Loom Space: Integration of Partial Structures

• Loom Space is a concept that organizes partial structures into a coherent whole by applying a weaving logic across domains such as creative writing, multi-view image generation, and topology.
• In creative writing, it enables active co-creation using the SOSS (Shape, Observe, Stir, Select) cycle, allowing authors to orchestrate dynamic narrative simulations.
• In multi-view diffusion and topology, Loom Space enhances spatial coherence and enables the construction of latent representations and veering triangulations with measurable quantitative improvements.

“Loom Space” denotes several technically distinct constructs that share a common “weaving” logic: in human–AI co-creation it names a creative-writing environment in which authors Shape, Observe, Stir, and Select a living narrative simulation; in multi-view generative vision it names a shared, spatially continuous latent triplane into which per-view hypotheses are woven; and in low-dimensional topology it names a copy of $\mathbb{R}^2$ equipped with two transverse foliations satisfying cusp and tetrahedron axioms, from which locally veering triangulations of $\mathbb{R}^3$ are canonically built (Andersson et al., 19 May 2026, Federico et al., 7 Jul 2025, Schleimer et al., 2021, Baik et al., 2022). This suggests that the term is best understood not as a single standardized object, but as a family of structured media in which local data are combined into a coherent global configuration.

1. Terminological scope

The term appears in at least three research domains with non-identical meanings. In each case, however, the defining operation is not mere storage or display, but an organized integration of partial structures into an evolving whole.

Domain	Object called “Loom Space”	Core mechanism
Human–AI creative writing	The Loom space as the concrete realization of an “active medium” framing	SOSS: Shape, Observe, Stir, Select
Multi-view image generation	A shared, spatially continuous latent representation	Per-view splatting, cross-attention fusion, learned weaving
Topology and geometric group theory	A copy of $\mathbb{R}^2$ with transverse foliations satisfying axioms	Rectangles, cusps, tetrahedron rectangles, induced veering triangulation

In the creative-writing system Loom, the space is operational and interface-level: it is the environment within which authors orchestrate simulated narrative agents rather than merely judge outputs (Andersson et al., 19 May 2026). In LoomNet, the space is representational and latent: a discrete triplane $(A_{XY},A_{YZ},A_{XZ})$ encoding a consensus 3D scene embedding from which all views are rendered (Federico et al., 7 Jul 2025). In topology, a loom space is axiomatic and combinatorial: a foliated plane whose skeletal rectangles induce an ideal cellulation that is locally veering and whose realization is homeomorphic to $\mathbb{R}^3$ (Schleimer et al., 2021). A related construction starts instead from a veering pair of laminations on $S^1$, forms a quotient weaving, and obtains a loom space as the universal cover of the marked weaving (Baik et al., 2022).

2. Loom as an active creative medium

In “Material for Thought: Generative AI as an Active Creative Medium,” Loom is introduced to reject a decision-support framing in which generative AI acts primarily as an advisor whose suggestions must be accepted or rejected. Drawing on Schön’s theory of reflective practice, Loom instead treats the model as a malleable creative material, “akin to clay in a potter’s hands,” and repositions the human as an active orchestrator of possibility (Andersson et al., 19 May 2026).

Its operative cycle is SOSS. In Shape, the author defines the “conditions for emergence”: cast, personalities, goals, flaws, secrets, interrelationships, setting, and initial tone or genre constraints. In Observe, once shaping is committed, the system rolls the clock forward and each character—implemented as an independent LLM agent—speaks or acts in turn. In Stir, the author intervenes through the “Narrator Bar,” where stage directions such as “A storm breaks the power in the mansion” are perceived by all agents on their next turn. In Select, every stir event becomes a branch point in a version-control–style timeline, enabling replay of common history up to a fork, side-by-side comparison of divergent continuations, and commitment to the branch that appears richest (Andersson et al., 19 May 2026).

The system architecture is a single-page web application with four main panels corresponding to these phases. A stateless API layer routes character-turn requests to a hosted LLM, and a central state manager maintains three structures: a rolling “working memory” of the last five utterances across all agents, a set of “long-term memories” selected by impact score, and the branch timeline data structure. Each agent’s next utterance is generated from a prompt of the form

$$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$

Agents are defined by the tuple $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$. For each new utterance $U$, the system computes an impact score $S(U)$ via a lightweight LLM meta-evaluation and promotes $\mathbb{R}^3$0 into long-term memory when

$\mathbb{R}^3$1

The paper states explicitly that no explicit loss functions or convergence-proof metrics appear; the relevant convergence is qualitative, namely LLMs’ tendency toward sycophancy and overly neat conflict resolution (Andersson et al., 19 May 2026).

