Wan2.1-VACE-14B
Alibaba Cloud · China · 2025
Most open video models only make new footage. VACE also changes footage you already have - swap a subject, repaint a masked region, keep the rest of the shot - and it does all of it from one set of weights.
Wan2.1-VACE-14B, released on 14 May 2025 under Apache 2.0, is the editing member of the Wan 2.1 family. The other models in that generation start from nothing: text or a still image goes in, new footage comes out. VACE starts from footage that already exists and changes it, which is a different and harder problem. It covers three jobs from one set of weights. Reference-to-video takes a picture of a subject and generates a clip that keeps that subject consistent. Video-to-video editing rewrites an existing clip to a new instruction. Masked video-to-video editing does the same inside a region the user paints, leaving the rest of the frame alone. Before VACE these were usually three separate models, each trained for its own task, and combining them meant chaining tools and accepting the quality loss at every handover. The mechanism that unifies them is worth naming because it explains the model's flexibility. Alibaba's research group defined a single input format, the Video Condition Unit, into which editing targets, reference images and masks are all packed; a Context Adapter then injects the task concept into the generator as a formal statement about time and space. The practical consequence is that tasks compose - a reference subject and a mask and an edit instruction can be combined in one pass rather than in three - and the paper reports that the unified model matches task-specific models on their own ground. The 14B build supports both 480P and 720P, unlike the 1.3B sibling, which Alibaba recommends keeping at 480P because its training at the higher resolution was limited. The download is heavy: seven weight shards totalling about 63 GB, plus the shared umt5-xxl text encoder at 11.4 GB and the Wan video autoencoder. Diffusers conversions of both sizes followed on 4 June 2025. One small point of catalogue interest: this repository contains the LICENSE.txt file that its model card links to. Every repository of the following generation, Wan 2.2, links to the same file and does not include it - so the paperwork on this older model is in better order than on its successors. VACE remains one of the most downloaded models in the Wan line, well ahead of the video editors published since, and no editing model of this kind has appeared in the generations Alibaba has kept closed.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!