VACE is an open implementation from Alibaba Group’s Tongyi Lab that brings several video-generation and editing tasks into one model family. Its official code and model releases cover reference-to-video generation, video-to-video editing and masked video-to-video editing, with workflows that can combine those controls.
That is a more precise description than calling VACE a free replacement for a commercial video editor. It is a research-oriented model and toolchain for developers and technical creators. The official VACE repository includes inference code, preprocessing tools and Gradio demos, while the peer-reviewed ICCV 2025 paper explains the unified architecture.
What VACE actually unifies
VACE organizes different source materials—such as a reference image, an existing video and a mask—through a shared Video Condition Unit. A Context Adapter then injects the requested task into the model. In practice, the framework supports three broad modes:
- Reference-to-video (R2V): generate a video from a text prompt and reference material.
- Video-to-video (V2V): transform an input video globally using controls such as depth, pose or structure.
- Masked video-to-video (MV2V): modify a selected region for jobs such as inpainting, subject replacement or canvas extension.
The modes can be composed. The project presents examples such as animating a reference, replacing a subject inside a mask, extending a scene and continuing an earlier clip. These are model capabilities demonstrated by the authors, not a promise that every source clip will produce production-ready results without iteration.
Which VACE models are available
The official repository currently lists four releases built on Wan 2.1 or LTX-Video:
- VACE-Wan2.1-1.3B-Preview: an early Wan-based model targeting roughly 81 frames at 480 × 832.
- VACE-LTX-Video-0.9: an LTX-based release targeting roughly 97 frames at 512 × 768.
- Wan2.1-VACE-1.3B: the released 1.3-billion-parameter Wan variant, targeting roughly 81 frames at 480 × 832.
- Wan2.1-VACE-14B: the larger Wan variant, targeting roughly 81 frames at 720 × 1280.
Those sizes are the project’s recommended operating ranges, not universal input restrictions. The repository says inputs can use other resolutions, although results are expected to be strongest within the specified ranges.
“Open source” needs a licence check
The VACE code repository is published under Apache 2.0, but the downloadable models inherit their base-model licences. The project table identifies the Wan-based releases as Apache-2.0 and the LTX-Video release as RAIL-M. Anyone planning a commercial deployment should read the licence attached to the exact checkpoint rather than infer terms from the repository name.
This distinction is especially important for a pipeline that mixes VACE code, a base model, third-party preprocessors and community workflows. Each component can carry separate obligations.
Installation and hardware reality
VACE is not presented as a one-click consumer desktop app. The official environment was tested with Python 3.10.13, CUDA 12.4 and PyTorch 2.5.1 or newer. Users download a chosen base model, install the relevant Wan or LTX dependencies and prepare source material before inference.
The repository documents both single-GPU and multi-GPU commands. Its 14B 720p example uses an eight-process distributed launch, while the 1.3B model has a single-GPU example. That does not establish a universal memory requirement or render time: practical performance depends on the checkpoint, resolution, frame count, attention settings, precision and hardware. Claims that VACE always runs on a particular consumer GPU are therefore too broad unless they are tied to a reproducible configuration.
A realistic VACE workflow
- Define the task. Decide whether the job is generation, whole-frame transformation, masked editing or a composition of several controls.
- Prepare the inputs. Collect the source video, reference images, prompt and any mask or bounding-box information.
- Preprocess the material. VACE converts inputs into a source video, source mask and reference-image set expected by the model.
- Run inference. Choose the base family and checkpoint, then pass the prepared material and prompt to the appropriate pipeline.
- Inspect intermediate and final files. The scripts save processed inputs and model outputs, which makes it possible to diagnose a bad mask or mismatched prompt before another run.
The official VACE user guide recommends generating longer work in short clips and using the first-clip extension task to maintain continuity. It also warns that prompt extension can drift away from the source video during editing, because the extension system does not automatically understand every visual detail in an unknown input.
What the research does—and does not—prove
The ICCV paper reports that the unified framework was competitive with task-specific models across the authors’ evaluation set. That supports the architectural claim: one framework can handle a wide range of creation and editing conditions. It does not prove that VACE is best for every production, style or hardware budget.
The project’s own guide is appropriately cautious, describing the model as intended for experimental research validation and saying real-world performance is not guaranteed. Editors should expect to test masks, prompts and temporal consistency, then apply conventional review for visual artifacts, rights and disclosure.
The practical verdict
VACE is significant because it exposes a composable video-control framework, trained checkpoints and reproducible tooling in public. Its strongest audience is researchers, developers and technically comfortable creators who need more control than a prompt-only video generator provides.
For a studio choosing tools, the decision should rest on the exact model licence, available GPU capacity, control requirements and willingness to troubleshoot a research stack. Treat VACE as an open foundation for experimentation—not as proof that professional video editing has become automatic. Follow Kingy AI’s AI News for related model releases and the AI Launches tracker for new tools.
The Kingy Brief
Follow The Kingy Brief.
One consequential launch, one pricing, limit, or shutdown change, one hands-on test, one exact prompt or Test Pack, and one try / watch / skip verdict.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime