VideoX-Qwen: Data-Centric Instruction-Based Video Editing
JJiahang Li, Dingbao Shao, Xinyu Chen, Song Wu, Jiang Lin, Duo Li, Yuhang Liu, Jiaxin Hu, et al.
cs.AI · cs.GR
Quick verdict
Claims a 1.2M-record data pipeline plus Qwen-Wan editor winning 9/11 metrics; solid engineering, but small eval, judge bias, no ablations temper the claim.
Summary
The paper presents VideoX-Qwen, an integrated framework for instruction-based video editing combining a large-scale data pipeline with a unified model. The pipeline routes source videos selected by Qwen3-VL-235B into two synthesis paths — SAM3 masking plus Minimax-Remover for addition/removal, and Qwen-Image-Edit first-frame editing plus Wan-Animate propagation for replacement/attribute edits — followed by whole-video quality filtering and instruction enrichment, yielding over 1.2 million directional (source, instruction, target) records (>400K per task group) at an 89% automatic acceptance rate. The editor couples a LoRA-adapted Qwen MLLM with learnable queries for semantic conditioning and channel-concatenated frozen source-video latents injected into a Wan DiT trained with flow matching. A three-stage progressive recipe (21K/27K/500 steps) aligns the semantic interface, enables source-conditioned editing, then refines at 720-1280 resolution on 30K selected videos. On a 100-example benchmark against UniVideo and Kling O1, it attains the best mean on nine of eleven metrics (e.g., SSIM 0.7658 vs 0.5723/0.6701), while Kling O1 leads aesthetic and imaging quality. Limitations: no human evaluation or significance testing, the judge (Qwen2.5-VL-7B) shares lineage with pipeline models, possible reference-target circularity for SSIM/LPIPS, no ablations, and no code or data release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The two-route production pipeline constructs more than 1.2 million directional video-editing records, over 400,000 in each of the three task groups, with an 89% automatic acceptance rate.
- A unified editor combining LoRA-adapted Qwen multimodal semantic conditioning with channel-concatenated frozen source-video latents in a Wan DiT performs instruction-only editing at inference, requiring no masks, edited first frames, or expert generators.
- Three-stage progressive image-video training converts the corpus into an editor achieving the best mean on nine of eleven metrics against UniVideo and Kling O1 on 100 examples.
Strengths
Unified two-route data pipeline with directional pair reuse
Section 3.2 (Eq. 2) stores each accepted physical add/remove pair as two directional records, and all routes normalize into one (S, T, Y) interface (Section 3), enabling a single general editor rather than per-task models.
Unusually transparent training recipe
Table 2 gives per-stage records (approx. 3.0M / 2.6M / 30K), steps (21,000 / 27,000 / 500), batch 64, LR 1e-5, EMA 0.9995, and resolutions (480-832 then 720-1280); Figure 3 maps trainable components per stage.
Whole-video quality filtering, not first-frame only
Honest reporting of lost metrics
Weaknesses
No ablation studies anywhere
Small evaluation with tiny margins and no significance testing
Judge-family bias and unstated reference-target provenance
Thin baseline coverage
No code, data, or weight release; unverified enriched instructions
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
VideoX-Qwen (arXiv 2609.26015, Nanjing University with Jiutian Research) is a system paper coupling data production with model training for instruction-only video editing. It claims three things: (1) a two-route pipeline yielding over 1.2M directional (source, instruction, target) records, more than 400K each for addition, removal, and replacement/attribute editing, at an 89% automatic acceptance rate (Section 3.5, Table 1); (2) a Qwen-Wan editor needing only source video plus instruction at inference; (3) the best mean on nine of eleven metrics against UniVideo and Kling O1 on 100 examples (Table 3). The abstract's framing is data-centric; all evidence offered is one automatic benchmark.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is VideoX-Qwen?
VideoX-Qwen is a research system from Nanjing University and Jiutian Research for instruction-based video editing. It combines an automated pipeline that builds over 1.2 million paired training videos (source video, edit instruction, edited target) with a single editing model built on a Qwen multimodal encoder and the Wan video generator. At inference the model needs only the source video and a text instruction; no masks or edited first frames.
How big is the VideoX-Qwen dataset and how was it built?
The paper reports more than 1.2 million directional editing records, over 400,000 in each of three groups (addition, removal, replacement and attribute editing). It combines SAM3 segmentation with Minimax-Remover for add/remove edits, and Qwen-Image-Edit first-frame editing with Wan-Animate propagation for replacements and attributes, then filters whole videos with Qwen3-VL-235B. The automatic acceptance rate is stated as 89%.
How does VideoX-Qwen compare with UniVideo and Kling O1?
On a 100-example benchmark using eleven automatic metrics, VideoX-Qwen reports the best mean score on nine metrics, including SSIM 0.7658, LPIPS 0.1966, and instruction following 7.70 on a 10-point scale. Kling O1 scores higher on aesthetic quality (0.5248 vs 0.5067) and imaging quality (0.6973 vs 0.6294). No human evaluation or statistical significance tests are reported in the paper.
Is VideoX-Qwen code, data, or a model available?
The paper text does not state any release of code, dataset, or model weights. The pipeline also depends on several third-party models (SAM3, Minimax-Remover, Qwen-Image-Edit, Wan-Animate, Qwen3-VL-235B), so independent reproduction would currently require rebuilding the entire data and training stack.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 could not be inspected; architecture and qualitative claims are taken from captions and body text only.
- The extracted text has minor spacing artifacts (e.g., VideoX-Qwencombines) and the abstract is truncated mid-sentence; tables 1-3 appear intact.
- Evaluation-target provenance, inference hyperparameters, and compute figures are absent from the text, limiting verification of headline claims.
- arXiv v1 preprint (22 Sep 2026), apparently not yet peer-reviewed; cited baselines were not independently checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.