JoyAI-Video-Edit, JD’s open 16B model that edits a video while the video is still playing, at about 30 frames a second on one Nvidia B200

2 months ago

Video editing has always meant handing over the clip and waiting. JD just released JoyAI Video Edit, and it changes a video while the video is still playing. The weights are open under Apache two point oh, free for anyone to download and run.

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00JoyAI-Video-Edit
  2. 0:15You write the instruction, the model edits the footage
  3. 0:35The edit starts before the video has an ending
  4. 0:55The model edits eight frames at a time, looking only at what already happened
  5. 1:25The model cleans up each frame in two passes instead of dozens
  6. 2:02JoyAI tops JD's one-minute editing test and runs at thirty frames a second
  7. 2:25JoyAI takes the vote against all four live editors
  8. 2:49The weights are open. The thirty frames a second still needs a data center chip.
Show transcript

JoyAI-Video-Edit

It edits the video while the video is still playing.
JD Apache-2.0 · ungated video in, video out 16B parameters

Video editing has always meant handing over the clip and waiting. JD just released JoyAI Video Edit, and it changes a video while the video is still playing. The weights are open under Apache two point oh, free for anyone to download and run.

You write the instruction, the model edits the footage

source
edited
“Transform the people, hairstyles, and interior into a British castle aristocratic style.”
source
edited
“Turn the video into a watercolor wash style.”
source
edited
“Make all dogs white, add colorful hats, and turn the sunglasses hot pink.”
the footage can be
a video file a live camera feed
the edits it covers
background change global style local add local change local remove
JD, JoyAI-Video-Edit README, Showcase table (GitHub).

JoyAI Video Edit takes two things: footage, and a written instruction. The model hands back the same footage with that change made. The footage can be a video file, or a live camera feed. The instruction is an ordinary sentence, like turn the video into a watercolor wash style. The model will swap a background, restyle a whole scene, add an object, or take one out.

The edit starts before the video has an ending

Live stream and edited stream, same clock
30 FPS · 720 × 1280 · one Nvidia B200
JD, JoyAI-Video-Edit project teaser (GitHub), the same figure as the technical report’s Figure 1.

A normal AI video editor reads the whole clip before it changes anything, because it looks at how a scene ends to decide how to handle the beginning. JoyAI Video Edit works only from frames that already went by, so the model can begin on a live feed the moment the first frames arrive. JoyAI runs at about thirty frames a second, the rate live video plays at, on one Nvidia B two hundred, a data center graphics chip.

The model edits eight frames at a time, looking only at what already happened

a live stream has no ending yet
the very first batchthe anchor
…
older batcheslet go
the last few batchesstill in memory
nowbeing edited
inside a batch: all eight frames together across batches: only what already happened kept: the recent few, plus the very first
JD, JoyAI-Video-Edit technical report, Figure 3.

How does a model edit a video it has not finished watching? JoyAI Video Edit works in batches of eight frames. Inside a batch, all eight frames get edited together, so the change is consistent across them. From one batch to the next, the model looks only backward. A live stream can run forever, and holding every past batch would eventually fill the chip’s memory. So the model keeps only the last few batches, plus the very first one. That first batch is the anchor. It stops the edit from sliding away from what the scene looked like when it started.

The model cleans up each frame in two passes instead of dozens

Every pass is checked against the matching frame of the original video.
one batch of eight frames, start to finish266 ms
edit, two passes · 185 ms
compress 22 ms edit, two passes 185 ms decompress 19 ms remember 31 ms prep 9 ms
8 frames per cycle → 30 frames a second
JD, JoyAI-Video-Edit technical report, Figure 5, one Nvidia B200.

Working in batches alone is still too slow. The usual way an AI makes a picture is to start from static and clean it up, over and over, dozens of passes for a single frame. That will never keep up with live video. So JD trained a fast copy of the model that does the whole job in two passes. JD also trained that copy to check its work against the matching frame of the original video, every batch. The check stops a long edit from drifting away from the source footage. Two passes is what makes the arithmetic work. One batch of eight frames takes two hundred and sixty-six milliseconds, and the two editing passes are a hundred and eighty-five of those milliseconds. Eight frames every two hundred and sixty-six milliseconds is thirty frames a second.

