Producer Runs & Budgets

A Producer run walks your project through six stages in order and records every decision it makes:

1. Analysis — reuses your track analysis and alignment, or produces it if it is missing. A script-first project (no audio track) skips this stage entirely: nothing is transcribed or billed for it.
2. Planning— the treatment and storyboard: one scene per window of the track, cut to what your video provider can render. For a script-first project the film's length is the brief's target duration and the script is the only source
3. Generation — a still image for each scene, then a video clip animated from that image
4. Review — every clip is scored; clips under the quality target are repaired within the remaining budget
5. Assembly — clips are cut together against the original audio
6. Export — the final video is rendered and made available for download

On laptops and smaller screens the Producer shows one panel at a time with tabs (Conversation, Artifact, Plan & run) so the preview and its actions keep the full width; very wide screens show all three side by side. The Studio navigation folds into a menu button on small screens.

Runs execute on Filmgenic's production workers, not in your browser. You can close the tab; the run keeps going and the Producer panel shows its current stage, the credits spent so far, and any decision waiting for you when you come back.

Every run needs an explicit ceiling before it is created; a blank field is not an unbounded run. Both places you can start a run, the guided first run and the workspace's New runform, pre-fill the credit ceiling from the planner's estimate for this project and ask for:

• Credit ceiling — the maximum number of credits the run may spend, a whole number of one or more
• Currency ceiling — an optional monetary cap on provider costs, above zero, for plans that expose it

Budget is reserved before each paid step, so a run cannot overshoot its ceiling by a partially finished scene. When the remaining budget cannot cover the next step, a run created with the defaults pauses for your decision rather than stopping: you can raise the credit ceiling, approve the specific step it is asking about, or deliver the current draft. A run that has spent over 80% of its ceiling is flagged in the panel so the pause is not a surprise.

A retry of a failed run reuses every scene that already finished. Only the scenes that still need work are billed again.

Credits are what you are charged, and nothing about them changes here. Underneath them are the amounts the AI providers charge us for your film, and a project now keeps a partial record of those too: one line per paid video call, with the scene it belongs to and what it bought. Until recently the run page could only show a prediction, because no run had ever written down, in dollars, what one of its clips cost — the product tracked your credits, and its own provider bill arrived as a single monthly total nobody could break down by project, run or scene.

• Generated is not the same as delivered. A film is billed by the second of video generated, and some of those seconds never reach your cut — a take the run discarded, a scene the reviewer had repaired, a clip a later version replaced. Those seconds are recorded against the reason they were spent, so the difference between what was bought and what you received is a number rather than a feeling.
• A reused clip is recorded once, not twice. Where a clip is recorded at all, it is recorded under the identity of the provider job that made it, so a retried run, a resumed job or a second run over the same scenes writes that clip down once and counts its seconds once as delivered to your film. That is a property of the record, not a promise about the bill, and it reaches exactly as far as the record does. A paid call on one of the paths that write nothing down — the images, the voices, the planning and review — cannot be recognised a second time, because it was never recognised a first time. And asking for a fresh take of a scene is not reuse: it buys a new clip, and the record shows it as the second purchase it is.
• What is confirmed and what is not are kept apart. One rate — $0.40 per generated second of video with audio — has been checked line by line against a real provider invoice. What that invoice does not say is which of the video models we run produced those seconds: it names a billing line, and our code names a model, and nobody has yet matched the two. So the video we can value is valued at that real rate and still reported apart from anything confirmed, and it stays there until a future bill settles which model was billed. Every figure is a floor, not a total: the record can only cover calls it wrote down, so it is never presented as your invoice.
• Work we cannot price is listed, not counted as free. For some models — newer previews, and the lightest video tier — no invoice line has been matched to them, and we will not borrow a rate from a different one. Those calls are recorded with what they produced and no dollar figure, and the record says how many of them there were and why. A zero there would tell you the film cost less than it did.

Three honest limits. The record covers the video clips your film is made of — those generated by a run and those bought by a retry or a repair you ask for in the Studio, including takes that failed or were replaced. Video is the largest thing we pay for, but it is not all of it: the images, the voices, and the planning and review behind them are not recorded at all, so their spend is missing from the figures rather than counted as zero. Even within video, a line can go missing — if writing one fails the clip still stands and the failure is logged for us, not silently swallowed, but the figures under-report until someone acts on it. And the record begins where it was switched on — spend from before then cannot be broken down after the fact. It is a record of where the money went, and where we are not yet sure; it is not your invoice, and your credit balance remains the thing you are billed on.

The mode is fixed for the life of a run. Publishing to external platforms is disabled in every mode.

When a run needs you, it creates a checkpoint and the Producer panel shows Approve and Reject controls with the reason. Checkpoint types you may see:

• Brief — confirm the creative brief before planning starts
• Storyboard and Cost — confirm the scene plan or a cost projection where the mode requires it
• Risky content — a prompt or image tripped a provider safety filter and needs a rewrite or your explicit go-ahead
• Policy exception — the next step would exceed the ceiling; approving authorizes that one step only, at the projected cost shown
• Final review — the reviewed cut is ready; approve to assemble and export

Checkpoints on a run that has finished, failed or been cancelled stay visible as history and can no longer be decided; a delivery run is listed alongside the run whose clips it publishes.

