Update README.md

README.md

@@ -2,59 +2,58 @@
 license: other
 library_name: transformers
 base_model: Qwen/Qwen3.5-9B
-pipeline_tag: text-generation
+pipeline_tag: any-to-any
 tags:
-  - WeVisCut
   - video-editing
-  - long-form-video
   - agent
   - tool-use
-  - speech
+  - long-horizon
   - reinforcement-learning
   - qwen3.5
 ---
 
-# WeVisCut
+# WeVisCut: A Long-Horizon Tool Agent for Edited Videos
 
-WeVisCut is a 9B tool-using policy for long-form video production from a narration script. It plans and edits through a registered runtime: selecting tools, using their outputs, inspecting intermediate results, and repairing failures. It is not a standalone text-to-video generator.
+**WeVisCut** is a 9B tool-using agent that turns knowledge articles or narration scripts into multi-minute edited videos. At each step, it reads the current editing state and selects one of 12 tools for planning, narration, asset retrieval, composition, inspection, or repair. It is not a standalone text-to-video model; use it with the [WeVisCut runtime](https://github.com/Tencent/WeVisCut), which executes the model's tool calls.
 
-**Base model:** `Qwen/Qwen3.5-9B`  
-**Model:** [Hugging Face](https://huggingface.co/tencent/WeVisCut)  
-**Code:** [GitHub](https://github.com/Tencent/WeVisCut)  
-**Project page:** [WeVisCut](https://xd-mu.github.io/WeVisCut/)
+## Deploy
 
-## Training
+Requirements: Python 3.11 or later and an NVIDIA GPU with about 24 GB of memory for bf16 inference. Install a PyTorch build compatible with your CUDA version before the GPU requirements below. The setup script installs Node.js, FFmpeg, and the renderer.
 
-The policy uses shared parameters across three roles: **Orchestrator**, **Generator**, and **Critic**. The paper describes a three-stage post-training process:
-
-1. Supervised fine-tuning (SFT) and runtime-verified rejection-sampled fine-tuning (RFT).
-2. Guided on-policy distillation (Guided-OPD), with teacher interventions gradually reduced.
-3. Role- and context-factored DAPO (RCF-DAPO), using execution feedback to optimize decisions.
-
-The paper constructs executable trajectories from 3,750 completed narration-driven video projects and reports 134,771 training examples. The source videos are described as spanning technology and digital products, business and finance, natural science, and social sciences and humanities. These figures describe the paper's data construction; they do not imply that the source videos or dataset are available with this checkpoint.
+```bash
+git clone https://github.com/Tencent/WeVisCut.git
+cd WeVisCut
+pip install -r requirements.txt
+pip install -r requirements-gpu.txt
+bash scripts/setup_external_tools.sh
+pip install -U "huggingface_hub[cli]"
+hf download tencent/WeVisCut --local-dir checkpoints/weviscut-9b
+bash run_studio.sh
+```
 
-## Evaluation
+Open `http://<host>:8081/` to use the Studio. For UI-only mode without loading model weights, run `bash run_studio.sh --no-model`.
 
-The paper evaluates on 100 held-out script-to-video tasks, with five independent runs in the shared-tool evaluation. It reports five metrics: Script Content Fidelity (SCF), Visual-Narration Semantic Match (VNSM), Timeline Plan Realization (TPR), Long-Horizon Consistency (LHC), and End-to-End Rendering Yield (ERY).
+## Model
 
-In the common-harness comparison, the paper reports **2,377 generated tokens per successful output minute** for WeVisCut and **78,763** for Claude Opus 5. This is a policy-generation cost measure, not end-to-end runtime or hardware throughput. The paper's Overall Score is a descriptive mean of the five metrics, not an independently validated sixth metric; the reported results should not be interpreted as evidence of statistical equivalence.
+The policy is based on Qwen3.5-9B and is trained on executable editing trajectories. Its runtime provides an editable timeline with pause/resume, manual and chat-based revisions, per-shot rerendering, and version rollback. Video quality depends on the configured retrieval, TTS, and rendering services and available templates.
 
-## Intended use and limitations
+## Intended Use and Limitations
 
-Use WeVisCut for research on narration-driven editing workflows in a compatible tool runtime. The policy depends on registered tools, available media assets, and the runtime's state and validation contract; the checkpoint alone does not retrieve assets or render a complete video. Results may not transfer to unregistered runtimes, other asset distributions, or visual styles outside the current template library. Visual-narration evaluation also relies on a vision-language model and may be affected by evaluator bias or service changes.
+WeVisCut is intended for research and production assistance in long-form video editing. It does not guarantee factual accuracy or production quality; review generated narration and video before publication. The model weights require the WeVisCut tool environment to produce a video.
 
 ## Citation
 
-If you use this work, please cite the paper:
-
 ```bibtex
-@article{mu2026weviscut,
-  title  = {{WeVisCut}: The Whole Editing Room in One Agent --- From Narration to Long-Form Video},
-  author = {Mu, Chenyu and Weng, Wenming and Wang, Qinrui and Ye, Xingsong and Yu, Hao and He, Jun and Sun, Chong and Li, Chen and Lyu, Jing and Yang, Erkun and Deng, Cheng},
-  year   = {2026}
+@misc{weviscut2026,
+  title        = {WeVisCut: A Long-Horizon Tool Agent for Turning Knowledge Articles into Edited Videos},
+  author       = {WeVisCut Team},
+  organization = {Tencent},
+  year         = {2026},
+  howpublished = {\url{https://xd-mu.github.io/WeVisCut/}},
+  note         = {Technical report and code release}
 }
 ```
 
 ## License
 
-The model card metadata currently uses `license: other`. Check the model repository and the licenses of the base model, tools, and media assets before use; do not assume they share the same license.
+The WeVisCut code is released under the [Apache License 2.0](https://github.com/Tencent/WeVisCut/blob/main/LICENSE). Model weights and third-party components are subject to their respective license terms.