Story Weave: Multimodal Generative–interactive Narratives with Playable 2D/3D Assets

To introduce Story Weave, a generative-interactive framework that combines a latent world model, a softcontrol narrative planner, and constraint-aware asset binding with the ability to generate both 2D and 3D images.Story Weave works differently than pipelines that only show pictures once the text is finished or that lock designers inside pre-written branches.It optimises the story's content, appearance, sound, and where the user can act, beat by beat.Core modules include (i) a Multimodal Narrative Graph for style and coherence control, (ii) World Weave for changing state dynamics across text, visuals, audio, and scene geometry, (iii) Plot Weave for planning with risk-aware goals and maximum entropy, and (iv) Bind/Render Weave for asset selection that keeps compatibility and rendering that is differentiable.On a set of interactive episodes, Story Weave gets 0.872 ± 0.028 coherence (-graph energy, ↑), 0.742 ± 0.021 text-visual CLIP Score, 0.708 ± 0.019 text-audio alignment, 0.904 ± 0.018 style consistency, and 92.1% ± 2.5% arc completeness (Table 1).Interactivity gets better when there is 1.86 bits of choice entropy, 0.31 bits of information gain per choice, 3.4 mean branch depth, 23.6 minutes of session length, 0.812 preference-fit, 7.8% dropout, and CVaR _0.9=0.18disengagement.The visual quality and performance stay the same at FID 14.2, LPIPS 0.121, Chamfer 0.42×10⁻³, 62.5 FPS, 5.1 GB VRAM, and 320ms P95 load.Human studies show MOS scores of 4.52/4.21/4.63/4.44 (quality/agency/visual/consistency) and SUS scores of 83.1.The preference above baselines is 72% and 78% (Table 6, p≤0.004).Figs.1-6 depict the coherence-novelty frontier, cross-modal retrieval, pacing cumulative distribution functions (CDFs), risk-utility analysis, training dynamics, and latent narrative mappings.