Co-Director in AI, Film & Corporate Governance

• Co-Director is a term for systems or relationships where multiple directing influences coordinate, seen in AI activation steering, video storytelling, film previsualization, and corporate governance.
• It uses explicit factorization methods—such as vector interpolation in neural networks and weighted graph projections—to manage multi-agent control across technical domains.
• Applications include creative collision in transformer models, agentic generative video storytelling pipelines, and analytical frameworks in board interlock studies, highlighting both cooperative and asymmetric dynamics.

In recent technical literature, co-director does not denote a single fixed object. It names several domain-specific constructs that share a common structural idea: coordinated control by more than one directing influence. In LLM activation steering, a co-director setting is the simultaneous mixture of two directors’ activation-space signatures; in agentic video generation, Co-Director is a hierarchical multi-agent system that optimizes high-level creative choices across an end-to-end storytelling pipeline; in film tooling, related systems function as virtual co-directors for previsualization, storyboard alignment, shot design, and Foley control; and in corporate network analysis, a co-director relationship is a graph edge between two directors who sit on the same company board (Sahoo et al., 15 Jun 2026, Song et al., 27 Apr 2026, Nan et al., 16 Mar 2026, Wei et al., 27 Jul 2025, Li et al., 20 Mar 2026, Sancheti et al., 24 Mar 2026).

1. Terminological range and disambiguation

The term is used in at least four technically distinct senses.

Sense	Operational meaning	Representative work
Activation-space co-direction	Simultaneous mixture of two directors’ persona vectors in a transformer residual stream	(Sahoo et al., 15 Jun 2026)
Agentic video storytelling	Hierarchical multi-agent orchestration of global creative strategy and local multimodal refinement	(Song et al., 27 Apr 2026)
Film-production assistance	A virtual or interactive co-director coordinating script, scene, blocking, camera, storyboard, or sound decisions	(Nan et al., 16 Mar 2026, Wei et al., 27 Jul 2025, Li et al., 20 Mar 2026)
Corporate-network analysis	Two directors linked through shared company board membership	(Sancheti et al., 24 Mar 2026)

A further source of ambiguity is the word director itself. In soft-matter physics, the director is the orientational field of a nematic medium, written as $\mathbf n(\mathbf r,t)$; that usage concerns liquid-crystal order rather than collaborative authorship, agency, or governance (Turiv et al., 2015, Safdari et al., 2021). A common misconception is therefore to treat all uses of director as semantically adjacent. The literature instead shows a terminological split between cinematic or organizational “directing” and the nematic director field of continuum physics.

2. Co-direction in activation steering

In "Creative Collision," a “Co-Director” setting means that generation is steered not toward a single cinematic persona, but toward a simultaneous mixture of two directors’ activation-space signatures—Steven Spielberg and Martin Scorsese—inside a 14-billion-parameter open-weight decoder-only transformer with 40 layers (Sahoo et al., 15 Jun 2026). The persona vectors are extracted from a curated screenplay-derived contrast corpus by mean-difference activation contrast against a shared neutral baseline, with mean-pooled residual-stream activations $h^{(l)}(x)\in\mathbb{R}^d$. The layerwise vectors are

$$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$

$$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$

followed by $\ell_2$ normalization. The mixed co-director vector is a linear interpolation,

$$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$

and the inference-time intervention is

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$

applied over $\mathcal{L}=\{20,21,\dots,38\}$ with $\alpha\in\{0,0.25,0.5,0.75,1.0\}$ and $\lambda\in\{0.5,1.0,1.5,2.0\}$ (Sahoo et al., 15 Jun 2026).

The paper evaluates this regime on five axes: moral valence, generation coherence, surface style, directional dominance, and vector geometry. Its central finding is that co-direction is not a balanced blend. At the key layer 28, the cosine similarity between the normalized Spielberg and Scorsese vectors is approximately $h^{(l)}(x)\in\mathbb{R}^d$0, so they are neither antiparallel nor orthogonal. In output space, Spielberg exhibits strong directional dominance: at $h^{(l)}(x)\in\mathbb{R}^d$1, the stylometric classifier assigns generated text to Spielberg with probability near $h^{(l)}(x)\in\mathbb{R}^d$2 for $h^{(l)}(x)\in\mathbb{R}^d$3, $h^{(l)}(x)\in\mathbb{R}^d$4, and $h^{(l)}(x)\in\mathbb{R}^d$5, remains around $h^{(l)}(x)\in\mathbb{R}^d$6 at $h^{(l)}(x)\in\mathbb{R}^d$7, and reaches only about $h^{(l)}(x)\in\mathbb{R}^d$8 at $h^{(l)}(x)\in\mathbb{R}^d$9, where the Spielberg component disappears entirely (Sahoo et al., 15 Jun 2026). The authors interpret this as asymmetric representational power, plausibly linked to pretraining priors and helpfulness- or safety-oriented tuning.

