layer-video-fixes

Layer Video Fixes

Overview

A clip that is 90 percent right is worth more than a fresh roll. Rerolling throws away the motion, the timing and the performance that already work, and buys a new set of problems. Layer exposes a use case for most single defects, so the job is to name the defect, pick the use case that fixes only that, and leave everything else as it was.

Generating a clip from scratch is layer-video; cutting finished clips together is layer-video-timeline. The loop, the spend gate and polling are in the layer skill.

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.

Quick reference

Each route is list_base_models with filter.use_case set as below; take the first result.

Symptom	use_case
Clip ends too soon, action is cut off	video_extend
A moment is missing in the middle	insert_video
One element is wrong (colour, prop, background)	video_editing
The wrong object is in shot	video_object_swap
The wrong character, or an off-model one	video_character_swap
Right performance, wrong subject	motion_transfer
Lighting fights the scene it is cut into	relight_video
Motion stutters, or the shot needs slow motion	video_interpolation
Too small for delivery	upscaling
Grain or compression noise	denoise_video
Soft or motion-blurred	deblur_video
Standard range master needed as HDR	sdr_to_hdr_video

Read the model before the first call

These models disagree about inputs more than generation models do. The schema does not fix which guidance slot a repair model reads the source clip from (init_video or reference_video), nor where a replacement subject goes (reference_image, face, or an element_* type). Call get_base_model on the chosen model and use the slots it lists; never assume one because another model took it. When a fix needs two inputs at once, narrow the search with filter.guidance_file_types, for example ["init_video", "reference_image"], which keeps only models that accept both.

Guidance files are {file_id, type} objects in guidance_files. A clip from an earlier run already has a file_id; pass it directly. Upload only footage that is local, with request_file_upload_url.

Repair from the original

Every pass through a model costs detail. Three fixes chained on each other look worse than three fixes each made from the source, and the losses compound in faces and fine texture first. Keep the original file_id as the master, apply each fix to the master where the fixes are independent, and chain only when a later fix genuinely needs the earlier result (upscale last, after the content is settled, since an upscale of a clip you then edit is wasted).

To fix only part of a clip (slow motion on the last two seconds), cut that span out first with a render_video_timeline edit (start_from_in_seconds and duration_in_frames, layer-video-timeline), then repair the cut file.

Sound, length and frame rate

• Sound. Set keep_audio: true to carry the source's audio through a fix. Without it, assume the output may be silent, and check before delivering.
• Length. Extensions, inserts and interpolation change a clip's duration, and some swaps do too. Report the new length to the user, especially when the clip sits in an edit cut to music. An interpolated slow-motion shot is longer by design; a swap that came back shorter is a defect.
• Frame rate. fps is honoured only by the models that take it (get_base_model says so). Ask for interpolation to reach a target rate rather than generating at one the model ignores.
• Size. For a reframe, pass both width and height. With either one missing and a source attached, the output geometry is taken from the source file.

For an upscale, upscale_ratio sets the factor (2.0, 4.0). creativity and resemblance trade invented detail against fidelity; keep resemblance high on footage with faces or text.

Pricing

The forge guide prices a video edit by its source's length and an upscale by its source's dimensions. An estimate without the source attached prices a different run, so pass estimate_forge_price exactly the inputs execute_forge will get. Several fixes on clips of different lengths are several estimates, never the first one multiplied. Sum the job and ask before spending above 20 CUs.

Worked example

"The hero's 6-second clip is perfect except the sword turned into a staff halfway through. Also make it 4K for the trailer."

1. The clip is a previous run's output, so it has a file_id. Keep it as the master.
2. Defect one is a wrong object: list_base_models with filter.use_case: "video_object_swap". Take the first result and call get_base_model for its source slot and replacement slot.
3. Defect two is size: list_base_models with filter.use_case: "upscaling" and filter.modality: "video"; get_base_model for its slot and ratio limits.
4. Price both before running either: estimate_forge_price for the swap with the master attached in its listed slot, the approved sword still as the replacement, keep_audio: true, and a prompt naming only the change: "Replace the staff in the hero's right hand with the approved longsword. Keep the motion, the timing and everything else unchanged." Then estimate the upscale with upscale_ratio: 2.0 against a 6-second source. Present the total; above 20 CUs wait for a yes.
5. execute_forge the swap; poll get_forge_run at poll_interval_seconds. Check the length is still 6 seconds and the sword holds across every frame.
6. Upscale the swap's output (file_id from step 5), the one fix that must come last. Poll, then report both file_id values and any change in length.

Common mistakes

• Rerolling a clip whose only defect a single use case could fix.
• Assuming the source slot instead of reading it from get_base_model.
• Chaining independent fixes off each other instead of each off the master.
• Upscaling first, then editing the upscaled clip.
• Losing the soundtrack by leaving keep_audio unset.
• Delivering a fixed clip without saying its length changed.
• Pricing several fixes from one estimate when their sources differ in length or size.
• Prompting a repair with a full scene description instead of only the change.