The significance of this formulation is methodological. The design implications extracted from Loom are to shift human effort from output evaluation to material orchestration, to make convergence visible and disruptable, and to treat friction as generative rather than punitive. The paper also states that the same principles apply beyond narrative writing to design ideation, musical improvisation, and data visualization, provided the goal is exploration of a creative space rather than post hoc trust calibration (Andersson et al., 19 May 2026).

3. Loom Space in multi-view diffusion

In “LoomNet: Enhancing Multi-View Image Generation via Latent Space Weaving,” the Loom Space is defined as “the shared, spatially continuous latent representation into which each per-view hypothesis is woven” (Federico et al., 7 Jul 2025). The setting is consistent multi-view image generation from a single image, where lack of spatial consistency degrades downstream 3D mesh quality.

The construction begins with $\mathbb{R}^3$2 per-view diffusion UNets, each producing at decoder level $\mathbb{R}^3$3 a feature map

$\mathbb{R}^3$4

for view $\mathbb{R}^3$5. For each spatial location $\mathbb{R}^3$6, a ray is cast from the camera center through pixel $\mathbb{R}^3$7, and $\mathbb{R}^3$8 points $\mathbb{R}^3$9 are sampled uniformly within a cube of side $\mathbb{R}^2$0 centered at the scene origin. Each sampled point carries an enriched feature vector

$\mathbb{R}^2$1

where $\mathbb{R}^2$2 is the ray direction, $\mathbb{R}^2$3 is the depth of $\mathbb{R}^2$4, and $\mathbb{R}^2$5 is a harmonic embedding. These features are then projected onto the three coordinate planes $\mathbb{R}^2$6 and inverse-bilinearly splatted to produce per-view splatted planes $\mathbb{R}^2$7.

For each orientation $\mathbb{R}^2$8, LoomNet maintains a learnable fusion plane

$\mathbb{R}^2$9

At each spatial coordinate $(A_{XY},A_{YZ},A_{XZ})$0, a single learnable query $(A_{XY},A_{YZ},A_{XZ})$1 attends to keys and values derived from all per-view splatted features, yielding a pixelwise cross-attention update. The three fused planes are then stacked into a tensor $(A_{XY},A_{YZ},A_{XZ})$2, and a weaving block applies time-conditioned normalization, self-attention, and an MLP residual block:

$(A_{XY},A_{YZ},A_{XZ})$3

The fusion-plus-weaving cycle is repeated multiple times at each decoder level, with the paper giving examples of 3, 4, 6, and 8 iterations. The communication module is inserted in each of the four decoder levels of every per-view UNet. Its cross-attention uses 8 heads with head size 64. At the final decoder level, latent rendering samples points along each outgoing ray from the fused Loom Space, interpolates their global features, concatenates them with the original pixel feature, and refines the result via an MLP decoder (Federico et al., 7 Jul 2025).

Experimentally, on Google Scanned Objects with fixed elevation at $(A_{XY},A_{YZ},A_{XZ})$4, LoomNet reports PSNR 21.60, SSIM 0.901, LPIPS 0.070, and 15 s, compared with Zero-1-to-3 at 17.79/0.796/0.201/7 s, SyncDreamer at 20.11/0.829/0.159/60 s, and EpiDiff at 20.49/0.855/0.128/17 s. For 3D reconstruction, the reported Chamfer Distance and Volume IoU are 0.0260 and 0.5366 for LoomNet, compared with 0.0543/0.3358, 0.0496/0.4149, and 0.0429/0.4518 for the same baselines. The ablations identify three points: replacing cross-attention fusion with mean fusion drops SSIM from 0.901 to 0.864 and raises LPIPS from 0.070 to 0.091 while reducing runtime to 12.5 s; halving $(A_{XY},A_{YZ},A_{XZ})$5 to 8 reduces PSNR only mildly from 21.60 to 21.39 while cutting 1 s; and removing harmonic positional encoding causes the largest collapse, with SSIM going to 0.844 and LPIPS to 0.128 (Federico et al., 7 Jul 2025).

