An old red pickup crosses Montana. A woman and her dog share its cramped cab. Several minutes later, that woman is wearing acid-citron vinyl, strapped to a missile, rising toward a spacecraft. The next song begins where the first film leaves her.
That connection is the spine of Clarapetra’s new two-part music-video story. “Train Leavin’” supplies the departure, ordinary companionship and impossible ascent. “Andromeda” takes Citron Clarapetra aboard a failing ship, gives her a small robot friend, and turns that friendship into the danger she has to escape.
The production developed through treatments, reference gathering, individual storyboard frames, motion generation, editing, critique, replacement shots and delivery checks. The archive shows how often a convincing image needed another decision before it could become a convincing scene. The tools made ambitious imagery available; human direction supplied the reasons for the images to follow one another.
Montana → desert → transformation → missile ascent
Latest connected cut: 3:49.2 · 1080p · 23.976 fps
Arrival → companionship → takeover → escape
R5: 3:27.52 · 1080p · 25 fps · 76 shots
Start with the songs. Give the world rules.
The first work was choosing what the story needed from its soundtrack. Early development compared possible songs for the Citron-and-Montana concept. “Train Leavin’” became the working soundtrack for the road film; “Andromeda” later became a full space film in its own right. The supplied songs and lyrics were preserved as source material, with replacement audio sources and mastering revisions stored separately.
Before generating a sequence, the production needed recognizable anchors. For the road story, those were Clara’s everyday face, Lucy’s actual markings and posture, and a newly designed battered red pickup. Driver and passenger positions mattered. So did the regular cab, long bed, brown bench, primer patch and dent. An attractive truck that changed body shape between shots would break the journey.
Citron had a second, equally specific identity: the glossy yellow-green costume, silver details, platform boots, sculpted hair, angular eye treatment, black nose and dark lips. Those references carried across the two films. The visual brief could change from rain-streaked glass to a white spacecraft corridor while the lead remained recognizable.
Train Leavin’: build the ordinary before the impossible.
The early red-pickup treatment divided the journey into eight sequences and 32 shots. Butte led to the southbound Divide Rest Area. Small actions followed: walking Lucy, offering water, sleeping on the bench. Night portraits, dust and a parked-truck storm gave the road its texture. The desert moon dance came before the fictional silo threshold and the Citron transformation.
The geography references helped ground the opening, while the weather and fantastic launch belonged to the film’s invented world. The unglamorous details were doing dramatic work. Without the dog, the water routine and the narrow bench, the later ascent would lose the companionship that gave it meaning.
Each shot received an individual widescreen image and a matching motion brief. The complete package also included eight printable sequence boards, a shot list and an interactive review. Its proposed 212.4-second allocation belonged to that early working soundtrack. It was a planning scaffold; it should not be read as the timing of the later 3:49 Part 1 cut.
The storyboard was the beginning of the edit.
The image-to-video handoff favored one action and one modest camera movement in a take. A wind portrait, a dog turning at the window, a hand reaching toward a pod: each could be selected and trimmed later. Transitions belonged in the edit. Asking one generation to perform several scenes would surrender the precise moment when one image should become another.
Motion introduced problems that a still could hide. The truck’s direction could reverse, a costume boundary could drift, a gesture could stop too early, or a character could freeze after the useful action. The working response was to review the take, choose its usable range, and regenerate only where an edit could not repair the problem.
The road film also accumulated story revisions. One review found that its crash and surreal journey did not connect clearly enough. Existing footage was recut around repeated train-and-cab imagery, awake rest-area action, and a clearer transformation, embrace and missile-decision chain. Performance and repetitive tumble coverage were tightened while the purple guitar, robot band, doubled moons and Citron identity remained part of the film’s vocabulary.
Andromeda: an atmosphere needed a consequence.
“Andromeda” began with “The Unanswered Room,” a 16-setup treatment for Citron in a pristine interstellar communications station. Two chairs and one occupant expressed distance. Controlled symmetry, tactile switches, violet glass and ceremonial gestures made the space beautiful, severe and faintly absurd.
The first film established that visual language, but the subsequent direction asked more of it. The production researched and reinterpreted imagery from Stanley Kubrick’s 2001: A Space Odyssey: monumental ship exteriors, the circular exercise space, HAL’s eye, closed pod-bay doors, memory-module removal and the white room. The “Daisy” shutdown reference informed a visual passage. The full Andromeda soundtrack stayed in place, with no added movie dialogue or “Daisy Bell” performance.
