手ブレ補正を全クリップに掛けたら、かえって酔う動画になった — clipwright v0.24〜v0.26 手ブレ補正を全クリップに掛けたら、かえって酔う動画になった — clipwright v0.24〜v0.26

clipwright, a Python video editing toolkit distributed on PyPI and developed in the open on GitHub, has moved through a sequence of releases from v0.24 to v0.26.0. The project sits in the Model Context Protocol (MCP) ecosystem, where it appears to expose editing operations as callable tools so that a language-model agent can assemble, trim, and adjust footage programmatically rather than through a manual timeline. The most instructive takeaway from this release cycle is a cautionary one: applying image stabilization uniformly to every clip can make a finished video harder to watch, not easier.

The problem described is counterintuitive but well understood in video engineering. Digital stabilization works by estimating camera motion frame to frame and then cropping, shifting, and warping each frame to cancel that motion. When this is done indiscriminately across an entire project, several side effects compound. Clips that were already steady get subtly warped for no benefit, footage with rolling-shutter distortion can develop a wobbling "jello" appearance, and the algorithm may leave behind slow residual drift that the original shake actually masked. Stitched together, clips with different motion characteristics produce inconsistent apparent movement, and that mismatch between what the eyes see and what the inner ear expects is a common trigger for motion discomfort. In other words, removing high-frequency shake while introducing low-frequency, unnatural motion can increase the sense of nausea.

The lesson that emerges is that stabilization is better treated as a per-clip, opt-in decision than a global default. Shaky handheld shots benefit from correction, but tripod shots, deliberate camera moves, and footage with strong parallax often look worse after processing. A more robust workflow analyzes each clip's motion profile first and applies smoothing only where the measured shake exceeds a threshold, ideally with a configurable crop margin so the tool does not zoom in further than necessary. This kind of conditional logic is exactly the sort of policy that an MCP-driven pipeline is positioned to encode, since the agent can inspect metadata and decide which operations to invoke per segment.

Across v0.24 through v0.26, the practical implication is that defaults and parameter handling matter as much as the underlying feature set. A tool that makes stabilization easy to apply to everything invites the exact failure mode the title describes. Exposing parameters such as smoothing strength, crop limits, and a per-clip enable flag, and surfacing them clearly through both the command line and the MCP tool schema, helps callers avoid blanket application. It is worth noting that the specific changelog entries for each version are best confirmed directly on the project's GitHub and PyPI pages, since the precise scope of each release is not fully detailed here.

For context, clipwright operates in the same territory as established media tooling. FFmpeg remains the foundational engine for most programmatic video work, and its filters such as vidstabdetect and vidstabtransform, along with the simpler deshake filter, perform two-pass stabilization in much the way described above. Higher-level libraries like MoviePy and PyAV wrap encoding and editing for Python developers, while desktop editors such as DaVinci Resolve and Adobe Premiere Pro offer warp-stabilizer features with manual override. What distinguishes an MCP-oriented project is the interface rather than the algorithms: by presenting editing actions as structured tools, it lets an assistant orchestrate a multi-step edit while a human reviews the result.

The broader industry move here is the steady expansion of MCP beyond text and data tasks into media manipulation. As more creative tools publish MCP servers, agents gain the ability to chain operations like ingesting clips, normalizing audio, color adjustments, and rendering output. That capability raises the stakes for sensible defaults, because an agent will faithfully apply whatever instruction it is given, including over-correcting an entire timeline. The clipwright experience is a small but concrete reminder that automation amplifies both good and bad editing choices, and that perceptual quality, not just successful processing, should guide how these tools are designed. Readers interested in the exact behavior should consult the repository and test stabilization on representative clips before enabling it project-wide.