Research video generation framework that keeps scenes consistent after edits using RGB and depth context memory banks.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
PermaVid is a research framework for consistent video generation under editing operations. When an edit changes a scene's appearance or layout, earlier stored context can become outdated. PermaVid addresses this with a multi-modal context memory that separates semantic appearance from geometric structure, plus an edit-aware memory update and retrieval strategy.
The method uses two memory banks: an RGB context memory that captures appearance-aware observations while implicitly encoding geometry, and a depth context memory that preserves geometry-only structure. A memory-guided video generation model fuses features from these mixed-modality references. The project page shows global edits (stable geometry, propagated appearance) and local edits (edited content recalled, unchanged geometry preserved), and the authors report better long-term semantic and structural consistency than state-of-the-art methods.
According to a YouTube creator summary, code has been released (about 29 GB), built on Wan 2.1 and VACE, along with a roughly 400 GB dataset. This was not confirmed on the fetched pages, and no repository, weights or license links were available. It is a research project, not a hosted product.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Video editing AI that uses two memory banks, one for appearance and one for 3D structure, to keep edits consistent over time. Code is released, about 29 GB, built on Wan 2.1 and VACE, with a 400 GB dataset.
Closest neighbours by meaning, not just by category.
Model
Research model for video relighting that jointly generates relit video and per-frame environment maps, no camera poses needed.
Model · Nvidia
NVIDIA research model that turns a single image into a queryable 3D Gaussian scene with semantics and object instances.
Model · Netflix (Eyeline Labs)
Netflix Eyeline research model that restyles video from an edited keyframe while preserving face identity and performance.
Model
Research framework for real-time, causal streaming video editing with diffusion, running at about 12.66 FPS.
Research
Research framework for long, consistent video generation using a persistent 3D world state as memory.
Model · SenseTime
Open-weights 8B image generation and editing model from SenseTime that works in pixel space without a VAE.
