How we make brand films · Video factory

Example film

Warm yellow memory beside cold white laptop grind

Locked story: yellow family bond → white laptop grind → math plate → Install / month returned.

Exact Fun Fact line: 2 Hours = 1 Month. This system exists to make that film.

Act 1 is warm tungsten, 90s to Y2K. Act 2 is cold white isolation late at night. Act 3 is the site math plate. Act 4 soft-sells Install and the month returned. One recurring object and one sound motif brand the craft. They are not a product name. Do not rewrite this story inside the tool pick.

Jobs

Five jobs on the line. Still lock. People and rooms motion. Type and math plate. Stitch and NLE. Motif audio.

A ~45s film is six to twelve clips plus plates. Not one model call.

Still lock owns faces, wardrobe, rooms, and the recurring object before any motion. People and rooms own Act 1 and Act 2 clips with identity hold. Type and math owns readable numbers. Stitch and NLE owns timeline, grade, and mix. Motif audio owns the designed hit under both lights. Native model audio is a temp bed only.

Run order

Stills to I2V to code plate to FFmpeg NLE to motif sound

Stills → I2V or gen clips → code plate for exact type → FFmpeg or NLE assemble → designed sound motif.

Class owns each step. Swap tools inside a class. Do not swap the order.

Approve stills under yellow and white light first. Animate from those stills. Render 2 Hours = 1 Month in the code plate class. Concat lossless, then grade and mix. Design the recurring sting in the NLE. Do not ask the video model to finish the mix.

Still lock

Approve yellow and white stills before motion

Freeze faces, wardrobe, rooms, and the recurring object as stills. Both light worlds.

Tool class: still generator, then those files as ingredients or start frames.

Yellow room and white room must read as one home under two keys. Name real furniture. A vague cozy room invents a catalog set. The recurring object must appear in both grades. Midjourney, Flux, Seedream, and Flow image models all sit in this class. Pick later. Lock the stills first.

Motion class

Kelvin strip from warm porch to cold laptop to math plate

People and rooms motion. Same faces across Act 1 and Act 2. Lighting jump stays one film.

Tool class: filmmaking UI or model API with refs, first and last frames, then NLE for the cut.

Options inside the class, not winners: Flow plus Veo or Omni. Seedance via Dreamina or an API such as fal. Seedance via Higgsfield. Stills to image-to-video to FFmpeg. Higgsfield multi-model when you want engine swaps in one UI. Test face hold on one Act 1 close before you buy a full timeline. Quality or Standard tiers often drop ingredients or extend. Read the tier matrix before you spend.

Code plate

2 Hours equals 1 Month as cream type on night ground

Exact line 2 Hours = 1 Month. Match site cream on night. Never trust a video model for this type.

Tool class: Remotion or other code render. Optional I2V metaphor under Act 4 only.

Visible Fun Fact stack: label Fun Fact, math line, then the body sentence from the site. Inter for display. Fraunces stays on the wordmark only. Center the stack. Hold the plate. Then leave it. Remotion is one renderer in the class. It is not a vendor lock.

Prompt discipline

Lock person, wardrobe, room, and recurring triad once. Paste the block into every clip.

One clip, one beat. Do not chain A then B then C in one generation.

One clip equals one beat, not A then B then C

Change only action, camera, and audio that belong to that shot. Keep the Fun Fact string exact. Negative blocks name what to exclude without writing no or don't. This discipline feeds every tool in every class. A late triad swap reads as a new film.

Assemble

Recurring object mug beside sound wave motif

Stitch in FFmpeg or an NLE. Object plus sound motif under both lights. Soft-sell Install late.

Classes next. Options under More. No winner named here.

People and rooms class beside type and math class

Class map. Still lock → still generator. People and rooms → Flow/Veo, Seedance, Higgsfield, or stills→I2V. Type and math → Remotion or code. Assembly → FFmpeg concat then NLE. Motif → designed audio in the NLE.

Criteria before any pick. Which shots must hold a face across Act 1 and Act 2? Which shots are type or math and must leave the picture model? Filmmaking UI or API plus NLE? Does Quality still support ingredients, last frame, and extend? Native finish resolution versus upscale? Cost of one keeper and one retry? Do credits expire in the shoot window? Is identity exportable if you change houses? Can you extend from a downloaded file or only a live URI? Where is the recurring sound designed? If the answer is the model, rewrite the answer. After the packs, test one family on a single Act 1 close. Know what stops you.