Restyling an Existing Clip with AI: Recast, Swap, Restyle
How OMNI Labs edits existing video clips using Higgsfield Genjutsu: the three edit modes, duration-based billing, whole-second trimming, and Seedance video-edit.
OMNI LabsUpdated
Key takeaways
Higgsfield Genjutsu provides three distinct edit operations: motion transfer, object swap, and clip restyle.
Billed seconds round up to the nearest whole second; trimming a 5.04-second source to 5 seconds saves 20%.
Restyle operations require source clips between 4 and 30 seconds, maintaining original audio without replacement.
The restyle API endpoint resides exclusively under the higgsfield/ path, whereas older endpoints used higgsfiled/.
Seedance 2.5 video-edit costs about $0.277 per second, making targeted clip trims far cheaper than editing full takes.
The three operational modes of Higgsfield Genjutsu
The Higgsfield Genjutsu edit runs in three modes. Motion transfer recasts an existing clip onto a new subject from 1 to 8 reference images. Object swap replaces one product, garment or prop with 1 to 8 reference images. Restyle redraws the whole clip in one of 20 presets, with 0 to 5 optional character references.
Modifying established footage through generative AI requires selecting the appropriate technical mechanism for the intended editorial outcome. Motion transfer preserves the underlying camera trajectory, physical velocity, and skeletal kinetics of the source footage while completely recasting the actor into a new synthetic subject. Object swap isolates specific geometry within the scene—such as a handheld luxury bag, a pair of sunglasses, or a jacket—and substitutes a new product without disrupting background lighting or actor performance. Restyle operates globally across the canvas, re-synthesizing the entire aesthetic through 20 curated stylistic presets.
With several characters, the reference images have no per-image subject mapping. We name who is who in the prompt in the same order as the image list. It is guidance, not a binding.
When multiple human subjects appear in a single source clip, multi-character conditioning presents distinct routing challenges. Because the Genjutsu submission schema does not provide explicit index bindings between specific character bounding boxes and uploaded references, directors must structure textual prompts to declare roles in exact alignment with the uploaded image sequence. While this positional notation provides strong semantic guidance to the diffusion model, it remains a heuristic rather than an absolute mathematical constraint.
Duration billing mathematics and whole-second trimming rules
Genjutsu billing: billed seconds equal the source length rounded up to a whole second, times $0.318 at 480p, $0.681 at 720p or $1.632 at 1080p. A 5-second 720p clip costs $3.41 and a 10-second clip $6.81.
Budget discipline in generative video post-production requires precise mathematical monitoring of source file durations. Unlike text-to-video systems that generate fixed duration blocks, Genjutsu calculates invoice costs based entirely on source media length. Billed seconds equal the source length rounded up to the next integer value. At high-definition resolutions, these costs scale: rendering at 1080p bills at $1.632 per second, whereas 720p delivery balances commercial visual clarity with economical turnaround.
Trim to a whole second before you submit. A 5.04-second clip bills as 6 seconds, which is 20% extra. We trim with ffmpeg at -t 5.
Failing to trim clip tails creates unforced operational waste. When an NLE exports a sequence with trailing frames—yielding a file of 5.04 seconds—the rounding ceiling bills the job as 6 full seconds. On a 720p job, that fractional tail incurs an additional $0.681 charge, adding a 20% cost penalty for four hundredths of a second of video. In our production pipeline, all candidate clips pass through an automated ffmpeg script using the -t 5 flag to ensure clean integer frame boundaries prior to submission.
Source file constraints and API endpoint paths
Restyle needs a source of at least 4 seconds, since a shorter clip fails input prep after it has been accepted. Clips over 30 seconds use only the first 30 seconds, audio included, and the file must be 200 MiB or less.
Submitting source footage that violates API constraints causes pipeline failures that burn engineering time. If a production team submits a clip under 4 seconds to the restyle endpoint, the API initially returns an HTTP acceptance status, only to fail input prep during back-end video processing. Similarly, footage exceeding 30 seconds uses only the first 30 seconds, discarding subsequent action, and files larger than 200 MiB are rejected immediately.