The co-director mixture is also nonlinear in moral behavior. At $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$0, moral valence is most negative at $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$1, with $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$2, then drifts back toward neutral by $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$3, where $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$4. At $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$5, the moral signal collapses across all $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$6, with $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$7 and variance below $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$8, which the paper attributes to perturbations moving residual states too far off the natural text manifold (Sahoo et al., 15 Jun 2026).

A practically important result is the coherence valley of directorial collision. For $$v_{\text{SPL}}^{(l)}=\frac{1}{N_{\text{SPL}}}\sum_i h^{(l)}(x_i^{\text{SPL}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$9, pure Spielberg and pure Scorsese steering produce the worst perplexities, reaching roughly $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$0 and $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$1 respectively at $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$2, whereas intermediate mixtures such as $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$3 or $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$4 remain around $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$5–9. The geometric explanation is norm reduction in the interpolated vector:

$$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$6

where $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$7. With $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$8, the mixed vector at $$v_{\text{SCO}}^{(l)}=\frac{1}{N_{\text{SCO}}}\sum_i h^{(l)}(x_i^{\text{SCO}})-\frac{1}{B}\sum_j h^{(l)}(b_j),$$9 has norm about $\ell_2$0, so a nominal steering coefficient induces a smaller effective perturbation than in pure steering (Sahoo et al., 15 Jun 2026).

Layerwise localization further identifies a shared “moral-tone substrate”. Both directors’ moral-tone effects peak at layer 28: approximately $\ell_2$1 for Scorsese and $\ell_2$2 for Spielberg under single-layer steering with $\ell_2$3. The paper argues that the two personas push on a common latent axis of narrative morality, while remaining asymmetric, entangled semantic directions rather than exact opposites (Sahoo et al., 15 Jun 2026).

3. Co-Director as agentic generative video storytelling

"Co-Director" also names a specific system for agentic generative video storytelling that turns high-level intent into a coherent multi-shot video with visuals, motion, and audio (Song et al., 27 Apr 2026). The framework formalizes storytelling as a global optimization problem over a compact creative parameterization,

$\ell_2$4

where the dimensions are Creative Strategy ($\ell_2$5Informational, Transformational, Comparative), Narrative Mode (Analytical, Vignette, Narrative Drama), and Aesthetic Archetype (Clarity/Energy, Cinematic Premium, Minimalist Focus, Kinetic Grit). This yields $\ell_2$6 creative combinations (Song et al., 27 Apr 2026).

The architecture has four stages: Orchestrator Agent $\ell_2$7, Pre-Production Agent $\ell_2$8, Production Agent $\ell_2$9, and Post-Production Agent $$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$0. The orchestrator parses user input and reference visuals, samples a creative configuration from a multi-armed bandit, and propagates it into every downstream stage through a shared conditioning context $$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$1. Pre-production produces a creative brief, storyline, visual assets, and storyboard. Production generates keyframes, clips, and audio:

$$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$2

and post-production assembles the final video $$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$3 (Song et al., 27 Apr 2026).

Global search is handled by UCB1, with an LLM-driven warm start and a factored reward

$$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$4

This factored reward is designed to improve credit assignment relative to a single scalar score. Local repair is handled by two self-refinement loops: a storyline refinement stage that scores hook quality, narrative cohesion, product integration, emotional engagement, and prompt adherence, and a multimodal keyframe refinement stage that jointly inspects the sequence of keyframes for identity drift, narrative flow, product appeal, and prompt adherence (Song et al., 27 Apr 2026).

The evaluation benchmark, GenAD-Bench, contains 400 scenarios built from 50 fictional brands, 4 products per brand, and 2 demographic scenarios per product, with 200 hillclimbing scenarios and 200 validation scenarios. On GenAD-Bench, Co-Director ($$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$5) achieves 82.1 VAF, 91.4 DA, 82.0 MA, 70.2 VQ, and 81.4 Avg, outperforming Random Search ($$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$6) at 75.7 Avg and the Base Agentic Pipeline ($$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$7) at 68.5 Avg (Song et al., 27 Apr 2026). In the reported human MOS study, Co-Director scores 4.22 VAF, 4.41 DA, 3.65 MA, 3.58 VQ, and 3.96 Avg, again leading the compared methods (Song et al., 27 Apr 2026). The ablations show that removing keyframe refinement induces a 9.8-point VAF drop, and replacing factored reward with scalar reward reduces average performance from 79.0 to 76.4 in the ablation setting (Song et al., 27 Apr 2026).

4. Virtual co-directors in previsualization, storyboarding, shot design, and sound

A broader cluster of systems uses the co-director concept functionally rather than nominally. "Mind-of-Director" is explicitly described as acting less like a one-shot generator and more like a virtual co-director that coordinates Script Development, Virtual Scene Design, Character Behaviour Control, and Camera Planning inside a game engine (Nan et al., 16 Mar 2026). It formalizes previz as

$$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$8

uses Discuss-Revise-Judge and Debate-Judge-Validation protocols, and produces a previz sequence in approximately 25 minutes per idea. Reported results include CLIP alignment 26.35 vs. 25.48, collision rate 0.83\% vs. 2.27\%, blocking loss 0.48 vs. 0.86, and camera collision rate 2.1\% vs. 13.7\% against named baselines (Nan et al., 16 Mar 2026).

"CineVision" addresses a different collaboration problem: director–cinematographer alignment during early pre-production (Wei et al., 27 Jul 2025). Its workflow integrates script input, SQL-style filtering, text-to-film database matching from MovieNet, storyboard regeneration, relighting, style emulation, and character or costume design. The system is built around a two-tier hierarchical menu and a weighting scheme

$$v_\alpha^{(l)}=(1-\alpha)v_{\text{SPL}}+\alpha v_{\text{SCO}},\qquad \alpha\in[0,1],$$9

as printed in the paper. In a 24-participant lab study, CineVision obtains a collaboration score of 6.5 (0.53), compared with 5.12 (1.54) for DALL·E 3 and 4.13 (1.46) for hand-drawn storyboards; it also scores better on usefulness and ease of use under the reported protocol (Wei et al., 27 Jul 2025). The paper is explicit, however, that CineVision does not yet provide production-grade camera metadata such as lens focal length, T-stop, camera height, sensor crop, or automatic continuity checks.

"FoleyDirector" transfers the directing metaphor to video-to-audio generation by enabling users to act as Foley directors through Structured Temporal Scripts (STS) (Li et al., 20 Mar 2026). Each 1-second segment receives a segment-level script encoded as

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$0

which is fused into a pretrained MMAudio backbone through the Script-Guided Temporal Fusion Module and Temporal Script Attention. The model also adds Bi-Frame Sound Synthesis for parallel in-frame and out-of-frame audio generation. On DirectorBench, FoleyDirector reaches Precision 0.4677, Recall 0.5962, F1 0.4819, compared with F1 0.2378 for MMAudio and 0.2451 for Hunyuan-Foley; on VGGSound-Director, it reports FD$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$1 1.17, ISC$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$2 14.84, IB 0.33, and DeSync 0.432 (Li et al., 20 Mar 2026).

"ShotDirector" is not presented as a co-director, but it is relevant because it isolates one directorial subproblem: multi-shot transition design (Wu et al., 11 Dec 2025). It combines parameter-level camera control with hierarchical editing-pattern-aware prompting and a shot-aware mask. On its evaluation set, it reaches Transition Confidence 0.8956 and Transition Type Accuracy 0.6744, outperforming the listed baselines and showing that directorial control can be injected at both geometric and semantic levels (Wu et al., 11 Dec 2025). Taken together, these systems suggest that co-direction in computational filmmaking is usually implemented through role specialization, structured intermediate representations, and iterative validation, rather than a single monolithic generative pass.

5. Co-director as a corporate-network relation

In corporate governance analysis, the term takes an entirely different meaning. "Who Sits Where?" defines a co-director relationship as a graph-derived relation between two directors who are connected through company board memberships, primarily by serving on the same company board and secondarily through longer alternating director–company paths (Sancheti et al., 24 Mar 2026). Formally, for a bipartite affiliation network

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$3

with incidence matrix $h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$4, the one-mode director projection is

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$5

with edge weight

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$6

A direct co-director pair is therefore any pair satisfying

$h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$7

Larger $h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$8 indicates a stronger co-director tie (Sancheti et al., 24 Mar 2026).

The paper builds a scraped dataset of approximately 54,216 directors, 87,187 companies, and 299,970 director–company affiliations, and uses Breadth-First Search (BFS) as the core extraction engine over company and director pages (Sancheti et al., 24 Mar 2026). It then constructs director and company projections, analyzes local third-degree neighborhoods, extracts maximal cliques, mines 5,972 maximal frequent director itemsets (MFDIs) and 4,742 maximal frequent company itemsets (MFCIs), and uses LLMs only for relationship enrichment and interpretation, not for primary structural detection (Sancheti et al., 24 Mar 2026).

The empirical picture is one of dense interlock. The study reports that 41% of directors hold directorship in exactly 1 company, 17% hold directorship in exactly 2 companies, and 58.6% hold positions in 2 or more companies. Among companies, 43% have 2 directors and 26% have 3 directors. The strongest MFDI pair co-occurs in 69 companies, and the discussion notes that at least one pair of directors sharing the same last name appeared in 30% of companies (Sancheti et al., 24 Mar 2026). The paper treats such structures as relevant to conflicts of interest, reduced board independence, concentration of control, systemic risk, and economic “oligarchy,” while also acknowledging benign mechanisms such as information transfer, governance learning, and resource access.

6. Shared technical motifs, misconceptions, and open directions

Across these disparate literatures, co-direction consistently appears when a system must coordinate multiple partially competing constraints: persona vectors inside a transformer residual stream, creative strategy and aesthetic mode across an agentic pipeline, script and cinematography in previz, segment-level sound events over time, or directorships across a many-to-many corporate network. A recurring technical response is explicit factorization. In Co-Director for video storytelling, factorization appears as $h_t^{(l)}=h_t^{(l)}+\lambda\cdot v_\alpha^{(l)},$9 and factored reward (Song et al., 27 Apr 2026). In Creative Collision, it appears as linear interpolation between distinct persona directions with measurable geometric asymmetry (Sahoo et al., 15 Jun 2026). In corporate networks, it appears as a bipartite graph with weighted one-mode projections (Sancheti et al., 24 Mar 2026).

A second motif is the separation of high-level semantic planning from lower-level realization. WorldDirector makes this explicit by decoupling semantic motion orchestration from visual generation and conditioning chunked video synthesis on $\mathcal{L}=\{20,21,\dots,38\}$0 to maintain persistent dynamic object memory (Wang et al., 2 Jul 2026). Director, in dynamic scene modeling, similarly embeds instance-consistent semantics directly into tracked Gaussians and supervises them with both identity logits and language embeddings, enabling open-vocabulary querying over a temporally coherent 4D representation (Jiang et al., 2 Apr 2026). A plausible implication is that co-director systems become more reliable when “what should happen” is separated from “how it is rendered.”

The literature also converges on several limitations. Co-direction is often asymmetric rather than democratic: Spielberg largely overpowers Scorsese in mixed activation steering (Sahoo et al., 15 Jun 2026). Collaborative film tools often remain pre-production systems rather than full production infrastructures: CineVision does not yet export the technical metadata needed for lens-accurate shot-list translation (Wei et al., 27 Jul 2025). Agentic storytelling remains compute-intensive, because each bandit iteration is a full video-generation pass (Song et al., 27 Apr 2026). Corporate co-director detection remains seed-dependent, coverage-limited, and structurally descriptive rather than causal (Sancheti et al., 24 Mar 2026). Systems that broaden control, such as FoleyDirector or WorldDirector, still depend on annotation quality, backbone quality, and the tractability of long-horizon coordination (Li et al., 20 Mar 2026, Wang et al., 2 Jul 2026).

A final misconception concerns lexical continuity. The director of liquid-crystal theory is the unit orientational field governing anisotropic continuum behavior, not a collaborative creative or organizational role. In that literature, the relevant objects are $\mathcal{L}=\{20,21,\dots,38\}$1, tactoid textures, anchoring, and electric-field response, including bipolar, quasi-bipolar, and frozen director configurations (Turiv et al., 2015, Safdari et al., 2021). That usage is terminologically adjacent but conceptually unrelated to co-directed generation, story planning, or board interlocks.

In current research usage, then, co-director is best understood as a family resemblance term for systems or relations in which more than one directing influence must be represented, coordinated, or resolved. The exact formalism depends on domain: vector superposition in activation space, bandit-guided hierarchical control in generative storytelling, multimodal role coordination in previsualization and sound design, or weighted co-membership in corporate graphs.