Reconfiguring Film Production Workflows amid Generative AI Evolution: A Longitudinal Case Study of Critterz

Generative AI has entered creative production not only as a tool for generating images and video, but also as a technology that changes how creative work is organized. Research on creative industries suggests that generative AI has the potential to accelerate creative processes and reorganize existing workflows while maintaining the importance of human judgment and creative agency in production [1], [2]. Therefore, the question of AI in filmmaking extends beyond whether a model can replace a specific production task. It also concerns how human creators reorganize their work when AI becomes integrated into the production system. Recent empirical research on AI-assisted video production describes workflows that are less linear than conventional production, characterized by repeated cycles of planning, generation, evaluation, revision, and regeneration [3]. Similarly, research on creative workers using generative AI demonstrates that previously distinct creative activities can become part of a more continuous process of interaction between humans and AI [4]. Further empirical research on generative cinema suggests that filmmakers actively negotiate where AI should enter the production process rather than employing it as a uniform production technology [12]. These studies indicate that the impact of generative AI on film production is determined not only by what the technology can generate, but also by its position within the workflow and the manner in which creators organize their work around it. Nevertheless, these discussions primarily explain human-AI collaboration under particular technological conditions. Less attention has been given to what happens when the AI tools themselves change during the production of a continuing film project. This issue becomes especially significant when the production cycle of a film project extends beyond the stable availability of the generative AI tools on which part of its workflow depends. While a project remains in production, models may be updated, replaced, restricted, or discontinued. At the same time, adapting to technological change should not automatically be understood as a cost-free solution. Transitioning between different models may require additional model testing, learning, production coordination, and the reevaluation of existing visual material. Research on generative video continues to identify spatial and temporal consistency as significant technical challenges, including inconsistencies in object appearance, motion, spatial relationships, frame-to-frame details, and visual style [13], [14]. In film production, technological adaptation therefore raises not only the question of whether another model can perform a comparable task, but also whether its outputs remain compatible with the creative and technical requirements of an ongoing project. These issues can be examined over time through the case of Critterz. Conventional animation and post-production tools were used together with DALL-E-generated visual material in the original 2023 short [5]. In 2025, the project was remastered shot-for-shot using Sora, bringing generative AI closer to direct moving-image production [6]. During the later feature-production stage, the project encountered the discontinuation of Sora and a changing model environment [7], [15]. The same continuing project therefore provides an opportunity to observe not only how generative AI entered film production, but also how the position of AI and the surrounding workflow changed as the underlying technologies evolved. This paper therefore asks: RQ: How does the evolution or discontinuation of generative AI tools reconstruct the production workflow of a continuing film project? Rather than evaluating AI solely through task replacement or production efficiency, the analysis focuses on changes across four dimensions in the three stages of Critterz: AI's functional position, human coordination and creative control, technological dependency, and adaptation risks. The objective is not to develop a statistically generalizable model for all AI filmmaking, but to identify analytical relationships that can subsequently be examined across other types of AI-assisted film projects.

