From Live Action to AI-Generated Video: A Comparative Study of Green- Screen Video-Guided and Text-Prompt-Based Workflows

Generative AI filmmaking research addresses character creation, style, and narrative expression [1]. Production shots must also preserve planned framing, movement, and character relationships. Spatiotemporal consistency research distinguishes frame quality from continuous video behavior [2]; a convincing still does not establish that intended action or subject presence persists throughout a clip. This study compares two storyboard-based workflows using shared character images and text, with only workflow A additionally receiving green-screen video. Building on workflow-oriented virtual-production research [3], it examines input preparation, departures from shot requirements, and implications for production review. The focus is whether generated shots preserve the intended composition, movement paths, and required subjects. The comparison concerns selected outputs from one commercial platform, rather than a new algorithm or a completed film.

Bazin connects staging, camera angles, and character arrangement to interpretation [4]; Bordwell discusses narrative space [5], and Burch distinguishes on-screen and off-screen space [6]. These concepts inform the observation categories rather than constitute a validated AI evaluation scale. They direct attention to how composition and movement convey the intended action. Three dimensions guide analysis: framing and viewpoint, including shot scale and visible viewing direction; blocking and spatial relationships, including movement path, placement, and interaction; and subject presence, including whether required characters appear during the clip. These dimensions relate generated outputs to specific storyboard requirements beyond the appearance of individual frames. Identity consistency is not scored. Planned setting and costume transformations are distinguished from departures from shot requirements.

CameraCtrl and Follow Your Pose investigate camera and pose guidance [7], [8]; ControlNet demonstrates spatial conditioning in image generation [9]. These methods contextualize visual constraints but do not establish Higgsfield’s implementation. Structure and content-guided synthesis separates source structure from target appearance [10]. FateZero and TokenFlow address video-editing consistency [11], [12], while VBench distinguishes subject consistency, spatial relationships, and temporal behavior [13]. The present study uses shot analysis rather than these algorithms or benchmark scores.

Thirty green-screen shot tasks were processed with Higgsfield’s Seedance 2.5 [14] during October 8-11, 2026. Outputs lasted 5 seconds at 1080p, 16:9, and 24 fps. Both workflows used identical storyboard text, character images, and generation settings; A added the corresponding video. Attempts varied, and one output per task and workflow was retained, yielding 30 pairs and 60 analytical videos, not the total number generated. All retained videos were assessed against applicable storyboard requirements. Four pairs were purposefully selected [15] to examine framing/viewpoint, walking path, foot-view direction, and two-character presence. Shot 1-4 are figure labels, distinct from source IDs S01—S30. Table 1 defines the comparison.

The researcher reviewed and revised a ChatGPT-assisted shot storyboard, prepared storyboard images, and filmed the planned framing and action with a Sony FX9 in the university’s green-screen classroom. Storyboard images supported planning and were distinct from the uploaded character reference images. A received the filmed clip with the shared text and character images; B received the same text and images without video. Images were linked through @-mentions. Thus, “text-prompt-based” denotes the absence of video reference, not strictly text-only input. Figure 1 and Table 2 summarize preparation, generation, and review. The storyboard defines requirements for framing, viewing direction, action, and subject presence. Green-screen footage presents these requirements over time, while character images provide visual references. Evaluation therefore focuses on storyboard compliance rather than resemblance to a single source frame. Both workflows share text and images; only A receives an additional video reference. The comparison cannot isolate the effects of information within that clip. Intended clothing and setting changes are acceptable, whereas altered movement paths or missing characters may violate shot requirements. Complete-clip review checks whether these requirements persist beyond the illustrated frame, without attributing observed differences to an unverified generation mechanism. The green-screen reference shows performed movement, whereas storyboard text describes the intended action verbally. Shared character images guide appearance but do not specify movement throughout the shot. A generated clip may match the planned composition at one moment yet depart from the required action later. Visual plausibility alone therefore does not establish that a shot meets production requirements. The analysis considers both composition and unfolding action, while treating requested changes to setting and clothing as intended transformations rather than generation errors.

Table 2 summarizes preparation and review. The storyboard requirements defined before generation provide the taskcriteria; source footage is a filmed realization, not the sole standard of success.