The paper also states two limitations. First, the weaving injects spatial coherence but does not explicitly enforce perfect geometric consistency; SyncDreamer’s volumetric fusion can sometimes yield smoother but semantically coarser geometry. Second, generation and 3D reconstruction remain separate stages. Proposed extensions include replacing the triplane Loom Space with voxel grids or deformable fields, conditioning fusion with textual or style embeddings for multi-modal editing, adding a temporal axis for dynamic scene synthesis, and unifying shape and appearance generation so that textured meshes or radiance fields are produced directly (Federico et al., 7 Jul 2025).

4. Loom spaces in low-dimensional topology

In “From loom spaces to veering triangulations,” a loom space is a copy of $(A_{XY},A_{YZ},A_{XZ})$6 equipped with two transverse nonsingular foliations $(A_{XY},A_{YZ},A_{XZ})$7 and $(A_{XY},A_{YZ},A_{XZ})$8, called the upper and lower foliations, satisfying two axioms: the cusp axiom and the tetrahedron axiom (Schleimer et al., 2021). The ambient setting is a plane $(A_{XY},A_{YZ},A_{XZ})$9, and by the Poincaré–Hopf theorem and Jordan separation, each leaf of either foliation is properly embedded and separates $\mathbb{R}^3$0.

The local combinatorics are organized by rectangles. An affine rectangle is an open set $\mathbb{R}^3$1 together with a homeomorphism

$\mathbb{R}^3$2

such that arcs with fixed $\mathbb{R}^3$3 lie in leaves of $\mathbb{R}^3$4, while arcs with fixed $\mathbb{R}^3$5 lie in leaves of $\mathbb{R}^3$6. Orientation of the foliations gives the four material sides “south,” “east,” “north,” and “west.” A cusp rectangle is obtained when $\mathbb{R}^3$7 extends continuously to $\mathbb{R}^3$8, sending the missing corner to infinity in $\mathbb{R}^3$9; analogous definitions give the other cusp types. A tetrahedron rectangle is one for which $S^1$0 extends to $S^1$1 for some $S^1$2. Intermediate types include edge rectangles and face rectangles (Schleimer et al., 2021).

The loom-space axioms are then: every cusp side of every cusp rectangle contains an initial nontrivial open segment lying in some ordinary rectangle; and every rectangle is contained in some tetrahedron rectangle. The basic examples listed are: an Anosov map on the torus after removing one periodic point and lifting eigenfoliations; pseudo-Anosov maps on higher-genus surfaces after removing singularities and lifting measured foliations; pseudo-Anosov flows without perfect fits after removing singular orbits, lifting, and quotienting by flow lines; and Agol’s veering ideal triangulations, whose lifted branched surfaces yield foliations on $S^1$3 satisfying the axioms (Schleimer et al., 2021).

Several structural results sharpen the definition. Rectangles form a basis for the topology of $S^1$4. In any tetrahedron rectangle there are exactly 4 face-rectangle subsets and exactly 6 edge-rectangle subsets; each face rectangle lies in exactly 2 tetrahedron rectangles, and every edge rectangle is contained in finitely many face rectangles. The Keane condition asserts that for a tetrahedron rectangle, the missing points on the south and north sides differ, and similarly on west and east. Cusps are equivalence classes of cusp rectangles under finite chains of cusp-side inclusions. Staircases based at a corner $S^1$5 define axis rays in the two foliations, and the astroid lemma states that the projection of exterior cusps onto either axis has no accumulation point in the interior but accumulates at both endpoints. A corollary is that cusp leaves are dense in both foliations, and non-cusp leaves are also dense (Schleimer et al., 2021).

5. Induced triangulations, quotients, and group actions

Given a loom space $S^1$6, the paper defines an ideal cellulation $S^1$7 whose model $S^1$8-simplices correspond to skeletal rectangles of codimension $S^1$9: vertices are cusps, edges are edge-rectangles, faces are face-rectangles, and tetrahedra are tetrahedron-rectangles. Gluing is determined by side inclusion of smaller skeletal rectangles into larger ones, and the $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$0-skeleton is deleted at the end (Schleimer et al., 2021).

