Poqr Mher: how an AI animated feature reached a film programme

A trailer is public and a festival lists runtime and sessions; the full film and sales results are absent from these materials.

The project manager links the Poqr Mher trailer. The Armenian Film Festival Australia's separate programme names director Hovhannes Margaryan, an 80-minute runtime and Armenian with English subtitles; it lists Sydney sessions on September 20 and 21 and later Perth and Melbourne dates. Readers can inspect the trailer and programme, but these sources do not offer the whole feature or verify each production shot.

A ticketed programme shows an exhibition route. It does not establish actual attendance at each session, ticket numbers, box-office income or profit. The manager also describes sponsors entering story scenes without publishing contracts, receipts or revenue splits. Distribution and commercial-partner paths are visible; money received remains unknown.

• Watch the project's public trailer ↗
• Inspect the festival film and sessions page ↗

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Poqr Mher project trailer · Armenian Film Festival Australia: Poqr Mher programme

The team reports 196 scenes and 23 members, with characters, locations, risks and dependencies scoped before intensive generation.

Inspired by the Armenian epic Sasna Tsrer, the film imagines the legendary Mher encountering students in modern life. The manager reports 196 scenes, a 23-person core team and 13 AI creators. When she joined, character and location generation had begun, but shot volume, difficulty and dependencies had not been mapped. She grouped characters by importance, ranked locations and planned storyboarding and production from that map.

She tiered locations P1, P2 and P3 by how many shots depended on them and how difficult they were to make. Storyboards, characters and locations had lock dates so some work could run in parallel. The point was to know which later shots a delay would affect, not to send an entire feature to models at once. The article does not publish adaptation agreements, the full script or all boards; it cannot establish the rights position of this particular adaptation.

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Armenian Film Festival Australia: Poqr Mher programme

Character specifications, location master shots, expression keyframes and boards anchored each shot instead of a single generic prompt.

The manager says characters had JSON descriptions of skin, hair, clothes, height and facial details; locations had interior, exterior, nature and building specifications; shared lighting used master prompts. The team then made location master shots, principal-character expression keyframes and boards. She describes chaining these materials in Freepik nodes for individual shots without publishing every JSON file, source asset or final shot input.

The team assigned creators by location: one person handled its angles, lighting variations and related scenes, learning where geometry or characters drifted. The manager believes this improved continuity compared with assigning work by film sequence. Her claim that consistency held is a team assessment. The public trailer shows excerpts, not enough to audit every scene of an 80-minute feature.

• Read the manager's character, location and node-system account ↗

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Poqr Mher project trailer

The generation total is team-reported; the director reviewed shots in context, and management tracked scene approvals per week.

The manager reports 3,286 video generations, mostly with Kling. For difficult motion, action or emotional close-ups, she says the team used Seedance at roughly ten times the per-generation cost. That is her project comparison, not a current price table or the project's final bill. Freepik nodes reused workflows, Milanote held shots and visual versions, and Asana held tasks and blockers. The director reviewed each shot against its board and neighboring shots.

The team planned separate reviews of stills and motion, then found that attractive stills often failed once animated through broken poses or flickering light. They dropped the separate still approval and strengthened character, location and board inputs before generation. The manager tracked approved scenes per week rather than treating generation count as progress. This is her first-party account of a production decision; approval logs, yield and detailed costs are not public.

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible

The manager describes specific failures and rework decisions, without publishing same-shot failed and repaired footage.

About two-thirds through production, the manager says the video models were unreliable for roughly ten days: quality fell, waits grew and previously useful prompts returned unusable footage. The team moved to backup platforms, made some scenes at lower quality for later rework and parked hard shots while creators completed others. She does not publish service incident records, every failed output or controlled before-and-after footage. This is the team's account of that period.

A more concrete failure concerned a prop: a book required across a sequence was missing from generated footage, so six scenes had to be redone. The team added a script-to-board prop check. Sponsors also changed imagery: they held a shopping-shelf scene while brands were uncertain, then rebuilt other scenes when an e-payment sponsor was confirmed. The transferable decision is to identify objects that must persist across shots and commercial details likely to change late, then decide which shots to delay or remake. The account does not show that a new prompt alone fixed these problems.

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible

The manager says the final soundtrack was unresolved four days before delivery, while director review remained the quality gate.

The manager opens by saying the final soundtrack was still unresolved four days before delivery. Her article does not detail recording, music rights, dialogue, mixing or subtitling. The visual-generation tools she names therefore do not describe the whole film. The festival lists Armenian with English subtitles, documenting how it is presented in the programme, not who made every audio or language asset or at what cost.

She describes a stretch when director reviews slowed under competing demands. The team did not substitute another approver; they made the delay visible and routed creators to ungenerated scenes. A striking image must still connect to neighboring shots, emotion and the director's intent. That is why approved scenes, rather than model outputs, were their progress measure.

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Armenian Film Festival Australia: Poqr Mher programme

Festival listings, reported sponsors and model-cost differences answer different questions and do not yield profit.

The public trailer and festival programme show a feature prepared for outside exhibition. The manager says the team explored candy, milk and juice sponsors and remade scenes for an e-payment partner, but gives no fees, contracts, invoices or receipts. The festival listing supplies no box-office settlement. Her per-generation Kling versus Seedance comparison also omits usage mix, subscriptions, labour, voice and release costs. Total production cost, break-even point and profit cannot be calculated.

A traditional epic may inspire a story, while a particular script, character design, music, voice, performer likeness, brand mark and generated asset can each carry separate conditions. The public account does not provide a rights ledger. Access to a trailer or workflow description is not permission to reuse the film's images, audio or production assets. A similar project should begin with material one may use and separate agreements for sponsor integration and exhibition.

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Armenian Film Festival Australia: Poqr Mher programme

Test cross-shot props, motion and editability in a small sequence before expanding to a feature or paid work.

A small trial could use a story you may adapt: one location, one main character and three connected scenes. Mark the character's appearance, props that must persist and the emotional change in each scene. Build a location master, character views and expression references, then a short board. Keep each generation's references, prompt, tool, cost and failure reason. Let one creator own the location, edit the shots into a sequence and ask another person to check props, action and emotion against the board. This three-scene exercise is BitShovel's suggestion drawn from the feature's account, not a recipe published by its makers.

If the sample may become a commercial project or public screening, add a rights ledger, review owner, backup tools, revision rules and cost records. Track audience response, bookings, actual ticket settlements and sponsor income separately. A programme listing is not a ticket sale, and a high generation count is not a count of approved scenes.

• Read the manager's original production account ↗

Sources for this section: Mariam Hakobyan: Producing a Film That Wasn't Supposed to Be Possible · Armenian Film Festival Australia: Poqr Mher programme