CoinVE-Edit: Multi-Instruction Video Editing Built on Wan2.1

Tencent's Smart Creation Platform team released CoinVE-Edit on August 18, a compositional instruction-guided video editing model that executes multiple editing instructions in a single forward pass. Built on the Wan2.1-T2V-14B video DiT and a Qwen3-VL-8B multimodal encoder, the 22B model ships together with the CoinVE-200K training dataset and the CoinVE-Bench evaluation benchmark.

CoinVE-Edit framework: an MLLM encodes each instruction with the source video, a mask head predicts per-instruction regions, and a residual-attention module steers the Wan2.1 DiT.

Multiple Edits in a Single Pass

Most open-source video editing models handle one instruction at a time. CoinVE-Edit accepts 2 to 5 editing instructions for the same clip and applies them simultaneously, with each edit confined to its designated region:

• Region-aware mask guidance — a lightweight mask head predicts a per-instruction spatial mask from MLLM visual tokens, keeping each edit inside the correct area
• Residual attention — a residual-attention module injects the per-instruction mask guidance into the DiT attention layers, so edits land in their own regions while the rest of the frame stays intact
• Compositional operations — Replace, Add, Remove, and Background Change can be combined freely in one pass

The CoinVE-200K Dataset

CoinVE-200K is a large-scale dataset of 200K+ 1080p video-edit pairs (up to 201 frames per clip), built through an automated data generation and quality filtering pipeline. Each sample carries 2 to 5 atomic editing operations across humans, objects, and backgrounds, with per-instruction and combined region masks. Existing datasets such as ReCo-Data and OpenVE-3M focus on single-instruction editing, which is what limits trained models in real-world multi-edit scenarios.

CoinVE-200K vs. existing video editing datasets: compositional, multi-instruction samples with per-instruction and combined masks.

Benchmark: CoinVE-Bench

CoinVE-Bench provides 361 multi-instruction test cases scored by Gemini 3.6 Flash. CoinVE-Edit leads the edit-accuracy metrics (SA / SPA / EP) and motion naturalness (MN) among open and commercial systems:

Availability

CoinVE-Edit does not have native ComfyUI support yet; it runs through the official Python inference pipeline built on DiffSynth-Studio. The checkpoint bundles the DiT LoRA (rank 128), MLLM LoRA (rank 256), learned image/video query embeddings, connector, VAE condition encoder, and mask head, and must be loaded on top of the Wan2.1-T2V-14B and Qwen3-VL-8B-Instruct base models. A Hugging Face demo space is available for trying the model.

The technical report and the project page cover the architecture, dataset pipeline, and benchmark in detail; inference scripts are in the CoinVE-200K GitHub repository.