Although actual film projects always include feedback and revision, traditional film production is generally structured around relatively distinct stages. Generative AI introduces additional feedback loops because generated outputs frequently require repeated evaluation, selection, modification, and regeneration. According to Kim, Lee, and Hyun [3], AI-assisted video production can develop into a nonlinear process in which planning, generation, evaluation, and revision are repeatedly connected. Similarly, Amankwah-Amoah et al. [1] argue that generative AI has the potential to reorganize creative workflows and establish new forms of human-AI collaboration. Further evidence that AI can occupy different positions within filmmaking, rather than functioning as a single production technology, is provided by recent empirical research on generative cinema. In an 11-week study involving 20 amateur filmmakers and 15 AI-infused films, Halperin, Flores Ruíz, and Rosner [12] identify four approaches through which filmmakers negotiate AI use: minimization, maximization, compartmentalization, and revitalization. Their findings demonstrate that creators actively determine how AI should enter production and how it should interact with existing creative practices. This implies that AI should not be regarded as a single production stage. As model capabilities evolve, its functional position may shift. Visual development may be the primary role of an image-generation model, whereas a video-generation model can participate more directly in moving-image production. The analytical problem is therefore not merely whether AI is used, but where AI is embedded within the production structure and how this position changes over time. Problems specific to temporal media are also introduced by the transition from image generation to video generation. Research on generative video has identified spatiotemporal consistency as an ongoing technical challenge. These problems can include inconsistencies in object appearance, motion, spatial relationships, frame-to-frame details, and visual style [13], [14]. Such limitations are particularly relevant to filmmaking because individual generated outputs must not only be visually acceptable on their own, but also remain compatible with the characters, environments, and visual language of the broader project. Existing research therefore explains both the nonlinear structure of AI-assisted production and some of the technical limitations of generative video. It does not, however, fully explain what happens when the technological system supporting an ongoing production changes while the project itself continues.

Human creative work is not necessarily eliminated by the expansion of generative AI. Bender [2] argues against understanding technological change through a simple task-replacement model and emphasizes the continuing significance of human creative agency in media production. Retkowsky, Hafermalz, and Huysman [4] further demonstrate that working with generative AI changes the organization of creative activity, as creators continuously respond to generated outputs, make judgments, and adjust the creative process. Human involvement can therefore operate at multiple levels in filmmaking. At the content level, creators evaluate, select, modify, and integrate individual outputs. At the workflow level, they determine where AI should enter production, which systems are appropriate for particular tasks, and how generated material can be incorporated into a coherent project. Similarly, Halperin et al. [12] show that filmmakers may minimize, maximize, compartmentalize, or revitalize AI use across different creative activities. This second level becomes particularly important when the technological environment changes during a project. Replacing a generative model may require creators to reevaluate the capabilities and limitations of alternative systems, repeat technical or creative tests, and determine whether new outputs remain compatible with material that has already been produced. Human-AI collaboration should therefore be examined not only through content generation, but also through the human coordination required to maintain production continuity under technological change.

The concept of dynamic capabilities provides another perspective through which changing AI technologies can be considered. Simón, Revilla, and Sáenz [9] show that effective AI integration depends on organizational capabilities that support interoperability, trust, mutual knowledge, and the development of adaptive routines. Although their research is not situated within film production, the dynamic-capabilities perspective provides a useful analytical lens for understanding a production environment in which the underlying AI technology may change during an ongoing project. Building on this perspective, this paper uses adaptive workflow to describe a production structure capable of evaluating alternative models, comparing their outputs with project requirements, reallocating production tasks, and adjusting the pipeline in response to changes in technological conditions. Adaptive workflow is not presented here as a universal theory of AI filmmaking. Instead, it is employed as an analytical concept through which the changes observed in Critterz can be interpreted. Adaptability, however, should not automatically be understood as a cost-free solution to technological dependency. While the capacity to transition between models may reduce dependence on a single technology, it may simultaneously increase the production team's requirements for testing, learning, coordination, migration, and consistency management. This creates a research gap between studies concerned with the adoption of AI and the longer-term problem of maintaining a production workflow under continuing technological change.

This study employs a longitudinal single-case study of Critterz. Case-study research is particularly useful for investigating emerging phenomena in which theoretical relationships and processes are still developing [10]. The case is suitable because Critterz provides multiple observable stages during which the role of generative AI changed while the creative project itself continued to develop. The 2023 short represents an early hybrid model that combined AI-generated visual material with conventional animation and post-production [5]. The 2025 Sora remaster represents a deeper integration of generative video into moving-image production [6]. The later feature-production stage provides evidence of technological disruption and subsequent workflow reconfiguration after the discontinuation of Sora [7], [15]. Rather than comparing unrelated films using different AI systems, these stages enable the study to compare changing technological configurations within one continuing creative project. The use of a single case also imposes an important limitation. Critterz cannot represent all AI-assisted film production because different conditions may apply to live-action films, feature animation, independent productions, or projects using different generative systems. Nevertheless, the analytical value of case-study research does not depend exclusively on statistical generalization; carefully examined cases can contribute to the development and refinement of analytical propositions [10], [11]. The objective of this study is therefore not to claim that the Critterz workflow represents all AI filmmaking, but to identify relationships that can subsequently be examined through cross-case research.