An approval is bound to the exact step and cost it was shown for. Raising the ceiling and approving a checkpoint are separate actions; neither silently authorizes more than you saw.

If a run pauses at its ceiling during review and you do not want to spend more on repairs, you can deliver the clips that exist. On the paused run's policy-exception checkpoint, choose Deliver current draft. A small delivery run (assembly and export only, no generation) appears in the run list, cuts every scene that has a clip against your audio, and publishes the video. The paused run is left exactly as it was, with its findings, so you can come back and approve the repair later.

The delivered file is labelled Draft with findings: it is complete, but the reviewer still flagged scenes, and the label says so wherever the video is listed. A run that finishes its review instead produces a file labelled Quality passed. Draft delivery is only offered for a budget pause during review; it never bypasses a safety or rights stop, and it needs every scene to have a clip first.

The Producer only acts on requests it can turn into an exact, validated command. Brief edits ("set tone to cinematic") apply while the brief is awaiting your approval. Production requests never start paid work from the chat: they answer with a proposal and a next step you take yourself.

• explain state: where the run is, which shots have clips, what is waiting for you.
• revise shot 4: close-up, keep the actor and costume: the scoped, costed proposal for that one shot and a link to apply it.
• show takes for shot 4 / use take 2 for shot 4: the recorded takes, and a free switch you confirm in the Studio.
• render a preview / export the film: what it costs and where to start it.

A vague request ("make it better") is not guessed at; the Producer says what it can do instead.

Every film Filmgenic made before this played pictures under music and said nothing. Tick Narrate this film on the brief and someone speaks over it.

• Pick a voice. Thirty voices, each listed with what it sounds like — informative, warm, breathy, gravelly — so you are choosing a delivery, not a name. Add a direction like "measured, observational" and it shapes how the lines are read without ever being read aloud.
• The planner writes the lines. One line per shot, budgeted to that shot's length at an unhurried pace, and told not to describe the picture you can already see. A line that comes back too long is trimmed at a sentence or a word — never mid-word, and never sped up to fit.
• Or write them yourself. Paste a script and the planner divides your words across the shots instead of inventing its own.
• The music gets out of the way. The narrator sits on their own track and everything else ducks under them while they speak, then comes back up.
• Re-runs do not pay twice. Each take is stored under a fingerprint of its line, voice, language and delivery, so an unchanged line reuses its recording and changing one line re-records only that shot.
• One shot failing does not lose the film. A line that cannot be spoken leaves that shot silent and is reported; the rest of the film is still narrated.

Narration works in any creative mode: it is what turns a documentary into a documentary and a story into a told one.

Choose Characters speaking to each other instead of a narrator and the film gets a cast. Write one character per line as Name :: Voice, and the planner writes what each of them says in each scene.

• Two people to a scene. A take is voiced by the model and it voices two, so at most two characters speak in any one scene. A three-hander is written as consecutive two-handed scenes — the planner is told this, and a scene that comes back with three speakers has the third reported and cut rather than quietly mixed into one voice.
• The exchange fits the shot. The word budget covers every line in the scene together, not each line, so two people talking do not overrun the picture.
• Only your cast speaks. A line attributed to someone who is not in the cast is dropped and reported, so the film never contains a voice you did not cast.
• Recasting re-records. Change a character's voice and every take they speak in is recorded again; everything else is reused.
• A character can have a face as well as a voice. Tie a cast member to a character asset in your creative library and the same person keeps both across the film.

Scenes where nobody talks are fine — they stay silent rather than being filled. In the editor the spoken track is called Dialogue rather than Narration, and each take records who is in it.

A project in Advertising mode carries an advertising block on its brief: the product, the offer, the exact call to action and the claims the film may make. Planning turns it into a deliberate beat structure (hook, problem, benefit, proof, call to action) and never asks the image or video model to render text, logos, prices or URLs.

• Claims need support. A claim that requires support must carry its source and be approved on the brief; a run cannot start while any such claim is unsourced or unapproved, and the reason names the claim.
• The CTA is a layer, not a picture. When the film is assembled, the call to action (and its destination) is composed as an editable text cue over the final beat, using the bound brand kit's typeface and colour when a kit is bound.
• The logo is the approved asset. A logo from the bound brand kit is placed as an overlay layer over the final beat, so it is exact and editable in the editor.

One approved commercial rarely ships as one film. The Variants section of the run panel (advertising projects) fans a concept out along the dimensions you declare — hooks, calls to action, aspect ratios — and copies everything else as is: product, claims, references, brand kit, providers.

• Project first, spend later. "Project the cost" lists every variant and the projected total from the same estimate a single run uses (clean run plus repair headroom per variant). Nothing is created and nothing can spend yet.
• One aggregate ceiling. Authorizing sets a single number for the whole family; each variant becomes its own project and run holding an equal share of it. Below the family's minimum (variants × clean-run credits) the authorization is refused, and so is one your account's credits do not cover.
• Shares cannot sneak past the aggregate. Raising one variant's ceiling is refused when it would take the family past what was authorized; widen the aggregate first. Every variant links to its own Producer page, and the family shows what has been spent and reserved across all of them.