Those recognizable motifs shaped the second version. The next story pass changed its emotional engine. Citron would arrive on the missile, be released with help from a friendly Microduck, and face a ship sabotaged by HAL. She would try to stop the computer, only to discover that the threat had moved into her companion. The earlier cosmic-bubble ending gave way to an escape with a cost: she leaves the corrupted robot behind.
Give the companion a design before giving it a betrayal.
The robot was grounded in the actual Pollen Robotics Microduck, which its maker describes as a 25-centimeter biped. The production used official product photographs to establish its compact scale, cream shell, orange details, exposed mechanism and camera arrangement. Pollen Robotics’ Microduck reference.
A dedicated character sheet covered design, front, profile, friendly behavior, scale and a fictional HAL-controlled state. One small correction was revealing: the side sensor had acquired an extra red glow. It was returned to black so a single red point inside the main camera could carry the takeover. The infection and attack are inventions for this story, not demonstrations of the real robot’s behavior.
Friendship had to come first. A small head tilt toward an open palm was enough to make the machine curious and approachable without inventing a smile. Later, a centered optical match cut could turn that same camera into a threat. HAL’s light dies; the companion’s camera becomes red. The hardware remains recognizable while its apparent intention changes.
Continuity is also a character’s state of mind.
Some faults were obvious: a wrong outfit, altered mouth geometry, a missing painted nose. Others were editorial. Clara could run urgently through a collapsing corridor and then enter the pod as though the emergency had ended. Every individual shot might look polished while their combination contradicted the story.
The revisions therefore checked behavior as closely as appearance. Once the escape began, slow walks, seated resets and recovery pauses were removed from the selected ranges. A ship jolt needed a bodily response. A blocked route needed an action and a consequence. The robot’s red eye had to stay active during pursuit. The pod sequence had to show approach, lunge, the hatch-close action, closure and launch in that order.
A faster cut had also lost the bite while removing takes with changing robot geometry. The next revision restored the complete action: the mechanical lower rim catches a fold of Clara’s boot, the material pulls taut, she yanks free, and the sprint continues. The R5 edit places that attack and release at roughly 2:10–2:13.4. The cause-and-effect chain matters more than the spectacle of a single attack frame.
The final R5 delivery records 76 shots drawn from 36 sources across 3:27.52. That count describes an actual edit, not 76 separate generations. Review covered identity, source ranges, critical actions and rendered joins. Some takes were rejected, some were regenerated, and some supplied only a tightly bounded crop or moment.
Make the ending of one film the opening of the next.
The decisive connection came after Andromeda’s escape story had taken shape: “Train Leavin’” should become Part 1 and end on the launch, with Clarapetra still strapped to the missile. “Andromeda” would be Part 2. The earlier return to the truck-crash aftermath no longer served that handoff.
The revised finale used two newly generated ascent views and a selected, previously unused part of the missile-to-ship approach footage. The take review caught an altitude jump. Intercutting the cloud and space portions kept the ascent moving forward instead of sending Citron back down through a layer she had already cleared.
The change was bounded. The preservation receipt confirms that the first 4,972 frames—about 3:27.4—match the source cut, and the full soundtrack is retained. Only the final 523 picture frames were replaced. The complete Part 1 runs about 3:49.2 at 23.976 frames per second. Andromeda’s 25-frame-per-second opening continues the next portion of the approach source, so the bridge advances the action rather than replaying the same five seconds.
Finish the films. Preserve the work that made them.
The production used several kinds of tools at different stages. Built-in OpenAI image generation produced early storyboards; those records do not expose an exact model identifier. Later Higgsfield production manifests record GPT Image 2.5 for stills and Kling 3.0 for motion, including sound-off generation. Higgsedit and OpenShot supplied editable assemblies, with scripts and FFmpeg handling bounded conformation, muxing and technical checks.
The song remained the timebase. In the latest Train Leavin’ finale, verification checked the retained video prefix, original AAC packets and decoded audio. For Andromeda, the archival export’s 24-bit PCM matches the selected master sample for sample. The directly master-derived web copy uses AAC encoding, so its compressed bytes are a different representation of that audio.
Treatments, prompts, stills, original clips, revised exports, editable projects and receipts live in the Clarapetra CMS. SHA-256 checks identify which file a review refers to; native project packages preserve the possibility of another edit. A visually reviewed cut, a technically verified archive and an artist’s approval remain different facts. The latest delivery checks do not establish that every frame is flawless.
