LiveEdit¶

Development line: project:liveedit · thread liveedit
Last event: 2026-06-30 · 1 dated since 2026-06-30 · Researched: 2026-09-06 · confidence: medium

What it is¶

LiveEdit is a text-instructed video-to-video editor for Linux and NVIDIA setups. It builds on Wan2.1-T2V-1.3B instead of running as a hosted service.

• Video-to-video editing from text prompts on an input stream.
• Causal chunk processing to keep background and unedited sections intact.
• Autoregressive mask caching for short clips and an extended path for long videos.

Development line¶

• 2026-06-30 — LiveEdit project resources were publicly linked. LiveEdit's official project page, code repository, and model checkpoints were already online. Author timelines place code and checkpoint guidance on 2026-06-24 and preprint v1 on 2026-06-25, so 2026-06-30 is not an author release date.

What changed¶

• 2026-06-30: LiveEdit's official project, source, and model endpoints were already public. The authors placed code and checkpoint instructions on 2026-06-24 and preprint v1 on 2026-06-25, so this is not an author release date.
• 2026-07-13: arXiv updated the paper to v2 without accompanying code or checkpoint releases.
• 2026-08-05: The authors documented long-video inference using rolling attention sinks and window-relative RoPE.

How to use this¶

1. On Linux with an NVIDIA GPU, create the documented Python 3.10 environment. Install the requirements and FlashAttention. Single-GPU inference works; training scripts require multi-GPU torchrun. — https://github.com/cp-cp/LiveEdit/blob/main/README.md
2. Download Wan-AI/Wan2.1-T2V-1.3B and the documented ar-forcing_002000.pt checkpoint into the paths expected by the repository. — https://huggingface.co/cp-cp/LiveEdit
3. Create a JSON file containing instruction and source_path. Run infer-local-ar-forcing.sh for the documented V2V path. — https://github.com/cp-cp/LiveEdit/blob/main/README.md
4. Run infer-token-pruning.sh with --save_mask to inspect the cache. Run infer-local-ar-forcing-long.sh for longer clips. — https://github.com/cp-cp/LiveEdit/blob/main/README.md

Best practices¶

• Use the documented 2,000-step ar-forcing_002000.pt checkpoint and its supplied script. Not every weight file in the repository has a standalone recipe. — https://github.com/cp-cp/LiveEdit/blob/main/README.md
• Keep SINK_SIZE smaller than LOCAL_ATTN_SIZE in long-video mode, and test at target clip length. The authors note output degrades past the short-sequence training horizon. — https://github.com/cp-cp/LiveEdit/blob/main/README.md
• Save mask visualizations during token pruning so we can inspect which regions the cache reuses. — https://github.com/cp-cp/LiveEdit/blob/main/README.md

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The 2026-06-30 entry supplies only its date and official URLs; the underlying claim is unavailable.
• No independent run has reproduced the reported 12.66 FPS or established a LiveEdit VRAM floor. That speed is an author-reported figure.
• The name LiveEdit is shared: qizhou000/LiveEdit is a separate CVPR 2025 vision-language model editing project. The cp-cp repository, Hugging Face checkpoint, project site, and arXiv preprint identify this video editing project.
• The 2026-06-24 and 2026-08-05 dates come from author README logs, not independent release tracking or third-party quality tests.

Sources¶

Agent brief¶

• Subject: project:liveedit, thread liveedit, 1 dated events 2026-06-30 → 2026-06-30.
• Practical note: Use the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above decide what is obsolete.