Layer Video Generation

Overview

Video models take a prompt the way a director takes a shot list, not the way an illustrator takes a brief. A video prompt describes what changes over the clip. A prompt that only describes the scene yields a near-still image with drifting artefacts, which is the most common failure here.

Pick the use case first: text_to_video from nothing, image_to_video to animate an existing frame, video_extend to continue a clip, video_editing to alter one (repairing a clip rather than rerolling it is layer-video-fixes). The loop and spend gate are in the layer skill. Assembling finished clips into one deliverable is layer-video-timeline.

If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add layerai/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Anchor on an image whenever one exists

image_to_video beats text_to_video almost every time a look has already been approved. The first frame fixes the character design, palette, and composition, so the model spends its capacity on motion instead of re-inventing the subject. Attach the still as init_image or first_frame depending on what the model accepts, and where a model takes both first_frame and last_frame, supplying both is the most reliable way to control where a shot ends up.

Generate the key frame with layer-image, approve it, then animate it.

Writing the motion

Name three things, in this order:

Subject motion: what moves, in what direction, at what speed. "The banner snaps taut in a gust, then settles" is a shot. "A banner" is a photograph.
Camera: static, slow push in, orbit, handheld drift, crane down. State "locked-off camera" when the camera must not move, because models add drift by default.
Rate: slow, deliberate, sudden, gradual over the full clip. Models otherwise pick a middling pace that reads as neither.

Model families disagree about conventions, so read get_base_model for the one you picked rather than reusing phrasing that worked on another family. Where a model exposes structured video_effects ([{type: "DOLLY_IN", weight}]), prefer that over prose for the same move. Choosing moves, lens, light, technique presets and pacing is layer-video-direction.

Duration, audio, and lip sync

Clip length is per-model and short, usually a handful of seconds. Do not plan a 30-second piece as one generation without confirming a model supports it.

Some models choose their own length and ignore duration_seconds entirely. When the clip has to land on an exact length, filter for fixed_duration, which excludes them. Without it a model may simply hand back a different length, and nothing in the response reads as an error.

generate_audio is a capability, so filter for it when the clip needs native sound rather than assuming silence can be fixed later. Dialogue and lip sync on a character are layer-talking-characters. filter.capabilities is an object of booleans, not a list of names: {generate_audio: true}.

Generation is rate limited per user: 600 per minute across all modalities, plus 300 video per minute. On RATE_LIMITED, wait the Retry-After interval, then retry; never cancel a run the user asked for to free a slot.

Drafts before finals

Where get_base_model reports capabilities.draft, iterate on drafts: draft: true renders a quick, low-cost take. Finish only the one you keep, with a second execute_forge: base_model_id set to the draft's draft.completion_base_model_id, batch_size: 1, the draft's file_id as a guidance file of type draft.completion_guidance_type, and no prompt or other generation parameters, since the completion keeps the draft's prompt, motion and length. The completion bills per second of the draft, so estimate it with the completion model and the draft's duration_seconds. A draft completes once and may expire (draft.status, expires_at), so finish a keeper promptly.

Subtitles

Burned-in captions are their own use case, video_subtitle, not something to prompt for. Read the model's subtitle_presets and pass subtitle_preset spelled exactly as listed, plus subtitle_alignment and colours where its capabilities report them. Dynamic presets bill higher, so estimate with the same preset. Captions in other languages, and dubbing, are layer-localization.

Multi-shot pieces

An ad, trailer, or cutscene is several shots. One generation holds its own continuity and a cut does not, so check first whether a model can render the piece in one. Otherwise lock the look with one approved still, then follow layer-video-continuity to chain shots and hold the character, and layer-video-timeline to cut them. Storyboard anything over three shots (layer-art-direction).

Loops

Filter for the loop capability first: it selects models that compose a seam-free loop, not ones that accept the word "looping" and return a clip that jumps at the cut. Pair it with fixed_duration for a specific length.

Failing that, generate the anchor frame and use it as both first_frame and last_frame where the model accepts both, so the motion at least returns to where it started.