This study draws on publicly available materials documenting the development of Critterz. These include creator interviews, production reports, official technology announcements, and industry reporting concerning the AI technologies used at different stages of the project [5]–[8], [15]. A structured qualitative comparison was conducted rather than a quantitative performance evaluation. The analysis was completed in two stages. First, the available case materials were organized chronologically according to significant changes in the generative AI technologies used by the project. This produced three analytical stages:1.DALL-E-assisted visual generation;2.Sora-based moving-image generation;3.post-Sora workflow reconfiguration.Second, each stage was compared across four analytical dimensions:Functional position of AI — the part of the production process that the AI system performs or supports;Human coordination and creative control — how creators select, evaluate, modify, and organize AI-generated content;Technological dependency — the extent to which the workflow becomes connected to the availability and capabilities of a particular AI technology;Adaptation risks — the additional testing, migration, coordination, or creative-consistency problems that may emerge as the technological environment changes.Because internal production metrics are not publicly available, this study does not conduct a quantitative performance comparison of DALL-E, Sora, or alternative models. Instead, the results concern observable changes in production organization across the three stages.

The original Critterz short, released in 2023, used OpenAI's DALL-E to generate visual material for its characters and environments. Production documentation indicates that the team repeatedly generated and refined visual material before incorporating selected results into a more conventional animation workflow [5]. The most significant aspect of this stage is not that AI replaced animation. Instead, AI was primarily integrated into the visual-development portion of the production pipeline. The process can be summarized as:Prompting → Generation → Human Selection → Refinement → Integration Human creators retained responsibility for determining which generated results were appropriate for the project and for incorporating those materials into subsequent production. The production account published by Animation World Network documents the continued use of conventional production technologies following AI-assisted visual development [5]. Dependence on DALL-E was therefore concentrated primarily in visual-material generation rather than across the complete moving-image production pipeline. The first phase can consequently be interpreted as a hybrid workflow with partial AI integration. AI altered the production process for part of the visual material, but it did not become the primary mechanism through which the complete moving image was generated.

In February 2025, Native Foreign released Critterz: Remastered, a shot-for-shot remake of the original short using OpenAI's Sora [6]. This represents a qualitative shift in the functional position of generative AI. Unlike DALL-E, which was primarily used for visual-material generation, Sora allowed generative AI to participate directly in moving-image production. The workflow therefore shifted from AI-assisted visual-asset generation toward direct generative-video production. Rather than exclusively providing static visual content that required subsequent animation, generation could now operate closer to the level of moving shots. This development did not imply that the project became independent of human creative control. The remaster retained the same script and performances while changing the technological tools through which the visual result was produced [6]. Sora's greater technical capability is therefore not the only reason this stage is significant. The more important transformation was structural: generative AI moved closer to the center of moving-image production. As the model's functional position changed, its capabilities and limitations became more closely connected to the production path. This stage therefore represents a transition from partial AI integration toward a more deeply integrated and iterative generative workflow.

