VIVID lifts Codex's video-editing score by 20.2 points
The visual harness gives coding agents video inspection and transcription without retraining, and its authors' benchmark scores still sit far below human editors.

Key takeaways
- VIVID is a visual harness that lets Codex, Claude Code and OpenCode work on video post-production without additional model training, per a paper submitted to arXiv on November 16, 2026.
- On AgenticVBench, a 100-task video post-production benchmark, Codex with VIVID averaged 58.6% against 38.4% for native Codex, a 20.2 percentage-point gain the authors report.
- The same harness lifted Claude Code from 32.4% to 55.1% and OpenCode from 27.4% to 37.1%, while human experts score 88.5% on the same tasks.
- Ablation runs show each component matters: removing VIVID's on-demand video inspection drops the Codex average by 16.0 percentage points.
VIVID is a harness that gives general-purpose coding agents the piece they were missing on video work: a structured way to look at the footage. The paper behind it, submitted to arXiv on November 16, 2026, reports that attaching VIVID to Codex lifted its average score on a 100-task video post-production benchmark by 20.2 percentage points, with no additional model training.
The gap it targets is documented. That benchmark, AgenticVBench, came out in May 2026 from the research group Philo Labs, and its headline finding was blunt: the best agent stack barely crossed 38% while human experts scored 88.5% on the same tasks. Video editing asks an agent to understand source material, find the right moments, and judge whether a rendered file actually works. Coding agents were good at the code half and bad at the looking half.
VIVID is an attempt to close that half.
What VIVID actually is
VIVID is a visual harness for agentic video editing, from Wenhan Yu, Zhaoxi Zhang, Zhilong Tan, Weikang Li and Lei Sha, at Beihang University and Peking University. The code is on GitHub under an MIT license.
It is not a model and it is not trained. VIVID extends coding agents that already exist. The paper's three evaluated configurations are GPT-5.6-sol on Codex, Claude Fable 5 on Claude Code, and GPT-5.5 on OpenCode. In each case the agent keeps its native shell, file and code-execution tools, and VIVID adds media tools on top.
The harness exposes six of them: catalog, inspect, transcribe, retrieve, edit and verify. inspect pulls frames from a time range or a spatial region and compares assets. transcribe returns segment-level or word-level text. retrieve searches saved observations. verify checks a rendered file's format and decoding status.
How the visual loop works
The design turns on one idea: an observation is useful only if you can find it again and know what it came from. VIVID stores every observation outside the model's context and tags it with a media version, a time coordinate and an evidence ID. If a file is rewritten at the same path, the older observations stay linked to the version they came from, so an agent does not confuse a frame from the current cut with a frame from an earlier one.
That supports a cycle the paper calls observe, reason, edit, render, review. An agent gets a sparse overview of the material, inspects candidate intervals in more detail, makes edits with its own tools, renders a file, then looks at the output and revises. The harness's own prompt reminds the agent that sampled frames are not continuous coverage, that a transcript is not proof of listening, and that passing technical checks does not establish content quality.
The case studies show the loop in motion. On a teaser task built from the film In the Mood for Love, Codex with VIVID sampled a 24-frame overview of a 5,640-second source, requested word-level transcription for chosen segments, assembled an 82.17-second montage and scored 82 out of 93 on the task's rubric. On a speech-highlight task, native Codex produced output the evaluator scored 0.21, with aspect-ratio distortion and a missing caption, while Codex with VIVID scored 1.00.
What the benchmark shows
AgenticVBench holds 100 expert-authored tasks across four families: Assembly (18), Repair (18), Sequencing (28) and Repurpose (36). Twenty industry experts averaging six years of experience wrote the tasks and their grading rubrics. The human reference score is 88.5%.
The paper runs each task three times per configuration, in a Modal sandbox with 4 vCPUs and 8 GB of RAM, a 30-minute wall-clock limit and a 200-iteration cap. Its reported averages:
- Codex: 38.4% native, 58.6% with VIVID, a gain of 20.2 percentage points.
- Claude Code: 32.4% native, 55.1% with VIVID, up 22.7 points.
- OpenCode: 27.4% native, 37.1% with VIVID, up 9.7 points.
Assembly scores highest for all three. Repair barely moves: Codex goes from 39% to 43%, and OpenCode does not change at all.
An ablation isolates the three components. Removing on-demand inspection drops the Codex average to 42.6, a 16.0-point fall. Removing evidence memory drops it to 47.8. Removing output review drops it to 44.0. The largest single effect is losing the ability to look.
The gains are not free. Codex uses 17.9% more tokens with VIVID, Claude Code 10.7% more, and OpenCode 53.3% more.
What VIVID does not prove
The evaluation is the authors' own. It is an arXiv preprint, not a peer-reviewed result, and no independent lab has replicated the numbers. The benchmark it uses comes from a different organization, Philo Labs, which is a point in its favor, but the harness is being tested by the people who built it.
The absolute numbers stay modest. Even the best VIVID configuration, 58.6%, leaves nearly 30 points between it and the 88.5% human reference. The paper states that "global narrative, audiovisual coordination, and repair quality remain areas for further improvement."
Repair is the clearest limit. Its gains are small, and the paper ties that to the need for methods VIVID does not supply, such as denoising, detail restoration and color correction. Seeing a defect and fixing it are different problems.
The scoring deserves a caveat too. Assembly, Repair and Sequencing use deterministic verifiers, but Repurpose is graded by a model aligned to expert judgment, and the benchmark team has not released that grader. A 58.6% average is partly the product of an automated judge.
This is also research code, not a product. The repository had no releases and zero stars when it was read, and full managed rendering requires Linux. The trade report this story is based on, from VideoTech Wire, drew on a single source document, so every claim here traces back to the paper itself.
What it means for a post house
The useful result may not be the score jump but the harness finding underneath it. On the same benchmark, holding a model fixed and changing only the scaffolding around it swung Assembly scores by as much as 20 points, the benchmark site reports. If that holds, the wrapper matters as much as the model, and a post house that builds a good tool loop gains more than one waiting for a better model.
VIVID itself is cheap to try and unfinished. It is MIT-licensed, needs Python 3.11 or newer and FFmpeg, and installs as a skill for Codex or Claude Code. The token overhead means it costs more per task than a bare agent; the paper estimates roughly $1.32 per run on the non-Repurpose tasks with Codex.
None of this replaces an editor. The paper is explicit that a well-reasoned edit plan does not always produce a correct output, which is the whole reason the review step exists. What VIVID adds is a way for an agent to notice the difference between the two.
The code is at github.com/yuwenhan07/VIVID under the MIT license; install it with bash skills/install.sh codex or bash skills/install.sh claude, then invoke it as $vivid or /vivid.
Sources
- videotechwire.com - the trade report this story is based on, itself drawn from one source document
- arxiv.org - the VIVID paper, submitted 16 November 2026
- arxiv.org - full text: results table, ablation, tool list, run settings, cost
- github.com - official MIT-licensed code repo, install steps, results
- arxiv.org - the AgenticVBench paper (the benchmark VIVID is tested on)
- agenticvbench.com - benchmark site: task families, human reference, the harness finding
- agenticvbench.com - leaderboard numbers for the native baselines
- philolabs.ai - benchmark maker's essay on verifiers and the human score