended control flow.
- Core properties: Live Artifacts retain generative specifications as inherent media properties, allowing outputs to change while preserving an underlying skeleton.Authors choose which elements remain static and which remain live, such as a recognizable layout, logo, or title.
- Core properties: Contextual and temporal responsiveness lets artifacts incorporate time, location, or web data into specifications and regenerate content over time or at consumption.URLs can provide external data for adapting visuals or other components to a viewer’s locale.
- Core properties: Cross-modal propagation links parent and child layers so changes in text can propagate to imagery or audio while preserving coherence.Relationships are intentionally limited, such as each layer referring to a single parent, to avoid complex or cyclic flows.
- Examples: Examples combine persistent generative logic, bounded propagation, and selective responsiveness so media can change while remaining recognizably the same artifact.The examples include context-aware promotional content, ambient inspiration, and personalized audio news.
- Positioning: Live Artifacts differ from programs because temporal change concerns when media is re-evaluated, not how general processes unfold through open-ended control flow.They remain inspectable as visible media structures rather than merely as outputs of execution.
4 Authoring Live Artifacts with LiveCanvas
LiveCanvas extends familiar visual layers into persistent, generative components that authors can compose, connect, and selectively regenerate. Its constrained dependencies and refresh model support contextual, coherent updates without requiring programming or general execution logic.
- Layer-based authoring: LiveCanvas makes layers both composition units and persistence units, with static layers storing media and live layers storing re-evaluable generative specifications.Authors can derive one layer from another, creating a single-parent dependency that remains acyclic and interpretable.
- Canvas organization: The canvas separates freeform exploration from output composition, while the layer stack manages only output layers, Z-order, and opacity.A merge layer can recompute selected layers into a unified visual composite while preserving spatial relationships and transparency.
- Layer states: Authors choose whether live layers become static, remain live, or regenerate automatically over time, defining what stays stable and what changes.Cycling among these states helps refine specifications and bound artifact identity and non-deterministic content.
- Contextual layers: External references such as location, APIs, or webpages provide contextual grounding during re-evaluation of live layers.In the coffee advertisement, geographic and weather data update the tagline and imagery.
- Dependencies: Dependencies define how variation propagates by injecting parent content into child specifications and triggering dependent layers when parents re-evaluate.A weather-responsive tagline can provide the beverage concept used to regenerate a derived illustration without authored if-then rules.
- Constraints: LiveCanvas prioritizes re-evaluation over execution and limits dependencies to an acyclic single-parent structure as a trade-off between flexibility and legibility.The system does not support authored branching, interaction logic, or multi-step procedural workflows.
5 User Study
The qualitative study found that LiveCanvas shifted creators from composing fixed outputs toward specifying persistent behavior, while preserving recognizable structure and enabling cross-modal variation.
- Study and artifacts: Six multimedia professionals used LiveCanvas to create varied Live Artifacts spanning different scenarios and expressive goals.The exploratory study examined how creators reasoned about persistence, dependencies, and authoring practice rather than task efficiency.
- Experiencing liveness: Live Artifacts supported variation within authored bounds, preserving recognizable identity, composition, style, and structure across realizations.Participants imagined targeted advertising, educational content, and situated installations adapting to audiences or context without manual duplication.
- Authoring behavior: Authors described Live Artifact creation as specifying constraints, relationships, and invariants that shape behavior over time rather than assembling a result sequentially.The linked-layer model encouraged reasoning about logical relationships instead of strict linear workflows.
- Authoring behavior: Participants treated spatial layout as a way to express logical relationships rather than merely optimize a final composition.Because live-layer content remained unfixed, layout became a looser constraint and attention shifted toward associations between elements.
- Tradeoffs: Participants valued cross-modality coordination but accepted reduced pixel-level determinism as a tradeoff for structural consistency.They considered unpredictable details potentially problematic in high-stakes scenarios while still valuing coordinated multimodal authoring.
- Authoring comparison: Compared with prompt-centric generation and programming, LiveCanvas offered visible constraints and direct manipulation for reasoning about generative media.Participants contrasted this with repeated prompt-based attempts to obtain suitable layouts and with implementation-focused programming concerns.
6 Discussion and Future Work
The discussion positions Live Artifacts as a media-native middle ground that preserves generative persistence and structural control without full programming logic. It also identifies boundaries involving determinism, interaction, latency, scalability, and longer-term evaluation.
- Specifications as a Media Property vs. One-Shot Generation: Live Artifacts embed selected generative specifications in media so persistence supports context-aware variation while preserving recognizable identity.This reframes generative logic as a media property rather than a discarded one-shot operation or software pipeline.
- Structural vs. Pixel-precise Control: The approach trades pixel-precise control for control over composition, layout, semantic relationships, and recognizable architectural intent.Pixel-final determinism remains important where precise reproduction or fully specified outcomes are required, including legal, scientific, and high-stakes commercial media.
- Expressiveness vs. Legibility in Authoring: Constrained dependencies keep layer relationships legible, balancing expressive power against direct manipulation and authorial comprehensibility.LiveCanvas occupies an intermediate form between prompt-centric generation and code-based systems by retaining persistence and propagation without full control flow or state.
- Augmenting Media vs. Simplifying Programming: Supporting gradual evolution, accumulated memory, or repeated-user responsiveness would require state retention, event handling, and branching behavior.These mechanisms would move beyond simple re-evaluation and may require a programming model, reintroducing constructs the approach seeks to avoid.
- Augmenting Media vs. Simplifying Programming: The boundary between media-native generative behavior and programming remains open, particularly for temporal continuity and limited user interaction.The unresolved question is what can be supported through constrained mechanisms without losing media-first legibility.
- Limitations: Cloud-generation latency limits real-time interaction more than ambient or background updates.The paper suggests local models for fast previews and cloud models for final output as a possible hybrid response.
- Limitations: Managing many live layers may become difficult as artifacts grow, motivating grouping and hierarchical organization to preserve dependency legibility.The study and gallery indicate that relatively few layers can already be expressive and compelling.
- Future work: Longer deployment in real-world contexts is needed to assess sustained value, organically emerging use cases, and LiveCanvas’s design tradeoffs.The current study provides early insights and preliminary feedback rather than long-term evidence.
7 Conclusion
The paper introduces Live Artifacts as persistent generative media that respond to time, context, or live data while preserving recognizable structure and cross-modal coherence. LiveCanvas makes these properties authorable through live, linkable visual layers in a media-first workflow.
- Conclusion: Live Artifacts retain generative specifications as an ongoing media property, enabling context-responsive regeneration with recognizable structure and cross-modal coherence.They occupy a design space between traditional media authoring and software development.
- Conclusion: LiveCanvas extends visual layers with live and linkable behavior, allowing authors to specify generative persistence through spatial composition and direct manipulation.The conclusion frames this as a media-native alternative to programming- or chat-centric abstractions.