The third phase introduces a different problem: what happens when a generative AI technology that has become important to production changes or becomes unavailable while the project itself continues? OpenAI discontinued the Sora web and app experiences on April 26, 2026, with the API scheduled for discontinuation on September 24, 2026 [7]. By this stage, Critterz had developed beyond the original short and the Sora remaster into a feature-film production. Reporting based on interviews with the filmmakers indicates that Sora had initially formed an important part of the feature's technological environment and that its discontinuation required the team to demonstrate that the surrounding production system could continue across a changing model landscape [15]. This disruption exposes a form of technological dependency. When AI is concentrated in a limited production activity, technological change affects a relatively contained portion of the workflow. As generative video moves closer to the center of moving-image production, however, changes in the underlying model can also affect the surrounding production arrangements. The feature-production stage of Critterz was also associated with Woven, an AI-assisted production system designed to provide a shared environment for collaboration and asset management [8]. Subsequent reporting describes Critterz as the first full production run of Woven and indicates that the system later incorporated other models following Sora's discontinuation [15]. Analytically, the significance of Woven is therefore not that it demonstrates the elimination of model dependency. Rather, it indicates a shift toward strengthening the production layer surrounding individual generative models. The observable workflow transition can therefore be summarized as: Existing Production Configuration → Tool Disruption → Evaluation of Alternatives → Model Testing → Workflow Adjustment → Continued Production This phase provides evidence of technological disruption, model reevaluation, and workflow adjustment [15]. However, the available sources do not provide internal quantitative data on the precise financial, learning, temporal, or labor costs associated with that adjustment. These potential costs are therefore treated as analytical risks rather than measured case outcomes.

Based on the documented production changes in [5]–[8] and [15], the longitudinal comparison identifies a distinct shift in the functional position of generative AI across the three phases.Three primary results emerge from this comparison. First, the functional position of AI shifts from visual-material generation toward direct moving-image generation. The technological role of AI therefore changes qualitatively rather than merely increasing in the number of tasks it performs. Second, deeper integration connects the production workflow more closely to the capabilities and availability of the underlying generative technology. In the DALL-E stage, AI remained concentrated in visual development. In the Sora stage, the generative model participated more directly in moving-image production. Third, technological disruption creates a need for workflow adjustment rather than the straightforward continuation of the existing production structure. The post-Sora phase demonstrates that when the technological environment changes, production may require renewed model evaluation, testing, and pipeline reconfiguration. The case therefore does not demonstrate a simple linear progression in which increasingly capable AI continuously improves production efficiency. Instead, it suggests that deeper technological integration can simultaneously create new forms of production dependency.

The function of generative AI changes across the three phases rather than simply increasing in quantity. In the DALL-E phase, AI functions primarily as a visual-generation tool. In the Sora phase, AI moves closer to direct moving-image production. In the later phase, the problem expands beyond content generation to include the organization of production under changing technological conditions. This is consistent with previous research suggesting that generative AI transforms creative workflows rather than merely automating isolated tasks [1], [3], [4], [12]. The Critterz case further suggests that AI should not be regarded as a fixed technological variable. Within the same continuing creative project, its functional position can shift as model capabilities and production requirements evolve. The unit of analysis should therefore extend beyond the question of what an AI model can generate and include the relationship between AI capability and production organization.

The second finding concerns technological dependency. Early AI integration can influence a limited production activity without making the wider moving-image workflow directly dependent on the model. Deeper integration changes this relationship. During the first Critterz phase, DALL-E primarily influenced visual-material generation, while conventional tools remained significant in subsequent moving-image production [5]. During the Sora phase, generative AI moved directly into moving-image generation [6]. The subsequent discontinuation of Sora illustrates how the continued availability of a model can become a production issue once the technology occupies a more central functional position [7], [15]. This suggests that production teams working with generative AI may need to evaluate not only how effectively a model performs under current conditions, but also how closely the wider workflow becomes tied to its continued availability. Technological dependency should therefore be considered alongside efficiency and generation capability when evaluating the long-term use of generative AI in film production.