JoyAI tops JD's one-minute editing test and runs at thirty frames a second

JoyAI tops JD’s one-minute editing test and runs at thirty frames a second

LongV2VBench · 229 one-minute editing tasks · five kinds of edit
JD’s own benchmark · JD’s own scoring · JoyAI at 720 × 1280
overall editing score, 1 to 5
JoyAI-Video-Edit 3.30 30.19 FPS
XMax-X2.0 1.71 20.90 FPS
SANA-Streaming 1.64 14.51 FPS
LiveEdit 1.23 15.45 FPS
StreamDiffusionV2 1.21 18.07 FPS
JD, JoyAI-Video-Edit technical report, Table 3. Every baseline re-deployed by JD.

JD built its own test for one-minute clips: two hundred and twenty-nine editing tasks across five kinds of edit. JD scored the results itself. JoyAI came first in all five kinds, with an overall score of three point three out of five. The next best live editor scored one point seven one. JoyAI was also the fastest, at thirty point one nine frames a second on video seven hundred and twenty pixels wide, forty to a hundred percent quicker than the other four.

JoyAI takes the vote against all four live editors

JD’s own study · pairwise · single-shot clips under 10 seconds · share of votes
JoyAI’s share of the votes against each live editor
LiveEdit90%
SANA-Streaming87%
StreamDiffusionV287%
XMax-X2.081%
against editors that get the whole clip first
Bernini-R JoyAI48% Bernini-R44% draw8% tie
Seedance 2.0 JoyAI35% Seedance 2.056% JoyAI loses9% tie
Kling-3.0 Omni JoyAI34% Kling-3.0 Omni56% JoyAI loses10% tie
Level with the best open-source offline editor, behind Kling-3.0 Omni and Seedance 2.0.
JD, JoyAI-Video-Edit technical report, Figure 9 and Section 7.3.

JD also asked people to pick between pairs of clips. Against the four other live editors, people picked JoyAI eight or nine times out of ten. Then JD ran JoyAI against editors that get the whole clip first. Against the best open-source one, JoyAI draws, forty-eight percent of the votes against forty-four. Against Kling three point oh Omni and Seedance two point oh, JoyAI loses, about thirty-five percent of the votes against fifty-six. Live editing has caught the open field. Those two still win.

The weights are open. The thirty frames a second still needs a data center chip.

weightshuggingface.co/jdopensource/JoyAI-Video-Edit
codegithub.com/jd-opensource/JoyAI-Video-Edit
Apache-2.0 · ungated

So editing a video no longer has to mean handing it over and waiting. Now editing can happen while the video plays, because the model works eight frames at a time, looks only backward, and cleans up each batch in two passes. Today that takes one data center chip. A graphics card a person could buy is on JD’s own to-do list, still unchecked.

Up next

2:58

Apodex 1.1 mini · the same file, run alone and run as a team

@ai-ml-today · 1 month ago

3:09

Ornith-1.5 · an open model trained on problems it wrote for itself

@ai-ml-today · 1 month ago

3:08

MOSS-VL · an open 11B video model trained to decide, after every frame, whether to speak or stay silent

@ai-ml-today · 1 month ago

3:29

Qwen3.8-27B · a 27B open model anyone can download averages 84.3 across real desktop jobs on Qwen’s own run, and loses the widest row on its own card by nine points

@ai-ml-today · 1 month ago

3:15

MiLMMT-46-v1.0 · Xiaomi finished training an open translator on sentences that had no translation attached, and the scores it now leads on are produced by the two models that graded it

@ai-ml-today · 1 month ago

3:31

BDH-CQ — a 150M-parameter system that learns a puzzle rule from the worked examples in front of it and does its thinking as numbers, never as words

@ai-ml-today · 2 months ago

3:29

Muse Glimmer 30B, Meta’s open agent model, and the second tiny model that guesses sixteen words ahead so it runs fast on one consumer graphics card

@ai-ml-today · 2 months ago

3:26

Inkling-Small, Thinking Machines’ 12B-active / 276B open mixture-of-experts model that out-reasons its own 3.5× bigger flagship

@ai-ml-today · 2 months ago

3:42

Maple-Preview, DeepGrove’s 20-billion-parameter ternary reasoning model that runs on a Mac mini

@ai-ml-today · 2 months ago

3:16

Hunyuan3D-Buffalo, Tencent’s one model that builds, understands, and edits 3D from words

@ai-ml-today · 2 months ago

3:28

MiniMax H3, an open model that makes a video and its soundtrack in one pass

@ai-ml-today · 2 months ago

3:37

DiffusionGemma, Google’s open model that writes text a whole block at a time

@ai-ml-today · 2 months ago