Motion assets

Generated assets make or break an animated site, and they cost real money per attempt. This skill is the production discipline around the models: decide everything that can be decided before spending, make every asset in one look, generate in the order that lets later assets anchor on earlier ones, check each one with numbers, and write it all down.

Tools

Related: alpha-cutout (clips that need transparency), scroll-film (prep.py seams scores joins between chained clips), cursor-rig (head-turn clips).

Needs Python 3.9+, pip install numpy opencv-python-headless, ffmpeg.

Before spending anything

Write a short plan and show it to the person paying:

1. The concept in two sentences, and what moves on the page.
2. The asset list: every image and clip, its size and aspect, its role on the page, and its exact prompt.
3. The model for each, chosen by capability (table below), and the cost of each and in total. Calibrate first: run one still and one short clip, read what was charged, then estimate the rest from that. Plans and prices change; never quote a price from memory.
4. What the visitor sees with reduced motion and on a phone.

Get a yes, then generate.

Choose models by capability

Observed while building with these models in 2026 (check current docs; behaviour changes):

• Video models drift a flat background colour within a single clip (pure magenta came back pale pink); key per frame.
• Models change small details mid-clip (a black cap turned green late in a serve). Watch the whole clip, not the first and last frames.
• Some higher resolutions and quality modes are gated by subscription plan. A rejected submission usually costs nothing: test the cheapest request that proves access.
• Rate limits are tight on many plans: two or three video jobs in flight. Submit in small batches and resubmit only the job that returned 429 once others finish.
• A "server not responding" error does not always mean no job was created. List recent jobs before resubmitting, or you pay twice.
• A request can come back as a suggested preset instead of a job; decline the preset and resubmit.

One look across every asset

• Write a style preamble (medium, palette, light, lens, texture) once and prepend it byte-for-byte to every image prompt.
• Make one hero still first, approve it, then pass it as the style reference to every later image.
• Use one video model for a whole chain of clips. Each model has its own motion, colour and grain; switching mid-film shows as a style jump at the seam.

Generation order

1. Stills: hero, then keyframes (every start and end frame of every clip), then supporting images. Review them side by side before any video.
2. Clips, each anchored on approved stills. For chains, anchor each clip's start on the exact extracted last frame of the previous clip, not on the still, so colour drift does not accumulate as a jump.
3. Check every clip: loop_check.py for loops, scroll-film/scripts/prep.py seams for chains, the whole clip by eye for detail drift.

Clip prompts that behave:

• Name the camera: "locked-off camera" or the exact move ("slow push in, no rotation").
• Name what must not move: "the body, background and lighting stay perfectly still".
• For loops: "constant, even motion; no slowing down at the end". Models ease to a stop by default, which hitches every loop. loop_check.py reports it.
• Repeat the plate colour for anything that will be keyed.

Fixing instead of regenerating

Log everything

Every generation goes in prompts/asset-prompts.md (and machine-readable in a JSONL log): the asset name, model, exact prompt, input images, settings, job id, cost, and which output was kept. It is the only way to regenerate one asset months later in the same style, and the honest record of what was AI-generated.