metadata
license: apache-2.0
pipeline_tag: video-text-to-text
library_name: transformers
This repository contains the Videollama3Qwen2ForCausalLM model, a reward model presented in the paper Learning Human-Perceived Fakeness in AI-Generated Videos via Multimodal LLMs.
The model is designed to detect human-perceived deepfake traces in AI-generated videos. It takes multimodal input and provides natural-language explanations, bounding-box regions for spatial grounding, and precise onset/offset timestamps for temporal labeling. It was trained on the DeeptraceReward benchmark, which is the first fine-grained, spatially- and temporally-aware dataset for annotating human-perceived fake traces.