JoyAI-Video-Edit¶

Development line: project:joyai-video-edit · thread joyai-video-edit-public-reference
Last event: 2026-08-06 · 1 dated since 2026-08-06 · Researched: 2026-09-05 · confidence: high

What it is¶

JoyAI-Video-Edit: an Apache-2.0, 16B autoregressive-diffusion video editor for causal, live video-to-video edits.

• Text, local, background, style, motion and reference-image edits.
• Live-camera or uploaded-video input over WebSocket.
• Claimed benchmark: 720p at about 30 FPS on one Nvidia B200; the consumer-GPU configuration is 840×480 at 24 FPS on RTX 5090.

Development line¶

• 2026-08-06 — JoyAI-Video-Edit public project references recorded. On 2026-08-06, public links appeared for a GitHub repository, a Hugging Face page, and a JoyAI Labs site. The references show active work, but omit a specific release version, capability, or launch status.

What changed¶

• 2026-08-06 — Deployment code, checkpoints, and a technical report released; initial DiT checkpoint was joyai_video_edit_dit_0804.pth.
• 2026-08-14 — DiT checkpoint joyai_video_edit_dit_0811.pth replaced the initial checkpoint for serving and improved reference-image-guided editing.
• 2026-08-15 — A live Hugging Face demo became available on RTX PRO 6000 hardware.
• 2026-08-24 — Official consumer-GPU support added an RTX 5090 configuration at 840×480 and 24 FPS.

How to use this¶

As of 2026-08-06, check the public code, model-distribution, and official-web references to verify the exact release state and usage guidance before adopting JoyAI-Video-Edit.

1. Clone the repository, create a Python 3.10 Conda environment, then install deploy/requirements.txt. — https://github.com/jd-opensource/JoyAI-Video-Edit
2. Build the in-tree joyomni_ops CUDA extension; for Blackwell hardware use CUDA nvcc 12.8 or later. — https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/DEPLOYMENT.md
3. Download joyai_video_edit_dit_0811.pth, the VAE files and the required MiMo-VL encoder into the documented checkpoints tree. — https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/DEPLOYMENT.md
4. Set the GPU-specific environment variables and run bash deploy/run_server.sh; open port 8080 or forward it over SSH. — https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/DEPLOYMENT.md

Best practices¶

• Use the current 0811 DiT checkpoint; 0804 is retained only for reproducibility and is not the server default. — https://huggingface.co/jdopensource/JoyAI-Video-Edit
• Give each GPU model a separate JOYOMNI_CACHE_ROOT when sharing a checkout, and use the documented per-GPU resolution/FPS settings. — https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/DEPLOYMENT.md
• Use SageAttention only on RTX 5090; leave it disabled on RTX PRO 6000 and B200. Enable low-VRAM mode on cards with 48 GB or less. — https://github.com/jd-opensource/JoyAI-Video-Edit/blob/main/DEPLOYMENT.md

Superseded by this¶

• 2026-08-14 — joyai_video_edit_dit_0804.pth is superseded for serving by joyai_video_edit_dit_0811.pth; retain the former only for reproducibility.

Still unknown¶

• The original JoyAI Labs demo URL could not be retrieved in this research environment, so its current availability is unverified.
• The claimed throughput is first-party benchmark evidence, not an independently reproduced performance result.

Sources¶

Agent brief¶

• Subject: project:joyai-video-edit, thread joyai-video-edit-public-reference, 1 dated events 2026-08-06 → 2026-08-06.
• Practical note: As of 2026-08-06, practitioners should treat JoyAI-Video-Edit as having public code, model-distribution, and official-web references, then verify the exact release state and usage guidance at those sources before adopting it.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.