The third analytical implication is that long-term AI film production may require adaptive capability rather than permanent commitment to a single model. Simón et al. [9] describe AI integration as a process involving organizational capabilities, interoperability, learning, and reconfiguration. Although this perspective was not developed specifically for filmmaking, it helps explain why technological change may require modifications to the routines surrounding AI use. Applied to film production, an adaptive workflow can be represented as: Project Requirements → Model Evaluation/Testing → Model Selection → Human Coordination → Production → Continuous Re-evaluation The purpose of this concept is not to argue that every AI film should employ multiple models. Rather, it describes a production structure that is less dependent on the permanent availability of a particular technological system. The relevance of this capability is visible in the Critterz case because the underlying technological environment changed while the larger production continued [15]. However, adaptability should not automatically be understood as efficient or cost-free.

The fourth analytical implication concerns a potential trade-off between technological dependency and adaptation cost. A production pipeline constructed around a single generative model may reduce the immediate need to evaluate multiple systems, but it can become vulnerable when that model changes or becomes unavailable. The Critterz case makes this side of the problem visible: Sora's discontinuation occurred after the model had already become important to the project's technological environment [7], [15]. A shift toward a more adaptive production structure may reduce dependence on a single model; however, adaptation may also introduce additional production requirements. Research on AI integration suggests that effective human-AI collaboration depends partly on organizational learning and the reconfiguration of routines [9]. Similarly, research on generative cinema indicates that filmmakers actively negotiate the position of AI across different production activities [12]. Changing the technological configuration of a film project may therefore require additional testing, learning, and coordination. A further issue concerns creative consistency. Research on generative video continues to identify spatial and temporal consistency as significant technical challenges. These include inconsistencies in object appearance, motion, spatial relationships, frame-to-frame details, and visual style [13], [14]. When the generative technology used by an ongoing production changes, these limitations create an additional compatibility question: the production team may need to evaluate whether outputs from the new technological configuration remain compatible with material that has already been produced. These observations lead this study to propose a dependency-adaptation trade-off. Dependence on a single model may reduce short-term adaptation work while increasing vulnerability to technological disruption. A more adaptive workflow may reduce this dependency, but may also require additional testing, learning, coordination, migration, and consistency management. These costs are not directly quantified in the Critterz case and should therefore be understood as analytical implications rather than empirically measured findings. Future production-based research is required to determine how significant these costs become in practice.

This paper employed the longitudinal case of Critterz to investigate how the evolution and disruption of generative AI technologies can reconstruct film-production workflows. The comparison of the DALL-E, Sora, and post-Sora phases demonstrates a movement from AI-assisted visual development toward direct participation in moving-image production and, subsequently, toward workflow adjustment under technological disruption. The results suggest that the impact of generative AI on filmmaking cannot be understood solely in terms of task replacement or production efficiency. As AI moves closer to the center of production, the workflow may become increasingly connected to the capabilities and continued availability of particular technological systems. The Critterz case therefore highlights technological dependency as an emerging production concern. At the same time, reducing technological dependency through a more adaptive production structure does not necessarily eliminate production risk. Instead, it may change the form of that risk. Model testing, learning, workflow migration, coordination, and creative-consistency management may become additional production requirements. Based on this analysis, this paper proposes a dependency-adaptation trade-off: long-term AI film production may need to balance the vulnerability created by dependence on a particular model against the costs required to maintain technological adaptability. This study is subject to several limitations. First, Critterz is a single case and cannot represent the full range of AI-assisted filmmaking. The results should therefore be interpreted as analytical propositions rather than statistically generalizable conclusions. Live-action film, feature animation, independent production, and other creative projects may operate under different technical and organizational conditions. Second, the study relies primarily on publicly available production materials, creator interviews, official announcements, and industry reporting rather than internal production records. This makes it possible to identify observable changes in production organization, but does not allow the quantitative measurement of the precise financial, temporal, or labor costs associated with model migration and workflow reconfiguration. Future research should therefore combine creator interviews, internal production records, and workflow data to compare multiple long-term AI film projects. Such research could determine whether the dependency-adaptation trade-off proposed through the analysis of Critterz also appears in other forms of AI-assisted filmmaking. It could also quantitatively examine the impact of technological adaptation on production time, labor requirements, creative consistency, and production cost.