The next audience-facing form is a serial: 15-second vertical chapters spanning the two songs before the full-film rollout. A first-episode launch record already exists, but that is a separate distribution step. This feature is published following the artist’s review and approval; the films retain their separate release states.
The practical lesson from these two videos is that human involvement is essential for the time being. The human had to notice the missing bite, the impossible change from panic to calm, and the ending that closed a story which needed to continue. Generation supplied material. Direction, selection and revision made the material carry a story.
Inside the production system
The original visual article is retained above. These added chapters trace the songs, specialist reviewers, connected tools and engineering behind the two films.
Before the camera: the artist’s lyrics and Suno.
The production begins earlier than the first storyboard. Clarapetra’s words, revisions and chosen musical performances determine what the films have to hold. The CMS keeps the supplied lyric texts as originals, including unusual spelling, punctuation, stanza breaks, stretched vowels and bracketed performance directions. That makes the writing process visible instead of reducing the song to a finished audio file.
“Train Leavin’” shows that process across three supplied versions. The first compact self-portrait moves through drifter, quitter, lost records and grifter. The second shifts toward lover, fighter, drifter and sinner. The expanded third version turns the passage into an eight-line accumulation. The sequence moves between desire, refusal, reinvention and damage while its repeated “Been” gives the singer a physical rhythmic anchor.
The expanded R3 passage is preserved exactly:
Been a drifter
Been a dreamer
Been a lover
Been a fighter
Been a stranger
Been a liar
Been a sinner
Been on fire
The extended-vowel spellings of “train” carry a delivery instruction inside the lyric itself. R2 introduces a raw emotional blues break; R3 explicitly adds a crescendo and guitar solo. “Turn off my / beacon” in an earlier version becomes “Leavin’ the / Station” in the expanded draft. These are choices about sound, shape and emotional movement. A separately stored clean display version serves readers without replacing the supplied performance text.
“Andromeda” works through a different pressure: phonetic speech, clipped internal echoes, names, commands and abrupt surreal objects. “Wish I wuz witch ya” gives the mouth a different movement from standardized prose. The strut, duck, truck and shifting body images create a musical vocabulary before the space film gives them a new visual context. Its closing “Instead I’ll miss ya” changes the promise of contact into absence. This is a reading of the supplied words, not a claim that every screen event was already dictated by a lyric.
Suno supplied the musical generation layer. The actual Train Leavin’ R3 WAV and corrected Andromeda FINAL WAV both carry “made with suno” metadata. The artist supplied and revised the lyrics, brought in the source audio, and selected or corrected versions; Codex preserved the record and supported later production. The archive establishes those inputs and replacements, rather than a complete history of every Suno prompt, discarded take or session setting.
That distinction mattered when Andromeda’s source changed from the first file to an ALT and then a corrected FINAL. Earlier files and masters stayed intact; the selected source received a separate album-oriented master. Its 48 kHz, 24-bit delivery became the film’s 207.52-second timebase. Writing, generated musical performance, source selection, mastering and picture editing were successive decisions, with a human choosing what each stage should carry forward.
OpenAI and Codex: turning a brief into coordinated work.
OpenAI appears in two distinct parts of this production. Codex served as the coordinating agent environment: it searched the library, developed treatments, delegated specialist work, wrote production manifests and assembly scripts, and compared review evidence. OpenAI image generation supplied visual material, first through the built-in image tool for early boards and later through the GPT Image 2.5 model identifier recorded in Higgsfield’s manifests. The early image receipts do not expose an exact underlying model version.
The coordinating agent translated the artist’s requests into jobs with explicit constraints. Preserve the supplied song. Keep the costume and identity references. Give each clip a specific action. Retain originals. Change only the requested picture interval. Those instructions became shot briefs, source indexes, edit decisions and checks that other parts of the system could read.
Agent integration meant connecting decisions to actual tools and files. A story agent could describe the missing action; an editing agent could inspect an unused source handle; the coordinating agent could turn that finding into a bounded native timeline change. Local Python and JavaScript scripts handled frame extraction, timing calculations, manifests and verification. The same work session could call the CMS and Higgsfield tools, then inspect the resulting media rather than leave the recommendation in chat.