Either way, keep the motion cyclical (a flag in wind, a torch flicker, a hovering idle); one-way motion, such as a character walking out of frame, cannot loop.

Worked example

"Animate this approved character key frame into a 5-second idle loop."

The frame exists from a previous run, so it already has a file_id.
list_base_models with filter.use_case: "image_to_video" and filter.capabilities: {loop: true, fixed_duration: true}, since this has to loop seamlessly and land on five seconds. Take the first result, then get_base_model to see whether it accepts last_frame and structured video_effects.
Prompt the motion only: "The character breathes slowly, cloak drifting in a light breeze, weight shifting subtly between feet. Locked-off camera. Gentle, continuous, no change in pose by the end of the clip."
Pass the same file as first_frame and last_frame so the loop closes.
Estimate. Video is priced well above images, so this will often clear 20 CUs and need an explicit yes.
Execute, poll at poll_interval_seconds, and check the seam by watching the end run into the start before calling it a loop.

Common mistakes

Prompting the scene instead of the motion, then reporting the model "barely moved".
Using text_to_video when an approved still already exists.
Leaving the camera unspecified, then fighting drift the model added by default.
Planning a long piece as one clip without checking a model's duration limit.
Assuming audio can be added later on a model that could have generated it natively.
Prompting for a loop instead of filtering for the loop capability, or leaving fixed_duration off when the length matters.
Running full-quality takes to explore when the model offers drafts.
Calling a clip a loop without watching the seam.