Generative video editing rates and duration billing structures · measured 24 Sept 2026
Video Edit Rail
Resolution Tier
Billing Rate Formula
Example Render Cost
Genjutsu (Motion/Swap/Restyle)
480p
$0.318 per ceil(second)
Rate card base tier
Genjutsu (Motion/Swap/Restyle)
720p
$0.681 per ceil(second)
$3.41 for 5 s ($6.81 for 10 s)
Genjutsu (Motion/Swap/Restyle)
1080p
$1.632 per ceil(second)
Master delivery tier
Seedance 2.5 Video-Edit
720p
~$0.277 / sec (input + output)
$2.22 for 4 s ($12.20 for 22 s)
Clips must be trimmed to whole seconds to prevent ceiling rounding penalties.
The restyle endpoint lives under the higgsfield/ path only. The older higgsfiled/ spelling still answers for motion and swap, but restyle does not exist there.
Legacy endpoint typos represent a common trap for API developers. Early SDK releases routed calls through an erroneous URL path with an inverted letter. While legacy motion transfer and object swap endpoints continue to accept calls routed through the misspelt path for backwards compatibility, the restyle service was deployed strictly under the corrected higgsfield/ namespace. Routing a restyle payload to the old URL fails because restyle does not exist there.
Parameter boundaries and post-production export deliverables
There is no aspect, fps, duration or seed parameter. Framing comes from the source, and without a seed the output cannot be re-rendered. We keep the take we like.
Generative video restyling sacrifices fine parametric control in exchange for structural adherence to source footage. The Genjutsu API accepts no aspect ratio declarations, frame rate overrides, or random seed integer inputs. Video dimensions, temporal pacing, and spatial composition are inherited directly from the uploaded footage. Because seed stability is absent, identical inputs produce slight generative variations upon repeated runs. Production teams must review takes immediately and archive successful outputs to cold storage.
Restyle keeps the source audio, and there is no option to replace it.
Audio preservation is automatic and immutable. Any dialogue, foley, or background ambience present in the source file is multiplexed directly into the re-rendered MP4 stream. Studios wishing to introduce new voiceovers or soundtrack elements must strip and replace audio tracks in external editing software after downloading the master render.
Outputs can include an After Effects JSX file, an FBX for Blender and a point cloud, not only a flat mp4. The files are kept for about 7 days.
For complex visual effects compositing, Genjutsu provides auxiliary 3D scene data alongside the rendered video. The pipeline can emit an After Effects JSX script that reconstructs solved virtual camera tracking, an FBX file containing camera trajectories for Blender, and a point cloud representation of scene geometry. These supplementary files are retained on cloud servers for approximately 7 days before automated deletion, requiring rapid post-processing retrieval.
Text-driven Seedance 2.5 editing and pipeline validation
Seedance 2.5 video-edit is the words-only route for a change described in text. Billed at 720p it is about $0.277 per second of input and of output, so a 4-second fix costs about $2.22, while editing a whole 22-second segment costs about $12.20, more than a re-roll.
When a shot requires a modification specified in plain language rather than image references, Seedance 2.5 video-edit provides an alternative pathway. However, billing applies symmetrically to both input duration and output duration. While executing a quick 4-second adjustment costs a modest $2.22, processing an extended 22-second sequence reaches $12.20. Submitting long takes to Seedance video-edit quickly becomes more expensive than generating a brand-new shot from scratch.
Seedance video-edit re-synthesises the whole clip, so nothing is pixel-locked. We trim to the defective 3 to 5 seconds, edit, and splice back.
Because the diffusion model recalculates every pixel across the frame, non-target regions may suffer subtle drift during processing. Rather than running an entire timeline cut through the model, editors trim the timeline to the isolated 3 to 5 seconds containing the defect, process the edit, and splice the corrected clip back into the master timeline.
Untested end to end: identity hold on our talent and the output quality of a Genjutsu restyle. The first paid run settles both, and until then we quote the price and wait for a yes.