The roles were focused but changed with the task. Early development used music supervision, creative direction and production design. Later passes added director/dramaturgy, editing/performance, identity/research, cinematography and VFX perspectives. These were AI agents applying separate craft lenses. The system did not replace the artist’s authority: the human supplied the words, corrected the audio source, changed the plot, demanded the missing bite, and decided that one film should hand its ending to the other.
The useful unit of progress was a concrete artifact: a treatment, a source-span recommendation, an accepted first frame, an edit-decision list, a rendered movie or a verified receipt. A persuasive response from an agent could still be incomplete. The work had to survive the transition from instruction to file, and from file to what could actually be seen and heard.
The reviewers had jobs—and their notes changed the cuts.
The review team was more than a final thumbs-up. Train Leavin’ received independent story, editorial and visual craft scrutiny. The consolidated review of the artist’s EDITV2 cut records a feedback round in which three specialists agreed to prioritize the ending’s cause and chronology, clearer evidence of Lucy’s anomaly, and less repeated discovery and farewell coverage. Later R3 and R4 reports examined the director’s intent, cinematography, production design, VFX and exported edit.
A specific cinematography finding illustrates the handoff. The existing missile-decision source had an unused beginning: Clara turned from the dog toward the missile before the range selected in the timeline. Slipping the same-duration source window earlier could restore the visible decision without generating a new image or changing the soundtrack. Another note found no awake handle in the sleep source, so it proposed continuing established walking and water actions instead.
Andromeda’s first team review used three lenses: director/dramaturgy; music-video editing and performance; and music marketing/audience psychology with Citron identity. The director and editor wanted a consequential late decision. The identity reviewer emphasized a characteristic performance gesture. Their synthesis was that the film needed observable emotional change. The later R3 consensus brought the producer, story, edit pipeline and identity/research roles into agreement around rescue, friendship, sabotage, infection and escape.
Later gates became more physical. Does the duck actually catch the boot? Does the fabric pull before she releases it? Does an urgent corridor run remain urgent at the pod? Does the camera light transfer in the final assembled sequence? A completed generation or correct costume could not answer all of those questions. The returned motion and exported cut had to supply the evidence.
The reports also state their limits. The R4 Train Leavin’ editor review inspected 238 decoded frames across seven changed sequences and scanned adjacent proxy-frame hashes; it did not perform a fresh continuous listening pass. Andromeda’s R4 director and identity reviewers used cut/midpoint and dense critical samples. Those are meaningful inspections, while artist approval and full audiovisual review remain their own stages. The reviewers’ judgments are not a real-world awards jury or a measured audience test.
Higgsfield: from references to jobs to usable motion.
Higgsfield was the connected production environment for later image and motion generation, with Higgsedit providing native picture assembly. Its generation layer let the agent carry a written shot brief and selected reference images into a specific model request, then retrieve job status, returned media and review galleries. The job ledger preserved the actual IDs and outcomes alongside the source briefs.
Andromeda R5’s manifest makes the technical choices explicit: gpt_image_2_5 stills at high quality, 2k and 16:9; kling3_0 motion in pro mode, eight-second duration and sound off. Those are the settings for that recorded pass. Earlier passes and the Train Leavin’ ascent used different clip durations. The generated clips supplied picture while the chosen song remained a separate source.
Reference roles were as important as prompt wording. Wardrobe and face references identified Citron. Separate files grounded the actual Microduck, spacecraft architecture, missile, HAL eye and escape pod. Each scene defined its first physical state, one principal action and the required consequence. The friendly robot could tilt toward a hand; the hostile one had to make readable contact with a boot. A prompt was a production specification, but the specification was tested against the resulting take.
The process had two gates: first-frame review and motion review. A still could pass face, costume and hardware checks yet produce a clip with a malformed lower rim, a missing black nose, an incomplete tug or a calm recovery step. Replacement jobs and selected ranges were therefore recorded separately. Rejected outputs stayed in the archive so a later reviewer could understand why a replacement existed.
The final ledger even records a provider-rejected keyframe that was not retried or moved to another provider. The coverage was redesigned around an empty structural-collapse view and an escape point of view. This was a production choice within the available material and boundaries. Generation created possibilities; the edit chose which ones could perform the required action.
For delivery, a selected clip needed a source index, checksum, normalized picture file and usable in/out range. That is how a gallery of striking images became an edit that could be reconstructed. The shot count describes cuts in the movie, not the total number of generations or prompts submitted.
