nextcf-anon-2026/nextcf

NExT-CF: Counterfactual Entity-Removal Benchmark for VideoQA

NExT-CF is a benchmark of 570 counterfactually edited videos derived from NExT-QA, used to probe whether Video Question Answering models rely on visual evidence or language priors. Each video has a target entity removed via diffusion inpainting (DiffuEraser), and the corresponding QA pairs are carried over from NExT-QA.

Tiered structure

• Tier 1 (direct): 389 edited videos in tier1_edited_videos/. Source clips are CC-BY, CC-BY-SA, CC-BY-NC, or CC-BY-NC-SA. Edited videos are released under CC-BY-NC-SA 4.0.
• Tier 2 (reproducible): 181 entries in reproduce_tier2/. Source clips are CC-BY-ND or CC-BY-NC-ND, whose "NoDerivatives" clause prohibits redistribution of modified versions. To stay license-compliant, we ship only the SAM3 prompt and QA pairs for these clips; run reproduce_tier2/reproduce.sh to regenerate the edits locally from the unmodified NExT-QA sources.

All 570 clips are included in the evaluation benchmark.

Contents

File	Description
metadata.json	All 570 entries — id, tier, prompt, license, QA
tier_assignment.csv	Per-clip tier + original license
tier1_edited_videos/	389 edited MP4s
reproduce_tier2/tier2_manifest.csv	181 ND clips, one row each
reproduce_tier2/reproduce.sh	Regeneration wrapper
ATTRIBUTION.csv	Flickr credits for all clips (mandatory under CC-BY)
LICENSE	CC-BY-NC-SA 4.0 (tier-1 edited videos)
LICENSE-METADATA	CC-BY-SA 4.0 (metadata, prompts, QA)
PROMPT_SELECTION.md	Selection criteria for SAM 3 prompts (stages 2–3 of the pipeline)

Attribution

Source clips are from YFCC100M via NExT-QA. Per-clip photographer credit is in ATTRIBUTION.csv. This dataset requires citing both:

• NExT-CF (this dataset) — see CITATION.cff
• NExT-QA: Xiao et al., "NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions", CVPR 2021.

Reproducing tier-2 edits

# 1. Download NExT-QA videos separately (license permits redistribution
#    of originals; see https://github.com/doc-doc/NExT-QA)
export NEXTQA_DIR=/path/to/NExTVideo

# 2. Clone the NExT-CF code repo
git clone https://github.com/<owner>/nextcf-code
export CODE_DIR=$(pwd)/nextcf-code

# 3. Run the reproduction script
bash reproduce_tier2/reproduce.sh

Takes roughly 15 GPU-hours for all 181 clips on an RTX 6000 (fp16).

Ethics and limitations

• Synthetic edits from diffusion models may contain artifacts.
• Source distribution inherits any biases of NExT-QA and YFCC100M.
• This dataset is research-only (NC clause). Commercial use is not permitted.