The resulting realization has strong global properties. Lemma 6.4 states that $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$1 is a noncompact, connected, orientable $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$2-manifold. Corollary 6.9 gives a taut structure by assigning angle $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$3 to each model edge coming from an edge rectangle that spans its containing tetrahedron rectangle and angle $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$4 otherwise. Corollary 6.10 shows that orienting the two foliations induces a canonical $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$5-coloring of the edges—red for east–west and blue for north–south—making $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$6 locally veering. Theorem 7.1 then proves that for any loom space $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$7, the realization of its induced locally veering triangulation is homeomorphic to $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$8 (Schleimer et al., 2021).

“Groups acting on veering pairs and Kleinian groups” supplies a complementary entry point. It begins with a veering pair $$\text{Character Profile} + \text{LongTermMemories} + \text{WorkingMemory} + \text{RecentStageDirections} + \text{Observed Dialog} \rightarrow \text{Next Action}.$$9 of lamination systems on $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$0, where each $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$1 is quite full and loose, the endpoint sets are dense and disjoint, and every non-leaf gap of one lamination interleaves with a gap of the other. From this one defines the stitch space

$$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$2

and then quotients by the weaving relation $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$3 to obtain the cusped weaving

$$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$4

The ordinary part is $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$5, obtained by removing cusp classes. Lemma 7.14 states that $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$6 is homeomorphic to $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$7 with two singular measured foliations induced from $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$8. The universal cover of the marked weaving is then a loom space, to which the Schleimer–Segerman functor applies, producing an ideal transverse veering triangulation $$(\text{Profile}, \text{WorkingMemory}, \text{LongTermMemory})$$9 of $U$0 (Baik et al., 2022).

This leads to a geometrization-type result. Theorem 17.15 states that if $U$1 is a veering pair and $U$2, then $U$3 is isomorphic to the fundamental group of an irreducible $U$4-orbifold of the form $U$5, with $U$6, endowed with a transverse veering triangulation; under a cofinite hypothesis on the $U$7-action, the orbifold admits a complete hyperbolic structure. The paper emphasizes that no ambient $U$8-manifold need exist a priori. As an explicit example, a pseudo-Anosov homeomorphism of the once-punctured torus yields stable and unstable laminations forming a veering pair whose construction recovers the usual veering triangulation of the hyperbolic mapping torus (Baik et al., 2022).

6. Comparative themes, limitations, and misconceptions

Across these uses, a recurring pattern is the conversion of many partial strands into one structured medium. In the creative-writing system, the strands are character profiles, working memory, long-term memory, stage directions, and branch history; in LoomNet they are per-view latent hypotheses projected onto three planes; in topology they are foliated rectangles, cusps, and skeletal inclusions; and in the lamination-theoretic construction they are linked leaves in stitch space modded out by a weaving relation (Andersson et al., 19 May 2026, Federico et al., 7 Jul 2025, Schleimer et al., 2021, Baik et al., 2022). This suggests that the shared content of the term is not a specific data structure, but an architecture of integration.

Several misconceptions are ruled out by the source texts. First, Loom Space is not a single universally accepted definition. The same phrase names an interface paradigm, a latent triplane representation, and an axiomatic foliated plane in different literatures. Second, the creative-writing Loom is not framed as prompt engineering plus output filtering; it is explicitly designed to move effort from correctness evaluation toward orchestration and reflection-in-action (Andersson et al., 19 May 2026). Third, LoomNet’s Loom Space does not explicitly enforce perfect geometric consistency, even though it improves spatial coherence and reconstruction metrics (Federico et al., 7 Jul 2025). Fourth, the topological loom space is not metaphorical; it is a sharply specified object with rectangle axioms, density statements, finiteness properties, and a functorial passage to locally veering triangulations whose realization is $U$9 (Schleimer et al., 2021).

The future directions also differ by field. The creative-medium paper generalizes its principles to design ideation, musical improvisation, and data visualization. LoomNet proposes alternative latent representations, multi-modal conditioning, temporal weaving for dynamic scenes, and direct generation of textured meshes or radiance fields. The topology paper points toward an equivalence between the category of loom spaces and the category of veering triangulations of $S(U)$0, and further directions include ergodic and counting properties, extensions of the astroid lemma to measured and singular-metric contexts, and aperiodic loom spaces. The group-theoretic paper extends the reach of loom-space constructions beyond settings where a $S(U)$1-manifold is given in advance, framing them as a route from abstract veering data on $S(U)$2 to hyperbolic $S(U)$3-orbifold groups (Andersson et al., 19 May 2026, Federico et al., 7 Jul 2025, Schleimer et al., 2021, Baik et al., 2022).