The written storyboard defined requirements before generation and served as the evaluation standard. Source footage provided a filmed realization. Intended changes to setting, clothing, and shoes were not coded as errors. The generating researcher reviewed complete clips and coded applicable requirements as met or unmet, without independent blinded assessment. Framing/viewpoint applied to 30 pairs, movement path to five, and required-character/body-part presence to 28. Food-only S18/S19 were excluded from the last dimension; hand, foot, and eye shots remained eligible. Four pairs received detailed analysis; screenshots illustrate these full-clip judgments.

Figure 2 presents representative source and output frames for four selected storyboard tasks. Each row places the green-screen footage beside the corresponding outputs from workflows A and B. The examples illustrate upper-body framing, walking-path presentation, foot-level viewing direction, and two-character composition. This arrangement supports comparison of visible shot scale, viewpoint, and subject placement. A screenshots were color-graded in DaVinci Resolve; color, fog visibility, and visual finish were not assessed. The frames illustrate the observations rather than independently establish temporal behavior; movement paths and whole-clip subject omissions were assessed from complete videos. Specifically, the graphic’s “subject fades out” label for Shot 2 refers to the walking-path deviation described below, not verified fading.

Shot 1 required upper-body framing and the planned viewpoint, retained by A; B instead produced a tighter facial close-up. Different screenshot moments do not establish a viewpoint error in A. Shot 2 required an extreme wide shot with straight left-to-right walking, retained by A. B approached from the distant center and turned left into the store, deviating in path rather than fading. Shot 3 required a lateral foot-level view: A retained the side view, whereas B showed the feet frontally. In Shot 4, A retained both required characters; neither appeared anywhere in the complete B clip, constituting whole-clip subject omission.The four shot comparisons illustrate how the two workflows preserve or depart from storyboard requirements for composition, movement, and character presence.

Table 3 reports compliance across applicable tasks and separates shared outcomes from pairs in which only one workflow met the requirement.

A = with video reference; B = without video reference. Movement-path tasks: S01, S02, S09, S25, S26. Food-only S18/S19 are excluded from character/body-part assessment; designated hand, foot, and eye shots remain eligible. Columns overlap only through the A/B totals: A meets = A only + both; B meets = B only + both. A showed higher descriptive compliance in all three dimensions. However, B-only successes and shared failures show that added video did not guarantee compliance. These counts describe retained outputs, not first-attempt success rates.

Video references provide a visual realization of requirements expressed in the shared inputs. The observed advantage for A concerns this retained corpus; it does not establish general superiority or reveal Higgsfield’s mechanism. Related guidance and editing studies [7], [8], [10], [11], [12] provide context rather than a tested explanation.

A requires reference filming; B uses the shared images and text directly. Shot acceptance should follow the storyboard and complete clip, checking framing, movement path, and required-subject presence. Prompt revision, reference replacement, resampling, and reshooting are possible responses, but their effectiveness and any time or cost savings were not tested.

Attempts and output selection were not standardized, and the generating researcher also selected and coded the outputs without independent confirmation. Counts therefore describe retained videos rather than average reliability or a controlled workflow effect. Only five tasks assessed movement path. Time-coded records, frame alignment, and quantitative temporal measures were not reported; graded A screenshots preclude matched-appearance comparison. A also received additional information, so differences cannot isolate the green background or a model mechanism. Replication requires prompts, assets, input-mode/version records, and generation and selection histories.

Across 30 storyboard tasks and 60 retained outputs, workflow A with added green-screen video references showed higher descriptive compliance than workflow B with shared text and character images alone. A preserved framing and viewpoint in 25 of 30 tasks, compared with 18 for B; movement paths in four of five tasks, compared with two; and required characters or body parts in 26 of 28 tasks, compared with 21. These findings suggest that video references can support planned shot design within the examined production setting, although B-only successes and shared failures show that they do not guarantee compliance. The study links storyboard requirements to workflow inputs and complete-clip review, providing a practical basis for assessing generated shots. Because generation attempts and output selection were not standardized and assessment relied on the generating researcher, the results describe the retained corpus rather than establish general workflow superiority. Further validation requires standardized procedures, independent assessment, and accessible generation records.