The CMS connected the agents to a durable production archive.
The Clarapetra CMS was the repository of record throughout this work. It held the original lyrics and audio, identity references, treatments, prompts, generated stills and clips, revision exports, native projects and delivery evidence. The website was a publishing consumer of those materials. The production did not have to rely on a temporary download, a provider URL or a conversation remembering which “final” was intended.
The integration used MCP, the Model Context Protocol, to expose defined capabilities to the agent. The actual local tool surface included cms_search and cms_get for finding and reading source records, cms_save_text for preserving exact writing, cms_ingest_file for importing files, and cms_verify for checking stored bytes. OpenAI’s MCP server documentation describes the general mechanism; these production records and the local CMS implementation establish the project’s concrete use.
The storage contract required an era, content kind and provenance. For these films that meant Pop Stars, the actual asset type, the tool or creator, the prompt or brief when known, source references, and a revision. Each ingestion returned an item ID, managed path, SHA-256 and receipt. A new version was another stored item; the earlier lyric, rejected clip or superseded movie stayed intact.
This gave agent handoffs stable references. “Use the corrected Andromeda master” could point to a specific CMS item instead of whichever similarly named WAV happened to be newest in Downloads. A reviewer could pin the film’s checksum; a finishing script could identify the selected master; a recovery package could map the media needed by the native timeline. Provenance carried across the transformation from generated source to normalized clip to selected range to export.
The CMS also kept the states separate. Stored meant that the file existed in the repository. Technical verification meant the specified integrity or media checks passed. A sampled visual review recorded what the reviewer had inspected. Artist approval, publication and live playback required their own evidence. The ingestion tool itself did not confer those later states, and the local CMS did not publish or deploy the films.
The same rule applies to this article. Its historical boards, selected film frames and newly authored technical diagrams are stored with source lineage. The review HTML and PDF are separately preserved before private delivery. The pipeline can explain exactly which assets a reader is seeing while leaving the publication decision with the human.
Native edits, frame arithmetic and an unchanged song.
The edit was represented as an EDL—an edit-decision list—with source indexes, source in-points and durations placed on a frame grid. Native Higgsedit assembly scripts created the project and picture render, while the road film also retained its OpenShot project. Recoverable packages included project data, referenced media, the EDL, source manifests, authoring scripts and review evidence. The movie was one output of that structure.
Source normalization solved practical differences before assembly. The initial Andromeda plan describes 1912×1080, 24 fps source clips conformed to 1920×1080 at 25 fps, with cover scaling and a crop rather than distorted stretching. Timestamp reset, removal of unwanted source audio, frame counts and checksums made the inputs predictable. Cadence conversion could repeat or drop frames, so a successful format check still had to be followed by motion inspection.
The scripts checked contiguous timing, source bounds and the provenance of derivative spans. Reusing one generated file for several cuts could work when the selected source intervals were distinct; repeating a moment to fill an unavailable action would contradict the intended performance. Final Andromeda delivery records describe 5,188 picture frames across 207.52 seconds. Frame counts, full decoding and native-package integrity checks made that measurable.
Audio finishing used a separate path. The picture render was paired with the selected master at the intended start position. The archival Andromeda MOV retained 24-bit PCM and was compared sample for sample with the master; the web MP4 used AAC at 320 kbps encoded directly from the selected WAV. The checks recorded no gain, fades, trimming, retiming or additional audio in that final mux. A compressed review stream and a lossless archive have different bytes even when they represent the same chosen program.
Train Leavin’ required a different engineering decision. Its latest Part 1 instruction protected everything before frame 4,972 and the existing complete soundtrack. The new Higgsedit tail was conformed and assembled against that boundary; the first 4,972 picture frames and original AAC packets were checked against the source. Only the final 523 picture frames changed. This let the new connection to Andromeda become concrete without reopening the entire cut.
The strongest part of the system is the meeting of evidence and judgment. A checksum can pin the file, a script can prove an interval or audio comparison, and a reviewer can identify a missing physical action. The artist can still say the scene feels wrong—or choose a different ending. Human involvement is essential for the time being because those judgments determine what the technical machinery is being asked to preserve.
The original eight sequence boards
All 32 original storyboard frames, arranged across the eight boards created during pre-production. Their timing and companion-pod ending belong to that early version. Click any board to inspect it; these are preserved development documents, not reconstructions made for this article.