--- name: layer-video description: "Use when generating video with Layer: text-to-video, animating a still image, extending a clip, adding camera motion, generating native audio or lip sync, looping animations, or planning a multi-shot ad, trailer, or cutscene. Also when a video prompt produces the wrong motion or the shot drifts off the source image. Keywords: txt2vid, img2vid, image to video, camera motion, video effects, loop, seamless, trailer, cutscene, lipsync." license: MIT --- # Layer Video Generation ## Overview Video models take a prompt the way a director takes a shot list, not the way an illustrator takes a brief. A video prompt describes **what changes over the clip**. A prompt that only describes the scene yields a near-still image with drifting artefacts, which is the most common failure here. Pick the use case first: `text_to_video` from nothing, `image_to_video` to animate an existing frame, `video_extend` to continue a clip, `video_editing` to alter one (repairing a clip rather than rerolling it is `layer-video-fixes`). The loop and spend gate are in the `layer` skill. Assembling finished clips into one deliverable is `layer-video-timeline`. If a sibling skill named here is missing from your available skills, ask the user to install it (`npx skills add layerai/skills --skill <name>`); unattended, proceed from tool schemas and flag the gap. ## Anchor on an image whenever one exists `image_to_video` beats `text_to_video` almost every time a look has already been approved. The first frame fixes the character design, palette, and composition, so the model spends its capacity on motion instead of re-inventing the subject. Attach the still as `init_image` or `first_frame` depending on what the model accepts, and where a model takes both `first_frame` and `last_frame`, supplying both is the most reliable way to control where a shot ends up. Generate the key frame with `layer-image`, approve it, then animate it. ## Writing the motion Name three things, in this order: 1. **Subject motion**: what moves, in what direction, at what speed. "The banner snaps taut in a gust, then settles" is a shot. "A banner" is a photograph. 2. **Camera**: static, slow push in, orbit, handheld drift, crane down. State "locked-off camera" when the camera must not move, because models add drift by default. 3. **Rate**: slow, deliberate, sudden, gradual over the full clip. Models otherwise pick a middling pace that reads as neither. Model families disagree about conventions, so read `get_base_model` for the one you picked rather than reusing phrasing that worked on another family. Where a model exposes structured `video_effects` (`[{type: "DOLLY_IN", weight}]`), prefer that over prose for the same move. Choosing moves, lens, light, technique presets and pacing is `layer-video-direction`. ## Duration, audio, and lip sync Clip length is per-model and short, usually a handful of seconds. Do not plan a 30-second piece as one generation without confirming a model supports it. Some models choose their own length and ignore `duration_seconds` entirely. When the clip has to land on an exact length, filter for `fixed_duration`, which excludes them. Without it a model may simply hand back a different length, and nothing in the response reads as an error. `generate_audio` is a capability, so filter for it when the clip needs native sound rather than assuming silence can be fixed later. Dialogue and lip sync on a character are `layer-talking-characters`. `filter.capabilities` is an object of booleans, not a list of names: `{generate_audio: true}`. Generation is [rate limited per user](https://layer.ai/docs/rate-limits): 600 per minute across all modalities, plus 300 video per minute. On `RATE_LIMITED`, wait the `Retry-After` interval, then retry; never cancel a run the user asked for to free a slot. ## Drafts before finals Where `get_base_model` reports `capabilities.draft`, iterate on drafts: `draft: true` renders a quick, low-cost take. Finish only the one you keep, with a second `execute_forge`: `base_model_id` set to the draft's `draft.completion_base_model_id`, `batch_size: 1`, the draft's `file_id` as a guidance file of type `draft.completion_guidance_type`, and no prompt or other generation parameters, since the completion keeps the draft's prompt, motion and length. The completion bills per second of the draft, so estimate it with the completion model and the draft's `duration_seconds`. A draft completes once and may expire (`draft.status`, `expires_at`), so finish a keeper promptly. ## Subtitles Burned-in captions are their own use case, `video_subtitle`, not something to prompt for. Read the model's `subtitle_presets` and pass `subtitle_preset` spelled exactly as listed, plus `subtitle_alignment` and colours where its capabilities report them. Dynamic presets bill higher, so estimate with the same preset. Captions in other languages, and dubbing, are `layer-localization`. ## Multi-shot pieces An ad, trailer, or cutscene is several shots. One generation holds its own continuity and a cut does not, so check first whether a model can render the piece in one. Otherwise lock the look with one approved still, then follow `layer-video-continuity` to chain shots and hold the character, and `layer-video-timeline` to cut them. Storyboard anything over three shots (`layer-art-direction`). ## Loops Filter for the `loop` capability first: it selects models that compose a seam-free loop, not ones that accept the word "looping" and return a clip that jumps at the cut. Pair it with `fixed_duration` for a specific length. Failing that, generate the anchor frame and use it as both `first_frame` and `last_frame` where the model accepts both, so the motion at least returns to where it started. Either way, keep the motion cyclical (a flag in wind, a torch flicker, a hovering idle); one-way motion, such as a character walking out of frame, cannot loop. ## Worked example "Animate this approved character key frame into a 5-second idle loop." 1. The frame exists from a previous run, so it already has a `file_id`. 2. `list_base_models` with `filter.use_case: "image_to_video"` and `filter.capabilities: {loop: true, fixed_duration: true}`, since this has to loop seamlessly and land on five seconds. Take the first result, then `get_base_model` to see whether it accepts `last_frame` and structured `video_effects`. 3. Prompt the motion only: "The character breathes slowly, cloak drifting in a light breeze, weight shifting subtly between feet. Locked-off camera. Gentle, continuous, no change in pose by the end of the clip." 4. Pass the same file as `first_frame` and `last_frame` so the loop closes. 5. Estimate. Video is priced well above images, so this will often clear 20 CUs and need an explicit yes. 6. Execute, poll at `poll_interval_seconds`, and check the seam by watching the end run into the start before calling it a loop. ## Common mistakes - Prompting the scene instead of the motion, then reporting the model "barely moved". - Using `text_to_video` when an approved still already exists. - Leaving the camera unspecified, then fighting drift the model added by default. - Planning a long piece as one clip without checking a model's duration limit. - Assuming audio can be added later on a model that could have generated it natively. - Prompting for a loop instead of filtering for the `loop` capability, or leaving `fixed_duration` off when the length matters. - Running full-quality takes to explore when the model offers drafts. - Calling a clip a loop without watching the seam.

SKILL.md